IoT framework for sports activity safety monitoring based on wearable sensors and CRNN spatiotemporal analysis

Abstract

The integration of wearable sensors and IoT technology provides new technical means for sports activity monitoring. However, existing solutions still have deficiencies in joint spatiotemporal feature modeling, edge-side lightweight deployment, and secure transmission of sensitive data. This paper proposes an IoT sports activity monitoring framework based on CRNN spatiotemporal analysis and secure transmission mechanisms, adopting a layered architecture to achieve end-to-end processing from data acquisition to intelligent services. The spatiotemporal analysis module extracts spatial features of motion signals through convolutional neural networks, models temporal dependencies using long short-term memory networks, and introduces self-attention mechanisms to enhance the representation capability of key motion patterns. Addressing the resource constraints of edge devices, 8-bit quantization and knowledge distillation techniques are employed for lightweight model compression, reducing parameters from 2.34 M to 0.58 M with embedded inference latency of only 47.3ms. The secure transmission module adopts AES-128 encryption algorithm and HMAC-SHA256 message authentication code to ensure confidentiality and integrity of data transmission. A comparative analysis of encryption-based security and data-driven privacy protection methods such as differential privacy is provided, clarifying the applicability and complementary relationships of different privacy protection strategies in resource-constrained wearable scenarios. Experimental results on four datasets (UCI-HAR, PAMAP2, WISDM, and self-built sports activity dataset) demonstrate that the proposed method achieves an average recognition accuracy of 95.99%, improving by 2.85 to 3.52% points over baseline methods. The end-to-end latency of secure transmission is 23.6ms, with energy consumption increase controlled within 12.4%. This framework provides a feasible technical solution for the design of intelligent sports monitoring systems with promising application prospects.

Introduction

The deep integration of Internet of Things (IoT) technology and artificial intelligence algorithms is reshaping research paradigms in sports science. Wearable sensor devices have become core tools for collecting athletes’ physiological parameters and monitoring motion states due to their non-invasive nature, portability, and real-time capabilities. Multi-modal sensing units such as accelerometers, gyroscopes, and heart rate sensors can continuously acquire dynamic information during human movement, providing a solid data foundation for motion performance analysis, injury risk warning, and rehabilitation progress assessment. The intelligence level of sports activity monitoring systems directly affects the scientific evaluation of training effectiveness and effective protection of sports safety. Traditional cloud computing models have obvious limitations in data transmission latency, privacy leakage risk, and network bandwidth occupation, prompting researchers to turn their attention to the technical route of edge intelligence and localized inference.

In the field of wearable sports monitoring, domestic and international scholars have conducted extensive fruitful exploratory work. De Fazio et al.¹ systematically reviewed the application progress of wearable sensors in rehabilitation parameter monitoring and motion performance tracking, pointing out that the selection of electromechanical transduction mechanisms has a decisive impact on monitoring accuracy. Seçkin et al.² comprehensively evaluated sports wearable technology from three dimensions: hardware architecture, software platform, and commercial products, emphasizing that privacy protection and data accuracy remain key bottlenecks constraining the development of this field. Deep learning methods have demonstrated powerful feature extraction capabilities in human activity recognition tasks. Zhang et al.³’s review research shows that hybrid architectures of convolutional neural networks and recurrent neural networks can simultaneously capture spatial features and temporal dependencies of sensor signals. Khatun et al.⁴ proposed a CNN-LSTM self-attention model that achieved excellent recognition accuracy on multiple public datasets, validating the effectiveness of attention mechanisms for focusing on key motion patterns. Wang et al.⁵ constructed a cloud-based deep learning sports health monitoring system, achieving automated prediction of athletes’ health status. Chen and Dai⁶ explored the synergistic role of artificial intelligence and IoT technology in sports risk factor identification, revealing the application potential of data-driven methods in sports safety assurance. At the data security level, Qi et al.⁷ systematically reviewed the application of blockchain technology in privacy protection of medical health IoT, providing technical reference for secure data transmission of wearable devices. Beyond cryptographic approaches, data-level privacy protection mechanisms have also attracted considerable attention. Wang et al.⁸ proposed a differential privacy-based security level protection scheme for intelligent terminals in IoT architecture, utilizing security level and level proportion as parameters to apply differentiated privacy protection intensities, demonstrating the effectiveness of noise perturbation techniques in preventing sensitive information leakage from shared data. Li et al.⁹ designed a local differential privacy protection mechanism specifically for wearable device data collection, combining the Laplace mechanism with random response to achieve a balance between privacy guarantee and data utility for physiological stream data. Schizas et al.¹⁰’s research shows that the TinyML framework can achieve local inference of machine learning models on ultra-low-power microcontrollers, significantly reducing edge devices’ dependence on cloud services. The above research has promoted the progress of intelligent sports monitoring technology from different perspectives, but existing solutions still have optimization space in spatiotemporal feature fusion modeling, lightweight model deployment, and end-to-end secure transmission.

Addressing the practical needs of wearable sports activity monitoring scenarios, this thesis proposes an IoT monitoring framework integrating CRNN spatiotemporal analysis and secure transmission mechanisms. The overall architecture of the framework is shown in Fig. 1, mainly comprising four functional modules: data perception layer, edge computing layer, secure transmission layer, and cloud service layer. The innovations of this research are reflected in three aspects: first, designing a spatiotemporal feature extraction module based on convolutional recurrent neural networks, capturing local spatial patterns of motion signals through multi-scale convolution kernels, modeling long-term temporal dependencies using LSTM units, and introducing self-attention mechanisms to enhance the representation capability of key motion features; second, employing model quantization and knowledge distillation techniques to achieve lightweight compression of deep learning models, enabling them to complete real-time inference tasks on resource-constrained embedded platforms; third, constructing a secure data transmission channel based on lightweight encryption algorithms, effectively preventing leakage and tampering of sensitive motion data while ensuring transmission efficiency; furthermore, this research provides a comparative discussion of encryption-based security and data-driven privacy protection methods such as differential privacy, clarifying the applicability boundaries and complementary relationships of different privacy protection strategies in resource-constrained wearable scenarios. The research findings of this thesis are expected to provide theoretical guidance and technical support for the design and implementation of intelligent sports monitoring systems.

Fig. 1.

Overall architecture of IoT-based sports activity safety monitoring framework.

Related work and basic concepts

Research on IoT-based sports activity monitoring

The application of IoT technology in the field of sports activity monitoring has experienced a development process from single-parameter acquisition to multi-dimensional comprehensive perception. Early motion monitoring systems mainly relied on independent sensor nodes to complete measurements of specific physiological indicators, with data processing requiring manual intervention, and both timeliness and accuracy of monitoring results having considerable room for improvement. The introduction of IoT architecture broke through the technical barriers of traditional monitoring modes, achieving organic coordination among sensing devices, communication networks, and computing platforms. Patalas-Maliszewska et al.¹¹ developed a motion activity consultation system based on inertial sensors, which uses machine learning algorithms to intelligently analyze collected motion data, providing athletes with personalized training advice and technical guidance. The core value of IoT sports monitoring systems lies in constructing a complete closed-loop chain from data acquisition to information services, making real-time perception, rapid transmission, and intelligent decision-making of motion states possible. Zhou et al.¹²’s proposed multi-sensor data fusion scheme collaboratively processes measurement information from accelerometers and gyroscopes, effectively improving the robustness of activity recognition in complex motion scenarios. Current IoT sports monitoring research shows a development trend of continuously deepening intelligence and continuously expanding application scenarios, with researchers exploring technical routes to sink edge computing capabilities to sensing terminals, aiming to reduce system response latency and energy overhead while ensuring monitoring accuracy.

Wearable sensor technology

Wearable sensors constitute the data perception foundation of sports activity monitoring systems, with their performance parameters directly determining the quality upper limit of subsequent analysis and processing. The continuous progress of MEMS technology has driven inertial measurement units toward miniaturization, low power consumption, and high precision. Current mainstream motion monitoring devices typically integrate multiple types of sensors such as three-axis accelerometers, three-axis gyroscopes, and three-axis magnetometers. Raj and Kos¹³’s research shows that acceleration signals can effectively reflect the intensity characteristics and periodic patterns of human motion, while gyroscope data excels at capturing dynamic information of limb rotation and posture changes. The collaborative use of both types of sensors can achieve comprehensive characterization of motion states. The choice of sensor placement positions has a significant impact on monitoring effectiveness, with wrists, waists, and ankles being the most common wearing positions in motion monitoring scenarios, with signals collected at different positions presenting differentiated feature distribution patterns. Luwe et al.¹⁴ designed a hybrid deep learning model targeting the characteristics of wearable sensor data, fully exploiting complementary information among multi-channel signals. Physiological parameter sensors also play important roles in sports monitoring applications, with indicators such as heart rate, blood oxygen saturation, and skin conductance reflecting athletes’ physical load states and fatigue levels. Dirgová Luptáková et al.¹⁵ employed Transformer architecture to process time-series data from wearable sensors, validating the effectiveness of self-attention mechanisms in modeling long-range dependencies. Technical details such as sensor data sampling frequency, quantization precision, and synchronization mechanisms all require refined design based on specific application scenarios.

Application of deep learning in human activity recognition

Human activity recognition tasks aim to achieve automatic determination of user behavior categories through analysis and processing of sensor signals. The introduction of deep learning methods has brought revolutionary technological breakthroughs to this field. Convolutional neural networks have been widely applied in sensor signal processing due to their powerful local feature extraction capabilities, with multi-layer convolution operations automatically learning hierarchical spatial patterns contained in raw data. Duan et al.¹⁶ combined edge computing with gated recurrent units to construct a distributed activity recognition framework, maintaining high recognition accuracy while reducing cloud computing load. Recurrent neural networks and their variant structures are specifically designed for temporal modeling needs of sequential data, with long short-term memory networks effectively solving the gradient vanishing problem of traditional recurrent networks through introducing gating mechanisms. Wang et al.¹⁷’s designed self-attention deep convolutional LSTM framework achieved excellent performance in badminton motion recognition tasks, proving the effectiveness of spatiotemporal feature joint modeling strategies. The cascaded combination of convolutional neural networks and recurrent neural networks forms the CRNN architecture, which can simultaneously capture spatial features and temporal dependencies of sensor signals. Koşar and Barshan¹⁸’s proposed novel CNN-LSTM architecture significantly improved the generalization capability of activity recognition through diversified feature extraction mechanisms. Han et al.¹⁹ employed heterogeneous convolutional neural networks to process sensor data from different wearing positions, achieving robust recognition performance in cross-device scenarios. Mim et al.²⁰ fused Inception modules with gated recurrent units, constructing the GRU-INC model that balances multi-scale feature extraction and temporal modeling. The performance advantages of deep learning models are built on the foundation of large-scale labeled data and sufficient computing resources, and how to achieve efficient inference on resource-constrained embedded platforms remains a hot research topic.

Secure transmission mechanisms in IoT systems

Motion and physiological data collected by wearable devices are highly sensitive, facing multiple security threats such as theft, tampering, and misuse during transmission and storage. Traditional security protection schemes are difficult to directly transplant to resource-constrained IoT terminal devices, making the design of lightweight encryption algorithms and efficient authentication protocols key technical support for ensuring data security. Chatterjee et al.²¹ proposed lightweight block cipher technology for medical health IoT scenarios, significantly reducing computational complexity and memory occupation while ensuring encryption strength. Yu and Park²² utilized physically unclonable functions to design efficient authentication schemes for wearable computing, effectively resisting security threats such as device spoofing and replay attacks. The decentralized characteristics of blockchain technology provide new solutions for trusted storage and traceability verification of IoT data. Rajasekaran et al.²³ constructed an anonymous privacy-preserving authentication framework based on blockchain, achieving secure management of user identity information in health IoT scenarios. Li et al.²⁴ combined blockchain with lightweight secret sharing technology, achieving a balance between privacy protection and efficiency in medical IoT data transmission. Pelekoudas-Oikonomou et al.²⁵ systematically analyzed the application modes of blockchain security mechanisms in medical IoT edge networks, providing reference for security architecture design of wearable health monitoring systems.

In addition to cryptographic and blockchain-based approaches, data-level privacy protection mechanisms have emerged as an important complementary research direction. Differential privacy, originally proposed for statistical databases, has been increasingly applied to IoT and intelligent terminal scenarios. Wang et al.⁸ designed a level-proportion-based differential privacy protection method that applies differentiated protection intensities according to the security levels of intelligent terminals, achieving configurable privacy guarantees while maintaining the statistical properties of the dataset for further utilization. Li et al.⁹ further extended differential privacy to the wearable device domain, proposing a local differential privacy scheme for physiological stream data collection that combines adaptive Laplace noise injection with data reconstruction techniques, effectively protecting heart rate and other sensitive health data while preserving data utility. These data-driven privacy approaches address a fundamentally different threat model from encryption: while encryption protects data during transit, differential privacy can mitigate privacy leakage from the analytical results themselves, such as inference attacks that may reveal individual behavioral patterns from aggregate analysis outputs. However, the noise injection inherent in differential privacy typically introduces accuracy degradation, posing challenges for real-time activity recognition on resource-constrained wearable devices where recognition precision is critical.

Alajlan and Ibrahim²⁶’s research shows that TinyML technology can achieve local inference of machine learning models on ultra-low-power edge devices, reducing the need for transmitting sensitive data to the cloud and reducing the risk of data leakage at the source. The design of secure transmission mechanisms needs to seek optimal balance among protection strength, computational overhead, and transmission efficiency. Moreover, existing solutions rarely provide a unified consideration of both transmission-stage security and analytics-stage privacy protection, and how to achieve multi-layered privacy protection while meeting the stringent resource constraints of wearable devices remains an open research problem that warrants further investigation.

Methods and framework

System architecture overview

The sports activity safety monitoring framework proposed in this thesis adopts a layered design philosophy, dividing system functions into four cooperating functional modules: data perception layer, edge computing layer, secure transmission layer, and cloud service layer. This architectural design fully considers the resource constraint characteristics of wearable devices, the real-time requirements of motion data, and the security protection requirements of sensitive information, achieving end-to-end processing from raw signal acquisition to intelligent analysis services. The overall system architecture is shown in Fig. 2, with standardized interfaces for data exchange between functional layers, ensuring loose coupling characteristics between modules and system scalability.

Fig. 2.

Overall architecture diagram of sports activity safety monitoring system.

The data perception layer undertakes the fundamental task of motion information collection, composed of wearable sensor nodes distributed at different parts of athletes’ bodies. Gopalakrishnan et al.²⁵’s research shows that inertial measurement units have broad application prospects in athlete physical fitness monitoring, and reasonable sensor layout can significantly improve the representation completeness of motion features. The edge computing layer is the core processing unit of this framework, responsible for completing computation-intensive tasks such as sensor data preprocessing, feature extraction, and activity recognition. Bourechak et al.²⁶ pointed out that the integration of artificial intelligence and edge computing is reshaping the technological form of IoT applications, and sinking intelligent inference capabilities to edge nodes can effectively reduce system latency and reduce bandwidth occupation. The secure transmission layer provides encryption protection and integrity verification services for sensitive motion data network transmission, ensuring data security in open network environments. The cloud service layer provides resource-intensive services such as large-scale data storage, complex model training, and cross-user data analysis, supporting long-term archiving and deep mining applications of motion data.

Data acquisition and preprocessing

The data acquisition quality of wearable sensors directly affects the performance of subsequent activity recognition algorithms. This framework selects three-axis accelerometers and three-axis gyroscopes as the main motion sensing components, supplemented by photoplethysmography sensors to achieve synchronous acquisition of heart rate signals. Raman et al.²⁷’s research confirms that collaborative use of multi-modal sensors can provide more comprehensive information support for sports activity monitoring. Key configuration parameters of the sensors are shown in Table 1, with sampling frequency settings requiring trade-offs between signal fidelity and power overhead.

Table 1.

Wearable sensor parameter configuration.

Sensor type	Measurement range	Sampling frequency	Resolution	Power consumption
Three-axis accelerometer	± 16 g	50 Hz	16bit	0.5mW
Three-axis gyroscope	± 2000°/s	50 Hz	16bit	0.8mW
Heart rate sensor	30-240 bpm	1 Hz	12bit	0.3mW

Raw sensor signals typically contain interference components such as measurement noise, baseline drift, and outliers. The task of the data preprocessing stage is to eliminate these interference factors and convert signals into a standard format suitable for deep learning model processing. This framework employs a fourth-order Butterworth low-pass filter to smooth acceleration and angular velocity signals, with the filter cutoff frequency set at 20 Hz, a parameter that can effectively filter out high-frequency noise while preserving the main frequency components of human motion. Filtered signals are segmented using sliding window techniques, with window length set at 2.56 s and adjacent windows using 50% overlap rate to ensure motion events are not cut by window boundaries. Zeng et al.²⁸’s proposed residual multi-feature fusion shrinkage network adopted a similar data segmentation strategy, validating the effectiveness of this method in activity recognition tasks.

Let the raw sensor signal sequence be , where T represents the number of time steps and C represents the number of sensor channels. The signal after low-pass filtering can be expressed as:

1

where H represents the impulse response function of the filter and * represents convolution operation. The filtered signal undergoes Z-score normalization:

2

where μ and σ represent the mean vector and standard deviation vector calculated on the training dataset, respectively. The normalization operation eliminates dimensional differences among different sensor channels, making features of each channel have the same numerical scale.

CRNN-based spatiotemporal analysis model

The core algorithm module of this framework adopts a convolutional recurrent neural network architecture to achieve intelligent recognition of sports activities, which simultaneously models spatial features and temporal dependencies of sensor signals through cascaded combination of convolutional neural networks and long short-term memory networks. Zhou et al.²⁹’s research shows that multi-level network structures can effectively improve human activity recognition performance based on wearable sensors. The overall structure of the CRNN model is shown in Fig. 3, including three core components: convolutional feature extraction module, LSTM temporal modeling module, and self-attention enhancement module.

Fig. 3.

Structure diagram of CRNN-based spatiotemporal analysis model.

The convolutional feature extraction module is composed of multiple cascaded convolution blocks, with each convolution block containing a one-dimensional convolutional layer, batch normalization layer, ReLU activation function, and max pooling layer. One-dimensional convolution operations extract local features along the time axis direction of sensor signals, with the mathematical expression:

3

where represents the feature map of the -th layer, and represent the weight matrix and bias vector of the l-th layer convolution kernel, respectively, represents batch normalization operation, and represents rectified linear unit activation function. This framework adopts a three-layer convolution structure with convolution kernel numbers set at 64, 128, and 256, respectively, and convolution kernel size uniformly set at 5. This configuration can effectively capture multi-scale spatial patterns in motion signals. Jameer and Syed³⁰’s research validated the discriminative ability of deep convolution features in wearable sensor activity recognition.

The LSTM temporal modeling module receives the feature sequence output from the convolutional feature extraction module and models long-range temporal dependencies through gating mechanisms. The core computation process of LSTM units includes four key components: forget gate, input gate, candidate memory, and output gate, with mathematical expressions as follows:

4

5

6

7

8

9

where , , represent activation values of forget gate, input gate, and output gate, respectively, represents cell state, represents hidden state output, represents Sigmoid activation function, and represents element-wise multiplication operation. This framework adopts a two-layer stacked LSTM structure with hidden state dimension set at 128, and the two-layer structure can learn more abstract temporal representations.

The introduction of the self-attention enhancement module aims to enable the model to dynamically focus on key time steps in the input sequence, improving recognition capability for important motion patterns. This module generates attention weight distributions by calculating similarities among query vectors, key vectors, and value vectors, with the computation process expressed as:

10

where , , represent query matrix, key matrix, and value matrix, respectively, represents the dimension of key vectors, and dividing by prevents dot product results from becoming too large and causing the Softmax function to enter saturation regions. Velykoivanenko et al.³¹’s review research pointed out that attention mechanisms demonstrate significant performance advantages in wearable device data analysis.

Secure transmission mechanism design

Motion and physiological data collected by wearable devices involve user privacy and require effective security protection measures during network transmission. Samuel et al.³²’s proposed federated learning and blockchain-driven medical IoT system provides important reference for the security mechanism design of this framework. This framework employs lightweight symmetric encryption algorithms to achieve data confidentiality protection, message authentication code technology to ensure data integrity, and challenge-response protocol-based bidirectional authentication mechanisms to ensure the authenticity of identities of both communicating parties. The overall process of the secure transmission mechanism is shown in Fig. 4.

Fig. 4.

Secure transmission mechanism flow diagram.

Considering the limited computing resources and energy supply of wearable devices, this framework selects the AES-128 algorithm as the core component of data encryption. Xu et al.³³’s proposed blockchain-enabled privacy-preserving authentication protocol validated the feasibility of lightweight encryption schemes in medical IoT scenarios. The AES encryption algorithm processes 128-bit data blocks, with the encryption process including four basic operations: byte substitution, row shift, column mixing, and round key addition, executing 10 rounds of transformations. Let the plaintext data block be and the encryption key be , then the encryption process can be expressed as:

11

where represents the ciphertext data block and represents the encryption function with as the key. This framework adopts counter mode to achieve streaming data encryption processing. This mode supports parallel computation and requires no padding operations, making it particularly suitable for processing continuous sensor data streams.

Message authentication codes are used to verify the integrity of transmitted data and prevent data from being tampered with during transmission. This framework employs the HMAC-SHA256 algorithm based on hash functions, with the computation process expressed as:

12

Where M represents the message to be authenticated, represents the SHA-256 hash function, and represent outer padding constant and inner padding constant, respectively, Å represents XOR operation, and | represents string concatenation operation. The receiver verifies data integrity by recalculating the HMAC value and comparing it with the received authentication code.

It is important to note that the security design of this framework primarily targets transmission-stage threats, ensuring confidentiality and integrity of sensor data during network communication. However, privacy risks may also arise from the analytical results themselves. Activity recognition outputs and behavioral pattern statistics transmitted to or stored in the cloud could potentially reveal sensitive information about users’ daily routines, health conditions, or physical capabilities. Data-level privacy protection mechanisms, such as differential privacy, offer complementary protection by injecting calibrated noise into shared data or model outputs, thereby preventing inference of individual records from aggregate analysis results⁸. Compared with encryption-based approaches adopted in this framework, differential privacy and federated learning address different threat models: encryption protects data in transit with zero impact on recognition accuracy, while differential privacy can mitigate privacy leakage from analytical results at the cost of introducing noise that typically degrades recognition performance by 1–5% depending on the privacy budget⁹. Federated learning avoids centralizing raw data but incurs significant communication overhead, which poses challenges for bandwidth-constrained wearable devices. Given the stringent latency requirements (sub-50ms inference) and energy constraints of the target deployment scenario, this framework adopts lightweight encryption as the primary security mechanism, which introduces manageable energy overhead (12.4% increase) without compromising recognition accuracy. Differential privacy and federated learning are recognized as promising complementary directions for future enhancement, particularly for scenarios involving multi-user data aggregation and cloud-side analytics^34,35.

Addressing the deployment needs of deep learning models on edge devices, this framework employs model quantization and knowledge distillation techniques to achieve lightweight compression of CRNN models. Li et al.³⁶’s review research systematically summarizes mainstream technical routes for deep neural network model compression. Model quantization technology reduces model storage space and computational complexity by lowering the numerical precision of network parameters. This framework adopts an 8-bit fixed-point quantization scheme to convert original 32-bit floating-point parameters to 8-bit integer representation. The quantization process can be expressed as:

13

where represents original floating-point weights, represents quantized integer weights, represents quantization step size, and n represents quantization bits. Hoefler et al.³⁷’s research shows that reasonable quantization strategies can achieve several-fold compression effects with controllable model accuracy loss. The deployment process of the lightweight model is shown in Fig. 5.

Fig. 5.

Lightweight model deployment flow diagram.

Knowledge distillation technology transfers model capabilities by training small student models to mimic the output behavior of large teacher models. Let the soft label output of the teacher model be and the student model output be , then the distillation loss function can be expressed as:

14

where represents cross-entropy loss function, represents true labels, represents KL divergence, T represents temperature coefficient, represents balance factor, and represents softened probability distribution calculated at temperature T. Comparative analysis of model compression strategies is shown in Table 2.

Table 2.

Comparative analysis of model compression strategies.

Compression Strategy	Compression Ratio	Accuracy Loss	Inference Speedup	Application Scenario
8-bit quantization	4×	< 1%	2–3×	General edge devices
Knowledge distillation	3–5×	1–2%	2–4×	Model structure optimization
Structural pruning	2–4×	1–3%	1.5-2×	Redundant parameter elimination
Combined compression	8–16×	2–4%	4–8×	Extreme resource constraints

This framework applies the above compression techniques in combination, achieving efficient deployment of CRNN models on embedded platforms, meeting the latency requirements of real-time activity recognition.

Experiments and results

Experimental setup

The experimental verification work of this research was completed on a software-hardware collaborative testing platform. The server side is configured with Intel Xeon Gold 6248R processor, 128GB memory, and NVIDIA RTX 3090 graphics card, with Ubuntu 20.04 LTS operating system and PyTorch 1.12 and TensorFlow 2.10 deep learning frameworks. The edge device side selects the STM32H743 microcontroller as the embedded inference platform, which features an ARM Cortex-M7 core with 480 MHz main frequency, 1 MB Flash storage, and 1 MB SRAM memory, meeting the deployment needs of lightweight deep learning models. Wearable sensor nodes adopt MPU-9250 nine-axis inertial measurement unit and MAX30102 heart rate sensor module, with data acquisition frequency set at 50 Hz, meeting Nyquist sampling requirements for human motion signals.

Experimental data comes from three public human activity recognition benchmark datasets and one self-built sports activity dataset. The UCI-HAR dataset (available at https://archive.ics.uci.edu/dataset/240/human+activity+recognition+using+smartphones) contains sensor records of 30 volunteers performing 6 types of daily activities, totaling 10,299 samples. The PAMAP2 dataset (available at https://archive.ics.uci.edu/dataset/231/pamap2+physical+activity+monitoring) covers multi-modal sensor data of 9 subjects completing 18 types of physical activities, with data collection points distributed at wrist, chest, and ankle positions. The WISDM dataset (available at https://www.cis.fordham.edu/wisdm/dataset.php) collected acceleration signals of 51 users performing 6 types of basic activities using smartphones. The self-built dataset (available at https://www.scidb.cn/s/MvmE7z) was collected for sports training scenarios, including 6 typical exercises: running, jumping, squatting, push-ups, sit-ups, and stretching. A total of 45 athletes aged 18 to 35 years were recruited for data collection, with each person completing approximately 5 min of each activity type, ultimately obtaining over 120,000 labeled samples. Detailed information about the datasets is shown in Table 3.

Table 3.

Basic information of experimental datasets.

Dataset name	Number of subjects	Activity categories	Total samples	Sensor type	Sampling frequency	Data source
UCI-HAR	30	6	10,299	Accelerometer, gyroscope	50 Hz	https://archive.ics.uci.edu/dataset/240
PAMAP2	9	18	376,417	IMU, heart rate	100 Hz	https://archive.ics.uci.edu/dataset/231
WISDM	51	6	1,098,207	Accelerometer	20 Hz	https://www.cis.fordham.edu/wisdm/dataset.php
Self-built dataset	45	6	121,500	IMU, heart rate	50 Hz	https://www.scidb.cn/s/MvmE7z

Model training employs the Adam optimizer with initial learning rate set at 0.001, batch size of 64, and 100 training epochs. The learning rate scheduling strategy adopts cosine annealing, decaying the learning rate to 0.5 times the original every 20 epochs. Training data and test data are divided in a 7:3 ratio, using leave-one-out cross-validation to ensure reliability of evaluation results. Model performance evaluation employs four metrics: accuracy, precision, recall, and F1 score, with calculation formulas as follows:

15

16

17

18

where , , , represent true positive, true negative, false positive, and false negative counts, respectively.

Dataset description

The collection of the self-built sports activity dataset took three months to complete, with all participants signing informed consent forms and confirming no exercise contraindications through physical examination. Data collection devices were worn at subjects’ right wrist and waist positions, with sensor nodes at each position synchronously recording three-axis acceleration, three-axis angular velocity, and heart rate signals. After quality screening of raw data, samples with sensor failures and labeling errors were removed, with the final valid data retained accounting for 94.3% of the original collection volume. The distribution of activity samples of each type is shown in Fig. 6, with sample numbers relatively balanced across different activity categories. The ratio of the largest category to the smallest category is 1.15:1, avoiding the impact of class imbalance problems on model training.

Fig. 6.

Bar chart of sample distribution for each activity type in self-built dataset.

Time-domain feature analysis of sensor signals shows that different sports activities exhibit significant differentiated characteristics in acceleration amplitude, motion period, and signal variability. Running activity acceleration signals present obvious periodic fluctuations, with step frequency of approximately 160 to 180 steps per minute and vertical acceleration peaks reaching 2.5 g. Jumping activity signal characteristics manifest as short-duration high-amplitude pulses, with acceleration impacts upon landing exceeding 4 g but lasting only 0.1 to 0.2 s. Strength training activities such as squats and push-ups have longer signal periods, with single action duration between 2 and 4 s and relatively moderate acceleration amplitudes. Typical activity sensor signal waveforms are shown in Fig. 7, where significant differences in time-domain representations among various activities can be observed.

Fig. 7.

Comparison of sensor signal waveforms for typical sports activities.

Evaluation metrics

This research employs a multi-dimensional evaluation system to comprehensively assess the performance of the proposed framework. Evaluation metrics for the activity recognition module include overall accuracy and precision, recall, and F1 score for each category. Model efficiency metrics include parameter count, floating-point operations, inference latency, and memory occupation. Evaluation metrics for the secure transmission module include encryption throughput, end-to-end latency, and energy overhead. System comprehensive metrics include real-time satisfaction rate, data transmission success rate, and battery life. Detailed definitions of the evaluation metric system are shown in Table 4.

Table 4.

Definition of evaluation metric system.

Metric category	Metric name	definition	Unit
Recognition performance	Accuracy	Proportion of correctly classified samples to total samples	%
Recognition performance	F1 score	Harmonic mean of precision and recall	%
Model efficiency	Parameter count	Total number of trainable parameters in the model	M
Model efficiency	Inference latency	Time consumption of single forward propagation	ms
Security performance	Encryption throughput	Amount of data encrypted per unit time	KB/s
Security performance	Transmission latency	Time from data collection to reception	ms

Results and comparative analysis

Activity recognition performance analysis

The CRNN spatiotemporal analysis model proposed in this framework achieved excellent recognition performance on all four datasets. The overall accuracy on the UCI-HAR dataset reached 96.82%, 94.57% on the PAMAP2 dataset, 97.34% on the WISDM dataset, and 95.23% on the self-built sports activity dataset. Comparison results with existing mainstream methods are shown in Table 5. The average accuracy of this framework on the four datasets improved by 3.41% and 2.76% over baseline methods CNN and LSTM, respectively, and by 1.28% over the recently proposed CNN-LSTM hybrid model.

Table 5.

Comparison of recognition accuracy of different methods on each dataset (%).

Method	UCI-HAR	PAMAP2	WISDM	Self-built Dataset	Average
CNN	93.24	90.86	94.12	91.67	92.47
LSTM	93.89	91.53	94.78	92.34	93.14
CNN-LSTM	95.42	93.21	96.15	94.08	94.72
BiLSTM-Attention	95.87	93.76	96.58	94.52	95.18
Proposed method	96.82	94.57	97.34	95.23	95.99

The introduction of the self-attention mechanism contributed significantly to model performance improvement. Ablation experiment results show that after removing the self-attention module, model accuracy decreased by 1.43% on the UCI-HAR dataset and 1.67% on the self-built dataset. Visualization analysis of attention weights reveals the model’s focusing ability on key time steps. In running activity recognition, attention weights mainly concentrate on the ground contact and push-off phases of gait cycles, while in jumping activity recognition, attention weights show peak distributions at takeoff and landing moments. Ablation experiment results are shown in Fig. 8.

Fig. 8.

Bar chart of ablation experiment results comparison.

Recognition performance for each activity category varies to some extent, with confusion matrix analysis intuitively reflecting the degree of recognition confusion among different activities. The confusion matrix on the self-built dataset is shown in Fig. 9, with running and jumping high-intensity activities achieving recognition accuracies of 97.2% and 96.8%, respectively, performing most excellently. There is approximately 2.3% mutual misclassification rate between squats and sit-ups, as these two activities have similarities in periodic characteristics of sensor signals, making model discrimination relatively difficult. Stretching activity recall rate is 93.4%, with some light stretching movements misclassified as stationary states.

Fig. 9.

Heatmap of activity recognition confusion matrix on Self-built Dataset.

Model efficiency evaluation

The application of lightweight compression strategies enables CRNN models to achieve efficient deployment on resource-constrained embedded platforms. The parameter count of the original model is 2.34 M, reduced to 0.58 M after 8-bit quantization and knowledge distillation processing, achieving a compression ratio of 4.03 times. Model inference latency on the STM32H743 microcontroller improved from the original unable-to-run state to requiring only 47.3ms for single inference, meeting the response time requirements of real-time activity recognition. The impact of different compression strategies on model performance is shown in Table 6.

Table 6.

Performance comparison of model compression strategies.

Compression configuration	Parameter count (M)	Model size (KB)	Inference latency (ms)	Accuracy (%)	Accuracy loss (%)
Original model	2.34	9360	-	95.23	-
8-bit quantization	2.34	2340	89.6	94.87	0.36
Knowledge distillation	0.78	3120	62.4	94.52	0.71
Quantization + distillation	0.78	780	53.8	94.16	1.07
Complete compression	0.58	580	47.3	93.89	1.34

Model inference performance on different hardware platforms shows significant differences. GPU platform has the lowest inference latency, requiring only 3.2ms to process 64 samples per batch. CPU platform inference latency is 28.7ms, meeting the needs of most real-time applications. Embedded platforms, limited by computing resources, have relatively longer inference latency, but can still reach practical levels after optimization. Inference performance comparison across platforms is shown in Fig. 10.

Fig. 10.

Inference latency comparison across hardware platforms.

Secure transmission performance analysis

The performance evaluation of the secure transmission module covers three dimensions: encryption efficiency, transmission latency, and energy overhead. The throughput of the AES-128 encryption algorithm on the STM32H743 platform reaches 1.28 MB/s, meeting the bandwidth requirements for real-time encryption of sensor data. The HMAC-SHA256 message authentication code computation speed is 0.86 MB/s, with single authentication operation latency of 0.47ms. The overall latency of the secure transmission module consists of encryption latency, network transmission latency, and decryption latency. The end-to-end latency in WiFi network environment is 23.6ms, and in 4G cellular network environment is 67.4ms. Secure transmission performance test results are shown in Table 7.

Table 7.

Secure transmission module performance test results.

Performance metric	WiFi environment	4G environment	Bluetooth environment
Encryption latency (ms)	3.2	3.2	3.2
Network latency (ms)	15.8	58.6	8.4
Decryption latency (ms)	4.6	5.6	4.1
End-to-end latency (ms)	23.6	67.4	15.7
Transmission success rate (%)	99.7	98.3	99.9
Energy consumption increase ratio (%)	12.4	18.7	8.6

The introduction of security mechanisms inevitably brings additional computational and communication overhead. Experimental results show that after enabling encrypted transmission, system overall energy consumption increased by approximately 12.4% to 18.7% compared to unencrypted state, with specific increases depending on network type and data transmission frequency. In typical usage scenarios, wearable devices equipped with 500mAh batteries can support approximately 6.2 h of continuous monitoring, a reduction of about 12.7% from 7.1 h battery life in unencrypted state. Energy consumption distribution analysis is shown in Fig. 11, with sensor acquisition accounting for 31.2% of total energy consumption, data processing 28.4%, security encryption 14.6%, and wireless communication 25.8%. Notably, the encryption-based security mechanism adopted in this framework introduces an energy overhead of only 12.4% in WiFi environment with no impact on recognition accuracy. In contrast, differential privacy mechanisms typically incur 1–5% recognition accuracy degradation depending on the privacy budget configuration, while federated learning approaches substantially increase communication overhead due to frequent model parameter exchanges between edge devices and cloud servers. These experimental observations further support the adoption of lightweight encryption as the primary security strategy for resource-constrained wearable deployment scenarios, while also highlighting the potential of integrating differential privacy at the cloud analytics stage where computational resources are less constrained.

Fig. 11.

Pie chart of system energy consumption distribution.

System comprehensive performance evaluation

The comprehensive performance of this framework in actual deployment environments was validated through two weeks of field testing. During the testing period, 12 athletes wore devices to participate in daily training, with the system cumulatively operating over 320 h, collecting and processing over 58 million sensor data points. The online accuracy of the activity recognition module was 94.16%, slightly lower than offline test results, mainly due to transition actions and compound actions in actual motion scenarios not covered by training data. Data transmission success rate reached 99.2%, with packet loss events mainly occurring in network switching and weak signal coverage areas. No security incidents occurred during the two-week testing period, with all transmitted data passing integrity verification.

Statistical analysis of system response time shows that the full-chain latency from sensor data acquisition to activity recognition result output presents an approximately normal distribution. In WiFi network environment, average response time is 76.4ms, 95th percentile response time is 112.3ms, and maximum response time is 187.6ms. Response time distribution is shown in Fig. 12, with the vast majority of requests concentrated in the 50ms to 100ms interval, meeting the real-time requirements of sports activity monitoring scenarios.

Fig. 12.

Histogram of system response time distribution.

Comparative analysis with existing commercial sports monitoring systems further validates the technical advantages of this framework. In activity recognition accuracy, this framework improves by approximately 8.3% over a commercial smart bracelet and 4.7% over a professional sports watch. In data security, this framework achieves end-to-end encryption protection, while both tested commercial products adopt plaintext transmission. In battery life, the energy consumption performance of this framework is basically on par with commercial products. Comprehensive comparison results are shown in Table 8.

Table 8.

Comprehensive performance comparison with commercial systems.

Performance metric	Proposed framework	Commercial Bracelet A	Commercial Watch B
Activity recognition accuracy (%)	94.16	85.82	89.43
Number of supported activity types	6	4	5
Data encryption	AES-128	None	None
Average response latency (ms)	76.4	52.3	68.7
Battery life (hours)	6.2	7.8	6.5
Transmission success rate (%)	99.2	99.5	99.1

Discussion and implications

Discussion

The IoT sports activity monitoring framework based on CRNN spatiotemporal analysis and secure transmission mechanisms proposed in this research demonstrates good performance in multiple dimensions, with experimental results validating the effectiveness and practicality of the proposed technical solution. From activity recognition accuracy perspective, the average accuracy of this framework on four test datasets reached 95.99%, improving by 3.52% points over traditional CNN methods and 2.85% points over standard LSTM methods. This performance gain mainly stems from the CRNN architecture’s joint modeling capability for spatiotemporal features of sensor signals and the self-attention mechanism’s dynamic focusing effect on key motion patterns. Ablation experiment results further confirm the contribution value of each technical component, with removal of the self-attention module causing accuracy decreases of 1.43% to 1.67%, indicating that attention mechanisms play an irreplaceable role in distinguishing similar motion patterns. The application of lightweight compression strategies reduced model parameter count from 2.34 M to 0.58 M, achieving a compression ratio of 4.03 times, while accuracy loss was only 1.34% points. This result shows that the combined application of 8-bit quantization and knowledge distillation techniques can achieve an optimal balance between model efficiency and recognition accuracy. The secure transmission module’s end-to-end latency in WiFi environment is 23.6ms, with energy consumption increase after enabling encryption controlled at 12.4%. These metrics are all within acceptable ranges, proving the feasibility of deploying lightweight encryption schemes on resource-constrained devices.

From the perspective of privacy protection scope, it should be noted that the security design of this framework primarily addresses transmission-stage threats through AES-128 encryption and HMAC-SHA256 integrity verification. However, privacy protection for wearable sports monitoring data is inherently a multi-layered problem. Beyond transmission security, the analytical results generated at the cloud side, such as activity recognition outputs and long-term behavioral pattern statistics, could also potentially reveal sensitive information about users’ exercise habits, health status, or daily routines. Data-level privacy protection mechanisms such as differential privacy can complement encryption by mitigating such inference-stage privacy risks through calibrated noise injection⁸. Nevertheless, the noise introduced by differential privacy inevitably degrades recognition accuracy, and the experimental results of this study indicate that even a 1.34% point accuracy loss from model compression already represents a noticeable performance trade-off for practical applications. Therefore, in the current resource-constrained wearable deployment scenario, encryption-based security achieves a favorable balance between protection effectiveness and system performance. For future multi-user data aggregation and cloud-side analytics scenarios where computational resources are less constrained, integrating differential privacy mechanisms at the data aggregation or model training stage represents a viable path toward more comprehensive privacy protection.

Comprehensive performance comparison of this framework with existing research methods is shown in Table 9, where competitive advantages in three key dimensions of recognition accuracy, model lightweight degree, and security protection capability can be observed.

Table 9.

Comprehensive performance comparison of this framework with existing methods.

Comparison dimension	Proposed framework	CNN-LSTM method	BiLSTM-attention method	Commercial system
Average recognition accuracy (%)	95.99	94.72	95.18	85.82–89.43
Model parameter count (M)	0.58	1.86	2.12	Not disclosed
Embedded inference latency (ms)	47.3	Cannot deploy	Cannot deploy	52.3–68.7
Data encryption support	AES-128	None	None	None
End-to-end secure transmission	Supported	Not supported	Not supported	Not supported
Real-time satisfaction rate (%)	98.7	–	–	99.2

It is worth noting that this framework’s online accuracy (94.16%) in actual deployment environments shows approximately 1% point performance degradation compared to offline test results (95.23%). This phenomenon is universal in wearable activity recognition fields, with causes mainly including three aspects: actual motion scenarios contain transition actions and compound actions not covered by training data, motion patterns among different users have individual variability, and slight deviations in sensor wearing positions affect signal acquisition consistency. The approximately 2.3% mutual misclassification rate between squats and sit-ups reflects the similarity of periodic strength training movements at the sensor signal level. This finding suggests future research could consider introducing more refined temporal modeling strategies or fusing additional sensing modalities to improve discrimination capability for similar movements. System response time distribution analysis shows 95th percentile response time is 112.3ms, meaning the system may still experience response latencies exceeding 100ms in extreme cases. For some application scenarios with extremely high real-time requirements (such as immediate warning of sports injuries), further optimizing edge inference efficiency and network transmission performance still has important research value.

Implications

The findings of this research provide several important theoretical guidelines and practical references for the design and implementation of intelligent sports monitoring systems. From technical architecture design perspective, the layered decoupled system architecture can effectively address the inherent contradiction between wearable device resource constraints and intelligent analysis computational intensity. The technical route of sinking lightweight inference capabilities to edge nodes demonstrates significant advantages in reducing response latency and protecting data privacy. This architectural philosophy can be extended to related fields such as health monitoring, rehabilitation training, and sports injury prevention. The successful application of CRNN spatiotemporal analysis models shows that deep learning methods have powerful feature learning capabilities when processing multi-channel time-series sensor data, while the introduction of self-attention mechanisms provides interpretable decision-making basis for models. Visualization analysis of attention weights helps sports science researchers understand the contribution degree of different motion phases to activity recognition. This interpretability feature has important practical significance for applying artificial intelligence technology to sports training guidance. The validation of model compression technology effectiveness proves the feasibility of achieving lightweight deployment of deep learning models with controllable accuracy loss. The combination strategy of 8-bit quantization and knowledge distillation provides reusable technical solutions for embedded artificial intelligence applications.

From application promotion perspective, the end-to-end secure transmission mechanism adopted by this framework fills the gap in data security protection of existing commercial sports monitoring products. With increasing user privacy protection awareness and improvement of relevant laws and policies, wearable devices with built-in security capabilities will gain broader market space. Furthermore, the integration of data-level privacy mechanisms such as differential privacy with transmission-layer encryption could establish a multi-layered privacy protection architecture, addressing both in-transit and post-analysis threats. Such a comprehensive approach would be particularly valuable as regulatory requirements for user data protection in sports and health domains continue to strengthen. This research’s experimental results show that under current hardware conditions, wearable sports monitoring devices can achieve activity recognition accuracy exceeding 94% and complete data encryption protection while ensuring battery life of over 6 h. This performance level can already meet basic needs of daily training monitoring, but still has gaps from the ideal goal of all-day continuous monitoring. Progress in battery technology and low-power computing chips will be key external factors driving this field’s development. The self-built sports activity dataset constructed in this research covers 6 typical training movements, providing reusable data resources for subsequent research. However, the coverage range of exercise types is still limited, and establishing more comprehensive sports activity benchmark datasets will help promote standardized development of this field and fair comparison among different research achievements.

Conclusion

This research addresses three core technical challenges in wearable sports activity monitoring scenarios: spatiotemporal feature modeling, edge intelligence deployment, and secure data transmission, proposing an IoT monitoring framework integrating CRNN spatiotemporal analysis and lightweight encryption mechanisms. The framework adopts layered architectural design, achieving end-to-end processing from sensor data acquisition to intelligent analysis services. The spatiotemporal analysis model based on convolutional recurrent neural networks extracts local spatial patterns of motion signals through multi-layer convolution operations, models long-range temporal dependencies using LSTM units, and introduces self-attention mechanisms to enhance representation capability for key motion features. Experimental results show that the proposed model achieved an average recognition accuracy of 95.99% on four datasets (UCI-HAR, PAMAP2, WISDM, and self-built sports activity dataset), improving by 3.52 and 2.85% points over traditional CNN and LSTM methods, respectively, and 1.27% points over CNN-LSTM hybrid model. The application of model compression strategies reduced parameter count from 2.34 M to 0.58 M, with inference latency on the STM32H743 embedded platform of only 47.3ms and accuracy loss controlled within 1.34% points. The secure transmission module employs AES-128 encryption algorithm and HMAC-SHA256 message authentication code, with end-to-end latency of 23.6ms in WiFi environment and energy consumption increase ratio of 12.4%, achieving effective balance among data confidentiality, integrity, and transmission efficiency. A comparative analysis of encryption-based security and data-driven privacy protection methods such as differential privacy further clarifies that lightweight encryption is the most suitable primary security strategy for resource-constrained wearable devices, while data-level privacy mechanisms offer complementary protection for cloud-side analytics scenarios.

This research still has several limitations that need to be improved in subsequent work, including limited exercise type coverage in the self-built dataset, model adaptability to individual differences needing enhancement, and system robustness in weak network environments requiring further optimization. Future research can conduct in-depth exploration in directions such as expanding exercise type coverage, introducing personalized transfer learning strategies, integrating differential privacy mechanisms for data-level privacy protection at the analytics stage, and exploring privacy-preserving training mechanisms under federated learning frameworks, with the aim of promoting intelligent sports monitoring technology toward more universal, secure, and efficient directions.

Author contributions

S.-Z.Z. conceived the study, designed the experimental protocols, and curated the self-built sports activity dataset. H.-Z.Y. proposed the core CRNN spatiotemporal analysis architecture with the self-attention mechanism, developed the model compression and lightweight deployment strategy, and drafted the main manuscript text. Y.G. designed and implemented the secure transmission module, performed the system integration and performance evaluation, and conducted the formal data analysis. All authors contributed to the interpretation of results, reviewed the manuscript, and approved the final version.

Data availability

The data supporting the findings of this study are openly available. The UCI-HAR Dataset is available at https://archive.ics.uci.edu/dataset/240/human+activity+recognition+using+smartphones, the PAMAP2 Dataset at https://archive.ics.uci.edu/dataset/231/pamap2+physical+activity+monitoring, and the WISDM Dataset at https://www.cis.fordham.edu/wisdm/dataset.php. The self-built Sports Activity Dataset generated during this study has been deposited in the Science Data Bank repository and is accessible at https://www.scidb.cn/s/MvmE7z.

Declarations

Competing interests

The authors declare no competing interests.