Table 1.

Previous works for human activity recognition and localization.

Method	Algorithm	High Points	Limitations	Proposed Solution
Hsu et al. [31]	The sensors measured the accelerations and angular velocities of the human body and transmitted them wirelessly to a computer. The computer then applied a series of steps to process the signals and classified them into different activities using a nonparametric weighted feature extraction algorithm and a principal component analysis method.	Two wearable inertial sensors were used while one was placed on the on the wrist and the other one on the ankle. Sensor data was wirelessly transmitted to the system.	System used only two sensors, which might not capture the full range of human motions and postures. It also requires a wireless connection between the sensors and the computer, which may be unreliable or unavailable in some environments.	More sensors are used to cover different parts of the human body, such as the torso, the backpack, the hand, and the pocket [32]. This allows us to capture more information and the diversity of human motions and postures. Moreover, smartphone embedded sensors have been used so as to recognize the human’s activities and locations, without relying on a wireless connection.
A-Basset et al. [33]	The system is based on heterogeneous human activity recognition (HHAR) and interprets the HHAR as an image classification problem. Their System encodes sensory data into a three-channel (RGB) picture representation and passes it through the system for the activity classification.	System Generated RGB images from HHAR data. Multiscale heirarchical feature extraction. Channel-wise attention unit.	The system was trained on small datasets that makes the generalizability of the system uncertain. Moreover, the computational and space complexity of the system is unclear that makes the scalability of the system uncertain.	Diverse and large datasets were utilized in training of the system that enhances the generalizability of the proposed system. As the system is trained on large datasets, it can handle bigger datasets while maintaining its computational complexity [34].
Konak et al. [35]	The system evaluates the performance of several sets of features taken from accelerometer readings and divides them into three classes: features related to motion, features related to orientation, and features related to rotation. Motion, orientation, and rotational information are used individually and in combination by the system to assess recognition performance. The analysis employs a number of categorization techniques, including decision trees, naive Bayes, and random forests.	Accelerometer based activity recognition. Categorization of the features into rotation, orientation, and motion related features.	Dataset used in the system was collected with the contribution of 10 subjects only that makes the generalizability of the system unceratin. Secondly, They used common machine learning classifiers for the activity reocgnition while advanced models may improve the performance of the system.	The proposed model uses the Extrasensory dataset for training that provides the data of 60 subjects. System achieves state-of-the-art performance over it and proves its ability to be more generalizable. Moreover, system uses a DNDF for the classification that ia an advanced classifier that possess the properties of both machine learning and deep learning classifiers.
Chetty et al. [36]	An innovative data analytic method for intelligent human activity recognition using smartphone inertial sensors was provided. The system used machine learning classifiers such as random forests, ensemble learning, and lazy learning and was based on an information theory-based feature ranking algorithm for the best feature selection.	Smartphone sensors-based activity recognition. Feature ranking based on information theory.	Common machine learning algorithms including lazzy learning, random forest, and ensemble learning were trained on a single dataset. Single dataset might not cover all of the scenarios and can cause system to decay its performance while working in realtime scenarios.	The proposed system is trained on two benchmark datasets that cover a diverse range of activities. Specially, the Extrasensory dataset was collected in wild scenarios when there was not restrictions on the subjects contributing to the data collection. This makes the proposed system more dependable as compared to their system.
Ehatisham-ul-Haq et al. [37]	The framework introduced a novel activity-aware human context recognition method that predicted user contexts based on physical activity recognition (PAR) and learnt human activity patterns in various behavioral circumstances. The method linked fourteen various behavioral situations, including phone positions, with five daily living activities (lying, sitting, standing, walking, and running). Random Forest and other machine learning classifiers were employed in the evaluation of the suggested strategy.	System uses human activity recognition to infer the context of the activity. System also integrates other information like location of the subject and secondary activites like walking while eating, and sitting and talking etc. being performed.	The system mainly depended upon the accelerometer data for the predcition of activities, locations, and secondary activites. While for the location estimation, GPS and microphone data can be a very good addition. Moreover, the system uses a simple random forest for the classification task that can misclassify the complex activites.	The proposed system uses smartphone acclerometer, smartphone magnetometer, smartphone gyroscope, smartwatch accelerometer, smartwatch compas, smartphone GPS, and smartphone’s microphone. By encorporating diverse sensors, the system increases its robustness for the activity recognition and localization. Moreover, DNDF is more advanced than a simple random forest and more reliable in terms of its predictions.
Cao et al. [38]	The system presented an effective Group-based Context-aware classification approach GCHAR, for smartphone human activity recognition. In order to increase classification efficiency and decrease classification errors through context awareness, the system used a hierarchical group-based scheme. GCHAR used context awareness and a two-level hierarchical classification structure (inter-group and inner-group) to identify activity group transitions.	The system proposes a solution to predict human activities along with their contextual information. Inner-group and inter-group classfication approach was adopted to enhance the system’s accuracy.	Their system used tri-axial accelerometer and tri-axial gyroscope to extract the data and process it for the activity classification as well as context awareness. Addition of more sensors can make the performance of the system better.	The proposed system utilizes diverse sensors for the activity recognition and localization. This property makes the proposed system more reliable as compared to their system.
Gao et al. [39]	The research presented a system for jointly recognizing smartphone location and human activity using motion sensors and the multi-task learning (MTL) technique. To combat the detrimental impacts of smartphone orientation change on recognition, the system used a novel data preprocessing technique that included a coordinate modification based on quaternions. The joint recognition model was created to produce results for multiple tasks using a single global model, thereby lowering processing requirements and enhancing recognition efficiency.	Multi-task learning techinuqe. Joint learning of human activities and smartphone location. Data processing based on quaternions.	Their framework used only the motion sensors that could be a drawback especially when classifying the locations of the smartphone.	The proposed system uses GPS and microphone data along with the motion sensors to make the location classification more accurate and reliable.
Fan et al. [40]	In the paper, a Context-Aware Human Activity Recognition (CA-HAR) method was proposed with the goal of identifying human behaviours even while the smartphone was not on the user’s body. The system combined several sensor inputs from the smartphone and used ripple-down rules (RDR) and deep learning to identify activities. In order to solve the on-body location issue, RDR rules were developed using a context-activity model that took into account additional contextual data.	Context-aware human activity recognition based on smartphone sensors. Aggregation of data from multiple smartphone sensors. Ripple down rules to rienforce the correct classification of the activities with context.	Real-time recognition performance may be impacted by the increased computational overhead caused by building and maintaining the context-activity model for RDR rules.	The proposed framework possesses the ability to accurately and robustly predict the human activities and the locations without the need of RDR rules. The extracts such features that generate distinctive representations of the activity examples and then trains a strong classifier such as DNDF for the activity and location classification. All these aspects, make the proposed system work better in challenging scenarios.