Smart Approach to Human Activity Recognition in IOT Using Hybrid-based Ensemble Classifier
Siju V. Soman, Jane Rubel Angelina Jeyaraj · Journal of Circuits Systems and Computers · 2025
In everyday life, HAR plays a critical role in obtaining precise insights into human activity from sensor data. Video-based systems have several drawbacks, including inadequate coverage, expensive prices and privacy problems, despite the fact that they may efficiently record physical activity. While sensor-based devices have the benefit of being able to record a variety of behaviors, they also have a number of drawbacks, such as poor contextual awareness and data noise. Both approaches have serious shortcomings. To overcome these challenges, this research proposed a novel Hybrid Ensemble Classifier (HEC) for recognizing activities in Smart Home HAR. Preprocessing uses a Bilateral filter to reduce noise and the HOG to capture multiscale information. To extract deep features from image frames, propose an Inception-based CNN with Bi-GRU. The process involves multiple convolutional layers to produce hierarchical information, with max-pooling reducing spatial dimensions while preserving crucial information. Global Average Pooling (GAP) is used to minimize overfitting. Bi-GRU is then incorporated to model temporal relationships and sequential information, reducing overfitting and improving pattern recognition. Finally, an Ensemble classifier determines the final prediction by majority voting among classifiers, ensuring the most accurate class label is chosen. The result reveals that the proposed technique is related to seven prevailing approaches, and it demonstrates superior performance with an accuracy of 99.9%. Additionally, it attains the maximum precision, recall and F1-score of 1.00 for ensemble classifiers. This approach significantly enhances HAR accuracy and robustness, offering a comprehensive solution to the limitations of existing methods.