Improved Accuracy and Computational Efficiency in Human Activity Recognition through Hybrid CNN-GRU Models

This paper proposes a novel Multiscale Convolutional Neural Network (MCNN) using CNN3D architecture for human behavior recognition. Unlike traditional models, it captures spatial and depth information through space-time and depth-separable convolution modules. Evaluated on the UCI HAR dataset, the MCNN model demonstrates higher accuracy (94%) and reduced complexity (3306 parameters) compared to CNN2D (93%, 9135 parameters). An extension combining CNN, GRU, and Bidirectional algorithms further improves accuracy (96%) with only 1162 parameters, optimizing feature extraction and reducing training complexity. The approach showcases significant advancements in efficiency and performance, making it highly suitable for activity recognition using smartphone-captured data