Optimized Depthwise Separable CNN-SVM Approach for Enhanced Human Activity Recognition

Human Activity Recognition (HAR) is paramount in several modern-day applications. However, the accuracy in recognizing the activity remains a challenge due to several factors such as natural variations in human movement, similar posture in various activities, physically challenged individuals, incomplete sensor data due to obstructions, unwanted signals affecting sensor readings, and more. This study introduces a hybrid model that combines the efficacy of Support Vector Machine in high dimensional space and potential of Convolutional Neural Network to capture the temporal and spatial pattern in data. It allows the model to handle complex data and generalize effectively thereby reducing the chances of overfitting. The two models are integrated through a combined decision function using a weighted average ensemble technique. The performance of the proposed model is evaluated on the UCI HAR dataset and compared with state-of-the-art models like SVM, Convnet, LSTM-CNN, CNN, CNN-BiLSTM, etc. In comparison to models like LSTM-CNN (95.80%) and CNN-BiLSTM (96.37%), the proposed model showed superlative performance with overall classification accuracy of 97.18% and average class accuracy of 99.05%. Moreover, it also exhibited higher precision, recall and F1-Score thus, giving a more reliable solution for advanced HAR tasks.