Predicting physical activity types in children aged 3–6 from video data using computer vision

Abstract

Objective

Assessing the types of physical activity (PA) in young children is crucial for exploring their relationship with health. Although existing machine learning methods have made certain progress, there are still deficiencies in non-contact, real-time dynamic monitoring and recognition accuracy. This study aims to utilize computer vision technology to construct non-contact, real-time dynamic, and high-precision recognition models for the types of physical activities of children aged 3–6 years in different age groups and overall, to provide scientific and practical tools for monitoring children’s daily physical activities and preventing health problems such as obesity.

Methods

The portable Drift Ghost XL camera collected video data of 11 types of physical activities from 72 children aged 3 to 6 (average age = 4.7 ± 0.9). These activities included: sitting still, sitting activity, standing still, standing activity, walking, running, crawling, jumping, cycling, climbing, and stair walking. Each type was recorded for 8 to 10 min. A total of 18,870 valid samples were obtained using the time-slice method. The PA type dataset was labeled and constructed using Labelimg software and expanded fivefold through data augmentation. Based on Yolov11 object detection technology, computer vision recognition models for PA types of children aged 3–4, 4–5, 5–6, and 3–6 were established, respectively. Their performance was evaluated using metrics such as F1 score and mAP.

Results

The model of children aged 3–6 achieved an F1 score of 96.0%. Among the age-specific models, the model of children aged 3–4 and 4–5 achieved the highest recognition rate of 95.0%, while the model of children aged 5–6 reached 94.0%. Regarding recognition accuracy, the model of children aged 3–6 achieved 98.3%. Among the age-specific models, the model of children aged 3–4 had the highest recognition rate of 98.2%, followed by 98.0% for the model of children aged 4–5 and 97.6% for the model of children aged 5–6. Yolov11 achieved recognition rates ranging from 96.9% to 99.4% for 11 PA types in the model of children aged 3–6, as follows: 98.9% (sitting still), 99.1% (sitting activity), 98.2% (standing still), 97.9% (standing activity), 97.7% (walking), 98.4% (running), 98.3% (crawling), 97.7% (jumping), 98.7% (cycling), 99.4% (climbing), and 96.9% (stair walking).

Conclusions

Computer vision technology could effectively recognize PA types in children aged 3–6 under non-wearable, real-time dynamic conditions, with performance surpassing existing wearable device-dependent machine learning techniques. Age-specific models (children aged 3–4, 4–5, 5–6) also demonstrated excellent performance, confirming the method’s effectiveness and applicability across different developmental stages of early childhood.

Physical activity (PA) is indispensable in promoting young children’s physical health [1, 2]. It serves not only as a vital means to enhance cardiopulmonary function, promote musculoskeletal development, and maintain a healthy body weight, but also as a key approach to developing fundamental motor skills and improving physical fitness. The intensity and type of PA are two critical dimensions for evaluating its health benefits [3–5]. To ensure the health benefits of PA in young children, it is necessary to adhere to evidence-based PA guidelines for macro-level guidance and develop objective and accurate monitoring methods to assess PA type and intensity [6, 7]. However, most existing studies and guidelines primarily focus on PA intensity and volume, while neglecting an equally important dimension—PA type. At present, objective monitoring methods for PA intensity based on wearable devices such as accelerometers have become increasingly mature and standardized [8, 9]. Nevertheless, accurate identification and objective monitoring of complex PA types remain a key challenge and bottleneck in current research.

Existing studies indicates that different PA types may have differentiated positive impacts on the physical health of young children. For example, walking improves cardiorespiratory endurance [10], while running helps reduce body mass index and mitigate childhood obesity [11]. Crawling enhances physical fitness and plays a significant role in developing cognitive and coordination abilities [12]. Furthermore, as a complex full-body exercise, climbing improves overall motor skills by coordinating upper and lower limb muscle groups [13]. Studies have also confirmed that certain PA types show significant potential for rehabilitation and health improvement in specific populations [14, 15]. Therefore, accurately identifying PA types and establishing objective monitoring methods for young children’s PA are crucial for gaining deeper insights into the differential effects of various activities on children’s health.

While the importance of assessing PA types in young children is self-evident, the development of objective monitoring methods has significantly lagged [16, 17]. Traditional approaches such as questionnaires and behavioral observations are characterized by high subjectivity and labor demands. With the rapid advancement of artificial intelligence technologies like machine learning and improved accessibility to raw accelerometer data, wearable device-based machine learning methods have become mainstream tools for objective monitoring PA types [18–24].

In recent years, five research teams have utilized accelerometers embedded in wearable devices and applied machine learning methods such as Random Forests (RF), Decision Trees (DT), and Support Vector Machines (SVM) to develop monitoring approaches for young children’s PA types [18–22]. The highest recognition rates have consistently exceeded 80%, with Brønd et al. [19] achieving over 90% accuracy using a DT approach. These previous research outcomes demonstrate the feasibility, effectiveness, and considerable potential of applying artificial intelligence technologies for identifying PA types in young children. However, this approach presents three limitations: Firstly, due to inherent constraints of the monitoring method, the devices require direct skin contact, which may cause discomfort in young children and lead to non-compliance issues such as forgetting to wear or refusal to wear the devices, thereby compromising data continuity and authenticity [25, 26]. Secondly, the recognized PA type categories remain narrow and lack sufficient differentiation. Existing models are primarily optimized for basic locomotor PA like walking, running, and jumping, while neglecting more complex action categories, such as distinguishing between “sitting still” and “sitting activity” [18–22]. Finally, current methods generally overlook the developmental characteristics of young children. Children’s movement patterns evolve rapidly with age, generic models struggle to accurately adapt to the specific motor characteristics of each age group within the aged 3–6 range, leading to fluctuations in recognition accuracy across different age stages.

Given the limitations of existing methods, the accurate identification and long-term monitoring of PA types in young children remain significant challenges. In recent years, the rapid advancement of computer vision technology has provided a novel solution to this problem. This technology can directly perceive and interpret human movement through video data, featuring non-contact operation and real-time dynamic monitoring capabilities [27, 28], enabling automated, high-precision recognition of complex PA types. Recent studies have introduced computer vision and other deep learning techniques into the field of PA recognition for adults, achieving remarkable success [29–34].

However, in the field of PA types recognition for children aged 3–6, there are still a series of unique and challenging issues. These challenges mainly stem from the fundamental differences in physiology, behavior, and environment between this group and adults. Firstly, the PA patterns of children are not yet established, showing higher randomness, variability, and instability. Their body proportions and motor control abilities differ significantly from those of adults, resulting in standard posture estimation models or action recognition models trained based on adult data often performing poorly on children. Secondly, in video data, the small body size of children leads to less target pixel information, coupled with frequent, irregular, and rapid body position changes, as well as the common mutual occlusion in activity behaviors, which places higher demands on the robustness of detection and tracking algorithms. Moreover, the kindergarten environment is complex with variable lighting, further increasing the difficulty of recognition. Thirdly, for the assessment of children’s PA health, accurately distinguishing fine-grained activities that are visually highly similar but differ in energy consumption (such as “sitting still” and “sitting activity”, “standing still” and “standing activity”) is crucial. This subtle difference places extremely high requirements on the model’s representation learning ability, as well as the latest detection architectures like Yolov11, which possess strong multi-scale feature fusion and fine-grained distinction capabilities in complex scenes. Building on this foundation, our study pioneers the application of computer vision technology to monitor PA types in young children, aiming to develop a high-accuracy, non-contact, real-time dynamic recognition model specifically for children aged 3–6. This work provides an innovative and effective scientific tool for refined monitoring of children’s PA types and developing targeted health promotion strategies.

Subjects and methods

Object of study

This study focuses on developing a PA type recognition model, with participants children aged 3–6. Based on previous research [20–22], the effect size was calculated to range from 0.51 to 2.03. Accordingly, G*Power software was used for sample size estimation, in which the F-test was selected, with an effect size = 1, α = 0.05, and power = 0.95. The calculated minimum sample size was 44 participants. To account for potential data loss during experiments, the sample size was increased by 30% relative to previous calculation, resulting in a minimum of 57 participants. According to the instance sample size requirements for deep learning [35], each target PA type category should contain at least 1,000 instances. This study involves 11 target categories, requiring 1,000 × 11 = 11,000 instances. Given an average of 10 instances per image, the minimum number of images required is 11,000 instances / 10 instances per image = 1,100 images. Therefore, a minimum of 1,100 image samples was necessary for model construction.

Two public kindergartens in Nanchang City, Jiangxi Province, China were selected, totaling 72 children aged 3–6 were enrolled as the subjects (4.7 ± 0.9 years old), which met the subjects sample size requirements of this study. Among them, 60 participants were involved in model construction (train/val./test), while the other 12 served as an entirely independent external validation set for evaluating the model’s generalization ability. A total of 18,870 sample image data were collected, meeting the image sample size requirement of 1,100 images. The study protocol was approved by the Medical Research Ethics Committee of The Second Affiliated Hospital of Nanchang University (approval number: 研临审 [2020] No.125). The specific contents include the project name, research category, task source, research content, etc.

Inclusion Criteria: (1) Children aged 3–6; (2) Children in good health, without physical developmental or motor impairments; (3) Parental/guardian signed informed consent obtained, covering the study’s purpose and nature, voluntary participation, procedures, potential risks and benefits, and privacy protection. Exclusion Criteria: (1) Children with overweight, obesity, or growth disorders; (2) Children those who did not complete anthropometric measurements.

Experimental protocol

Physical activity type selection

This study systematically identified PA types from the Children’s Activity Rating Scale [36] and the Observational System for Recording Physical Activity in Children-Preschool Version [37], categorizing activities into five broad classes based on movement characteristics, body position, and movement speed. (1) Stationary with no movement, e.g., lying, sitting, standing, or squatting; (2) Stationary with movement, e.g., sitting, standing, or squatting combined with limb or trunk activity; (3) Locomotion activities including low-speed, moderate-speed, and high-speed, e.g., walking, running, jumping, crawling, climbing, stair walking, rolling, cycling, etc.

Based on this framework, 10 specific PA types were initially identified: stationary without movement, stationary with movement, walking, running, jumping, crawling, climbing, stair walking, rolling, and cycling. To achieve a more refined classification, the research team further subdivided the first two categories through discussion: “stationary without movement” was divided into “sitting still” and “standing still”, while “stationary with movement” was split into “sitting activity” and “standing activity”. To validate the practical applicability of the classification system, researchers conducted on-site observations of children across all grade levels in kindergarten settings during the hours of 8:00 to 16:00. By comparing observed behaviors with the 12 predefined PA types, it was found that the “rolling” behavior occurred with very low frequency and was prone to confusion with “crawling”. Consequently, 11 representative PA types for young children were established: sitting still, sitting activity, standing still, standing activity, walking, running, jumping, crawling, climbing, stair walking, and cycling. Detailed descriptions of each PA type are provided in Table 1.

Table 1.

Classification of PA types and activity description for young children

	Intensity	Type	Motion description	The activity of collection	Detailed description of the activity trials
1	Resting	Sitting Still	No movement: sitting/lying/leaning/prostrating/kneeling for more than 5 s, with only slight movements of the head, fingers, and feet, while the trunk remains still.	Watch TV	Sit/lie/stand on a chair/the floor/ a mat, or stand and watch TV
				Read / Repeat reading	Sit on the floor/chair/mat and read a book
				Play mobile phone / tablet / computer	Sit/learn/lay on a chair/floor/blanket, and play with a mobile phone/tablet
2	Low intensity	Sitting Activity	No movement: sitting with the trunk or limbs in motion (swinging, rocking, bending, twisting, kicking, hitting, throwing, digging sand, performing a song)	Play while sitting or squatting.	Sit or squatting to play with toys / building blocks / jigsaw puzzles / dolls / mud / sand
3	Resting/Low intensity	Standing Still	No displacement: standing still for more than 5 s with slight movement of limbs or trunk.	Stand and watch TV	Stand in front of the TV and watch it
4	Low intensity/ Moderate intensity	Standing Activity	No displacement: standing with trunk and limbs moving freely	Play while standing	Stand by the table and play with toys / building blocks / jigsaw puzzles / dolls
				Dance standing up	Dance along with the dance teacher on the screen
5	Low intensity	Walking	Movement	Take a walk	Take a walk in the corridor or on the playground
6	Moderate-to-Vigorous intensity	Running	Movement	Run	Under the guidance of the teacher, run around the playground at their own pace
7	Moderate-to-Vigorous intensity	Jumping	Movement/No movement	Jump in place	Continuously jumps in place
				Hop on one foot	Continuously single-foot jumps
				Skip rope	Various forms of rope skipping could be performed either in place or while moving.
8	Moderate-to-Vigorous intensity	Crawling	Movement	Crawl	Crawl on the ground or the surface of an object using hands and knees in different ways of contact
9	Moderate-to-Vigorous intensity	Climbing	Movement	Climb	Climb up or down by alternately grasping the rope or ladder with both arms
10	Moderate-to-Vigorous intensity	Stair Walking	Movement	Take the stairs	Walk on the stairs
11	Moderate-to-Vigorous intensity	Cycling	Movement	Ride a bicycle	Ride a bike on the playground

Data collection process

This study utilized the portable camera Drift Ghost XL (Foream Network Technology Ltd., China) to collect video data on young children’s PA types. All data collectors received standardized training to ensure proficiency in equipment operation procedures and movement demonstration standards. Researchers conducted on-site assessments of kindergarten indoor environments (e.g., activity rooms, corridors) and outdoor environments (e.g., playgrounds, recreational areas). Based on these assessments, optimal fixed camera positions were planned to ensure comprehensive coverage of the target children’s main activity areas while minimizing obstructions.

Under the guidance of the data collectors, 60 participating (4.7 ± 0.9 years old) completed semi-structured PA trials covering 11 PA types within two weeks. Data were collected separately for each age group (children aged 3–4, 4–5, 5–6). The duration of each participant’s test for each activity type was 8 to 10 min. Additionally, 12 children aged 3 to 6 years old were selected as external test samples for data collection. Detailed descriptions of the activity trials are provided in Table 1, and sample collection illustrations are shown in Fig. 1. All data collection procedures complied with ethical review requirements, including obtaining signed informed consent from guardians and implementing facial blurring to protect the children’s privacy.

Fig. 1.

Example images of collecting PA types for paticipants (Standing still and Cycling)

Statistical analysis

Data preprocessing

After excluding invalid data, 18,870 valid image samples were ultimately retained. Image data were annotated using LabelImg software, with the minimum bounding rectangle centered on the children’s torso to label the 11 PA types along with their corresponding bounding boxes. The annotation process was independently performed by two researchers to ensure reliability. In cases of disagreements, the results were reviewed by to the research team for discussion and final determination. Upon completion of annotation, XML files were generated containing the image names, dimensions, and location information of the 11 different PA types within each image.

Data augmentation

To enhance the model’s generalization capability and effectively prevent overfitting during model training, data augmentation strategies were implemented after image annotation to expand the original dataset [38]. The specific methods and parameters are as follows: using a 50% probability of vertical/horizontal flipping, brightness enhancement of 1.2–1.5 times, Gaussian blur with a sigma range of 0–3.0, random translation within 15 pixels, scaling from 0.8 to 0.95 times, and an affine transformation with rotation ranging from − 30° to 30° to simultaneously enhance the images and bounding boxes of the object detection dataset. These techniques augmented the original images by a factor of five, ultimately expanding the dataset to 94,350 image samples. The study involved four distinct datasets: children aged 3–4, children aged 4–5,children aged 5–6, children aged 3–6. The dataset composition was as follows: 30,900 images for children aged 3–4, 32,525 images for children aged 4–5, 30,925 images for children aged 5–6, and 94,350 images for children aged 3–6. The distribution of each label category is presented in Table 2.

Table 2.

Distribution of augmented dataset samples

PA types	Number of labels for children aged 3–4	Number of labels for children aged 4–5	Number of labels for children aged 5–6	Number of labels for children aged 3–6
Sitting still	18,510	18,620	18,320	55,450
Sitting activity	18,360	17,850	18,505	54,715
Standing still	18,290	18,065	17,720	54,075
Standing activity	18,675	18,430	18,340	55,445
Walking	18,180	18,490	18,365	55,035
Runing	17,450	17,395	17,740	52,585
Crawling	18,050	17,725	17,855	53,630
Jumping	17,660	17,505	17,310	52,475
Climbing	18,430	18,400	18,070	54,900
Stair walking	18,665	18,045	17,720	54,430
Cycling	17,575	17,475	17,435	52,485
Total	199,845	198,000	197,380	595,225

Subsequently, all datasets were divided into the model training set, validation set, and test set. Since our video data was collected by having 10 children appear in the same frame (as shown in Fig. 1), the traditional individual-based partitioning method (allocating all images of a specific child to a separate set) was not feasible. To maximize the generalization ability of the model, multi-scenario shooting was adopted during the shooting process. Specifically, all images and labels were randomly shuffled in order and then distributed to the training set, validation set, and test set in a ratio of 6:2:2. The detailed segmentation results can be found in Table 3.

Table 3.

Distribution of augmented dataset samples

Group	Number of Train	Number of Val.	Number of Test	Total
Children aged 3–4	19,515	6505	6505	32,525
Children aged 4–5	18,540	6180	6180	30,900
Children aged 5–6	18,555	6185	6185	30,925
Children aged 3–6	56,610	18,870	18,870	94,350

Model training

This study employed Yolov11 [39], the latest object detection algorithm in the Yolo series, which effectively integrates accuracy, speed, and efficiency. Fig. 2 illustrates the network architecture of Yolov11, which consists of four components: the input end, backbone network, neck network, and prediction network. At the input end, images first undergo adaptive image scaling, normalization, and data augmentation, where they are resized to fit the model’s input dimensions. Within the backbone network, images are initially processed through alternating stacked convolutional layers, C3k2 layers, and SPPF modules for feature extraction. The features are then processed by the C2PSA module and passed to the neck network. The features are further processed in the neck network using C3k2 modules combined with upsampling and downsampling operations. Feature fusion is subsequently performed through the Path Aggregation Network to generate new feature maps. Finally, the new feature maps output by the neck network are fed into the prediction network, where the detection head produces the final results.

Fig. 2.

Structure diagram of Yolov11

Input

The input module primarily preprocesses image data into the format and size expected by the model before feeding it into the network. Yolov11 supports multi-scale input, a key feature that enables the model to adapt to image data with varying resolutions, thereby significantly enhancing its flexibility. Additionally, the input module performs image normalization to improve the model’s convergence speed and overall performance.

Backbone

The backbone network primarily serves to extract features from input image data. It progressively captures low-level, mid-level, and high-level features from images through a series of convolutional layers, pooling layers, and other network components. Compared to earlier versions, Yolov11 introduces the C3k2 module to replace the previous C2f module, thereby improving computational efficiency. By incorporating efficient components such as the C3k2 block and optimized architectural design, Yolov11’s backbone network achieves effective feature extraction and strong generalization capabilities.

Neck

The neck network, connected to the backbone network, integrates top-level and bottom-level features through a combined Feature Pyramid Networks and Path Aggregation Network. It processes these fused features to generate feature maps of different scales, which are subsequently fed into the detection head for predicting objects of three distinct sizes.

Prediction

The prediction network consists of three detection heads at different scales that process the multi-scale feature maps output by the neck network. The final output includes both bounding box coordinates and category information for the detected objects.

Experimental environment and evaluation indicators

Experimental environment

The experimental framework developed in this study is based on PyTorch, with Python as the programming language. The specific configuration of the experimental environment is detailed in Table 4.

Table 4.

Experimental environment

Name	Related Configuration
Graphics Card	NVIDIA GeForce RTX 4090D
Processor	Intel Core i9-13900 K CPU @ 5.40 GHz
Memory	24GB
System	Windows11

This study adopts the Yolov11s architecture, with specific parameters set as follows: epochs 100, batch size 4, and image size 640 × 640. The optimizer is SGD, and lr0 is set to 0.01. The parameter selection mainly follows the officially recommended configuration of Yolov11 and is manually fine-tuned based on the performance of the validation set to ensure stable convergence during training.

Model evaluation metrics

In the field of object detection, the evaluation of model performance typically involves three types of metrics: Precision (P), Recall (R), and the comprehensive mean Average Precision (mAP), which integrates both of the former.

1. P reflects the proportion of correctly identified positive instances among all retrieved results, representing the ratio of true positives to the total number of samples predicted as positive. Its calculation formula is shown in Eq (1).

   1

2. R represents the proportion of correctly detected positive samples relative to the total number of actual positive samples in the ground truth. Its calculation formula is shown in Eq (2).

   2

3. mAP is a key metric in multi-class object detection. A Precision-Recall (PR) curve can be plotted for each category, with recall as the horizontal axis and the maximum precision at each recall level as the vertical axis. Average Precision (AP) refers to the area under the PR curve for a specific category across all predicted images, where AP is calculated individually for each category. The calculation formula is shown in Eq (3). mAP is the average of AP values across multiple categories, and its calculation formula is shown in Eq (4).

   3
   4

To facilitate an intuitive understanding of the evaluation metrics, Table 5 illustrates the definitions of TP (True Positive), FP (False Positive), TN (True Negative), and FN (False Negative), where 1 represents a positive sample and 0 represents a negative sample. TP indicates cases where the predicted and actual results are positive samples, meaning the PA type is correctly predicted. FP refers to instances where the actual result is a negative sample but is incorrectly detected as positive, representing misidentified PA types in this study. FN denotes cases where the actual result is a positive sample but is predicted as negative. TN signifies instances where the actual result is a negative sample but is predicted as positive. Both FN and FP represent incorrectly identified PA types. In this experiment, a higher mAP value indicates better detection performance of the algorithm and superior model recognition capability.

Table 5.

TP, FP, TN, FN

Actual Results Prediction Results	1	0
1	TP	FP
0	FN	TN

In summary, this study employs Precision (P), Recall (R), Average Precision (AP), and mean Average Precision (mAP) as the evaluation metrics for the experimental section.

Study results

Results of data augmentation

Data augmentation demonstrated the most notable effect on the integrated dataset covering the groups of children aged 3–6. Table 6 illustrates the impact of data augmentation techniques on each dataset. The F1-score of the model children aged 3–6 increased by 2.0% points from 94.0% to 96.0%, while the mAP@0.5 also improved by 1.0% point from 97.3% to 98.3%. Data augmentation consistently enhanced the detection accuracy (mAP@0.5) across all age-specific datasets. Among them, the dataset of children aged 4–5 showed the most significant improvement in mAP@0.5, rising by 1.1% points from 96.9% to 98.0%. The dataset of children aged 3–4 and 5–6 also improved by 0.5 and 0.9% points, respectively. In terms of F1-score, the dataset of children aged 4–5 and 5–6 each increased by 1.0% point. It is worth noting that for the dataset of children aged 3–4, the F1-score remained stable at 95.0% before and after augmentation, while mAP@0.5 still improved, indicating that the model achieved higher localization and recognition confidence for positive samples while maintaining classification balance.

Table 6.

Data augmentation results

Dataset	Data Augmentation	F1 score /%	mAP@0.5/%
Children aged 3–4	Original	95.0	97.7
	Data Augmentation	95.0	98.2
Children aged 4–5	Original	94.0	96.9
	Data Augmentation	95.0	98.0
Children aged 5–6	Original	93.0	96.7
	Data Augmentation	94.0	97.6
Children aged 3–6	Original	94.0	97.3
	Data Augmentation	96.0	98.3

Results of overall recognition

After 100 training epochs, the model completed its learning process for recognizing young children’s PA types on the training set, and final outputs were generated from the test set. The training process is shown in Fig. 3. As shown in Table 7, the model trained on the integrated dataset of children aged 3–6 achieved the best overall performance, attaining the highest values among all models in F1-score (96.0%), mAP@0.5 (98.3%), and Recall (95.9%). The recognition results of the three age-specific models were very close. Among them, the models for children aged 3–4 and 4–5 reached an F1-score of 95.0%, while the model of children aged 5–6 achieved 94.0%, demonstrating exceptionally high performance. For mAP@0.5, the three models achieved 98.3%, 98.0%, and 97.6%, respectively, thereby confirming the model’s capability to achieve precise recognition across different age groups. As shown in Fig. 4, compared with the research on machine learning technology based on wearable sensors, this study is the first to apply computer vision to the 3-6-year-old age group and achieved competitively high recognition accuracy (mAP@0.5: 98.3%) on a new dataset that contains more detailed information about 11 PA types.

Fig. 3.

Changes in mAP during training process

Table 7.

Overall identification result

Dataset	F1 score /%	mAP@0.5/%	P/%	R/%
Children aged 3–4	95.0	98.2	95.3	95.4
Children aged 4–5	95.0	98.0	95.1	95.3
Children aged 5–6	94.0	97.6	93.9	94.4
Children aged 3–6	96.0	98.3	95.4	95.9

Fig. 4.

Recognition results comparison with other machine learning studies

Results of recognition in various types

The AP results (Table 8) demonstrate that the integrated model of children aged 3–6 achieved excellent recognition performance across all PA types, with AP values ranging from 96.9% to 99.4%. Among these PA types, climbing achieved the highest recognition accuracy at 99.4%, followed by sitting activity (99.1%), sitting still (98.9%), and cycling (98.7%), each with AP values exceeding 98.5%. Even stair walking, which had the lowest AP, still reached 96.9%. The model also exhibited strong capability in distinguishing visually similar PA types. For instance, in the model for children aged 3–6, both “sitting still” (98.9%) and “sitting activity” (99.1%), as well as “standing still” (98.2%) and “standing activity” (97.9%), achieved exceptionally high AP values.

Table 8.

Identification results of various types (mAP@0.5%)

Dataset	Sitting Still	Sitting Activity	Standing Still	Standing Activity	Walking	Runing	Crawling	Jumping	Climbing	Stair walking	Cycl ing
Children aged of 3–4	99.1	99.3	97.5	97.8	98.2	98.9	98.5	96.7	99.2	96.5	98.5
Children aged of 4–5	98.5	99.0	98.3	97.3	97.5	97.7	96.7	97.0	99.4	97.1	98.9
Children aged of 5–6	98.0	98.7	98.1	97.6	96.6	96.6	98.1	96.8	99.5	95.9	97.6
Children aged of 3–6	98.9	99.1	98.2	97.9	97.7	98.4	97.6	97.8	99.4	96.9	98.7

For the model of children aged 3–4, 4–5, and 5–6, the AP values for all activity types ranged from 96.5% to 99.3%, 96.7% to 99.4%, and 96.6% to 99.5%, respectively. Cross-age comparison of the age-specific models showed that climbing, sitting activity, and sitting still consistently maintained exceptionally high recognition accuracy (AP > 98.0%) across all age groups. For activities such as stair walking and jumping, slight fluctuations in AP values were observed across different age groups; yet all AP values remained above 95.9%, indicating consistently high performance.

Confusion matrix results indicate that except for the model of children aged 5–6, which had the highest error recognition rate for “walking”, the other three models all showed that the error recognition rate for “standing activity” was most commonly misidentified as “standing still,”, “jumping”, and “walking” (Fig. 5). Specifically, the model of children aged 3–6 shown the highest error recognition rate for “standing activity” : 3.0% of instances were misidentified as “standing still,” 1.0% as “walking,” and another 1.0% as “jumping.” For the model of children aged 3–4, “standing activity” also showed the highest error recognition rate: 5.0% of instances were misidentified as “standing still,” 2.0% as “jumping,” 1.0% as “walking,” and an additional 1.0% as “stair walking.” For the model of children aged 4–5, “standing activity” similarly exhibited the highest error recognition rate: 4.0% of instances were misidentified as “standing still,” 2.0% as “jumping,” and another 1.0% as “walking.” For the model of children aged 5–6, “walking” again showed the highest error recognition rate: 3.0% of the samples were not detected, 2.0% of instances were misidentified as “jumping,” 1.0% as “running,” 1.0% as “standing activity,” and an additional 1.0% as “stair walking.”

Fig. 5.

Confusion matrix normalized (a 3-6-year-old model; b 3-4-year-old model; c 4-5-year-old model; d 5-6-year-old model)

Ablation experiment

To verify the crucial effect of SPPF on the model’s recognition performance, this study conducted an ablation experiment by replacing the SPPF module with a lower-order SPP module on the dataset of children aged 3–6. After the replacement, the mAP@0.5 of the model decreased by 0.1. As shown in Table 9, the mAP@0.5 of the model on the comprehensive dataset of children aged 3–6 decreased from a baseline of 98.3% to 98.2%, with a decrease of 0.1% points. This indicates that SPPF provides stable and small performance gains for this task through its more efficient multi-scale pooling and feature fusion. The optimization might lie in better aggregating context information from different receptive fields, which is helpful for classifying PA types such as “sitting still” and “jumping”.

Table 9.

Ablation experiment

	Sitting Still	Sitting Activity	Standing Still	Standing Activity	Walking	Runing	Crawling	Jumping	Climbing	Stair walking	Cycling	mAP
SPP	98.9	99.1	98.1	97.9	97.7	98.5	97.7	97.7	99.4	97.0	98.7	98.2
SPPF	98.9	99.1	98.2	97.9	97.7	98.4	97.6	97.8	99.4	96.9	98.7	98.3

Model generalization ability verification

To evaluate the generalization ability of the constructed model, we further tested the best model on the training set that included data from children aged 3 to 6 years old. We also tested its performance in identifying the PA types of 12 children aged 3–6 who were not involved in the training process and compared the results with the manually annotated values. The results showed that the recognition accuracy of the model in the external validation set reached 95.0%. This proves that the model has good generalization ability.

Discussions

Main findings

Based on computer vision technology, this study developed a set of high-precision recognition models for PA types in children aged 3–6, including age-specific sub-models. Results demonstrate that the developed models exhibit outstanding recognition accuracy and robust generalization capability. Specifically, the integrated model of children aged 3–6 achieved an F1-score of 96.0% and an mAP@0.5 of 98.3%. The age-specific models also showed highly stable performance: the models for children aged 3–4 and 4–5 achieved F1-scores of 95.0%, while the model of children aged 5–6 reached 94.0%. Furthermore, the models accomplished high-precision recognition for all 11 PA types, with the highest accuracy observed for “climbing” (99.4%) and the relatively lowest—yet still substantial—accuracy for “stair walking” (96.9%). Confusion matrix analysis revealed that primary error recognition occurred between activities with similar movement patterns, such as mutual confusion between “standing activity” and “standing still,” attributed to the inherent visual similarity between these PA types.

Comparison with similar studies

In recent years, several studies have employed machine learning methods (e.g., RF, SVM, DT) combined with accelerometers to identify young children’s PA types. Compared to these approaches, our study demonstrates significant advantages in recognition accuracy, the granularity of identifiable PA types, and practical applicability.

Compared with previous similar studies, this research demonstrates significant advantages in recognizing PA types for young children. Ahmadi et al. [18] used an RF algorithm to classify five PA types in children aged 3–5, reporting a maximum F1-score of 86.4%. Brønd et al. [19] proposed a DT-based method in 2020, achieving an accuracy of 91.1% for five PA types, which was considered a high level at the time. In contrast, the integrated model of children aged 3–6 developed in this study achieved an F1-score of 96.0% and an mAP@0.5 of 98.3%, representing improvements of 9.6 and 7.1% points, respectively, over the aforementioned studies. This precise identification capability enables us to reliably distinguish between activities that are similar in energy consumption but different in PA types (such as “standing still” and “standing activity”), effectively addressing the blind spots of device-based methods in the analysis of similar PA types, and providing solid technical support for an in-depth understanding of the detailed characteristics of children’s behavioral development.

Regarding recognition category scope, existing accelerometer-based studies are limited by the feature representation capacity of sensor data, typically identifying only five basic activity types (sedentary, walking, running, light activities games, and moderate-to-vigorous activities games). Only one study has further subdivided PA types into six categories (sleep, rest, quiet play, low active play, moderately active play, and very active play) [22]. In contrast, our study successfully recognized 11 PA types with high accuracy, significantly enhancing the comprehensiveness and granularity in describing young children’s daily activity patterns. This improvement in classification granularity enables us to depict the behaviors of young children’s daily activities more comprehensively, laying a solid data foundation for future exploration of the specific associations between different behavioral patterns and health benefits.

Compared to accelerometer-based studies, which are often limited by challenges such as device wearability and data collection continuity—particularly in young children who frequently exhibit poor compliance and unstable data quality—the computer vision approach adopted in this study requires no wearable sensors and achieves high-precision PA recognition solely through video data. It demonstrates notable potential for non-contact and real-time dynamic monitoring.

Possible reasons for the results of this study

Technical advantages: from feature engineering to end-to-end learning

Deep learning-based computer vision frameworks demonstrate significant technical advantages compared to traditional machine learning approaches. Most existing studies follow the technical route of feature engineering combined with conventional classifiers. For instance, Hagenbuchner et al. [20] utilized manually extracted time-domain features with SVM, while Ahmadi et al. [18] relied on feature-selected metrics input into RF models. Such methods are constrained by the representational capacity of handcrafted features, making it challenging to capture subtle postural changes in children’s PA. This limitation also leads to potentially significant errors in recognizing similar PA types. For example, the recognition accuracy for walking and running in young children was only 61.1% and 68.8%, respectively [21]. In contrast, deep learning models could automatically learn multi-level feature representations directly from raw video data, including local joint movements, overall body posture, and rich contextual information from environmental interactions. This end-to-end learning approach enables a better understanding of the spatiotemporal characteristics of children’s PA, thereby demonstrating significant advantages in complex PA type recognition tasks.

Monitoring methodology: paradigm shift - from device dependency to visual perception

Compared to similar studies relying on wearable devices, the computer vision model developed in this study enables real-time dynamic monitoring without physical contact. Existing studies commonly used accelerometers and other wearable devices, whose recognition accuracy is primarily constrained by sensor placement and the number of devices configured. For PA recognition in young children, most studies adopt three wearing configurations: wrist-worn, hip-worn, or a combination of both [18–22]. Since acceleration signals vary significantly across different wearing positions, features extracted for the same PA type differ by location, leading to fluctuations in recognition performance. For example, when recognizing running and walking, hip-worn sensors yield accuracy rates of 85.1% and 80.2%, respectively, whereas wrist-worn sensors result in lower rates of 80.6% and 65.5%. Even with combined hip and wrist wearing configurations, accuracy rates only reached 88.1% and 79.3% [21]. In contrast, this study’s non-contact computer vision technology utilizes multi-angle video synchronization to comprehensively capture children’s PA movement trajectories, thereby effectively avoiding data loss caused by sensor misplacement or detachment.

Limitations

This study adhered to rigorous scientific research methodology; however, the following limitations should be acknowledged. Firstly, regarding data collection, although the sample size satisfies statistical requirements, the dataset was primarily obtained from controlled laboratory settings and lacks ecologically valid environments, such as outdoor spaces and home settings. The experiments of the application scenarios will be taken as the core part of model evaluation to directly quantify and enhance the environmental adaptability of the model. Future research should expand the scope of data collection to incorporate more diverse natural environment datasets. Secondly, the current model still demonstrates some degree of confusion between certain PA types with similar visual characteristics (e.g., “standing activity” and “standing still”). Subsequent research could improve the model’s ability to differentiate subtle movement variations by incorporating more refined feature extraction techniques, such as spatiotemporal modeling of skeletal keypoints and attention-based mechanisms.

Model practical deployment considerations

The computer vision model developed in this study was applied in the kindergarten environment to make progress in the following aspects: Firstly, the actual deployment of the model needs to adapt to the differences between the indoor and outdoor environments of kindergartens and optimize the real-time performance of edge devices. Through the design of a lightweight application interface, it can meet the needs of various users (such as teachers and children’s parents). Secondly, in terms of privacy and ethics, it is necessary to strictly follow relevant regulations to implement comprehensive data control and improve the dynamic informed consent mechanism for parents as well as the data withdrawal mechanism. Finally, in terms of technical paths, the human pose estimation (HPE) method has theoretical advantages in distinguishing fine-grained actions by extracting joint feature points [34]. However, due to irregular movements of children and occlusion, the stability of joint point detection is insufficient, and computational complexity is relatively high. This study adopts Yolov11, which does not need to extract joint points separately and is superior in multi-target detection, real-time performance, and adaptability to occlusion; thus making it suitable for complex scenarios in kindergartens. In the future, both advantages can be integrated so that HPE joint point features can enhance Yolov11’s fine-grained recognition ability further improving accuracy when distinguishing similar activity types.

Conclusions

The computer vision-based model developed in this study demonstrates excellent performance in recognizing PA types among children aged 3–6, effectively identifying 11 distinct PA categories while outperforming conventional machine learning methods in overall recognition efficacy. By adopting an age-specific modeling strategy, this approach successfully addresses the variations in motor characteristics across different developmental stages, with dedicated models for children aged 3–4, 4–5, and 5–6 each achieving recognition accuracy exceeding 97.0%. This work not only enhances the accuracy and adaptability of identifying children’s PA types but also lays the algorithmic foundation for building a practical, non-contact intelligent monitoring system for children’s physical activities. In the future, by integrating multi-target tracking technology and developing user-friendly application interfaces, this model can be incorporated into the intelligent management system of kindergartens, providing objective and continuous data support and intervention methods to promote the healthy development of children.

Abbreviations

DT

Decision Trees

PA

Physical Activity

RF

Random Forests

SVM

Support Vector Machines

Authors’ contributions

Xinxin Ma conducted literature research, designed the framework of the thesis, collected and analyzed data, and wrote and revised the thesis; Guanggao Zhao proposed the topic of the thesis, reviewed, guided and revised the thesis; Liqiang Su, Jiemei Fu, Shunli Sun, Zihao He, Ting Huang, Xuewen Hu, Yurong Li reviewed, guided and revised the thesis; Kun Jiao, Xiaohu Liu, XiaoyZhang, ZhixiDong, Jiachang Wang, Youdi Wei, Jiayi He collected, sorted out and verified the data.

Funding

This research was funded by The National Social Science Fund of China, grant. number 25BTY097.

Data availability

The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

The experimental protocol was established according to the ethical guidelines of the Helsinki Declaration and was approved by the Human Ethics Committee of The Second Affiliated Hospital of Nanchang University (approval number: 研临审 [2020] No. 125). Written informed consent was obtained from individuals or guardians of participants.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.