A Comparative Analysis of Human Activity Recognition Based on Spatio-Temporal Hybrid Models

Human activity recognition (HAR) plays a vital role in computer vision and human-computer interaction. This paper addresses the challenges of accurately recognizing activities by extracting spatial and temporal features from sequential data. We propose spatio-temporal hybrid model frameworks utilizing Long-term Recurrent Convolutional Networks (LRCN), Convolutional Long Short-Term Memory (ConvLSTM), and Convolutional Gated Recurrent Unit (ConvGRU) architectures are used to extract the spatial and temporal features to enhance the HAR. Various optimizers, including Adam, SGD, and RMSprop, were evaluated with different learning rates to improve model performance. Experiments on the UCF50 dataset demonstrate the superior accuracy of 98% achieved by ConvLSTM with SGD at a learning rate of 0.001. Adam provided optimal performance for LRCN and GRU with accuracies of 94% and 96%, respectively. The significance of the model based on different learning rates and optimizers is studied and justified in the ablation study