Video-Based Action Recognition of Spatial and Temporal Deep Learning Models

Action recognition is a fundamental task in computer vision with applications spanning surveillance, human–computer interaction, and video analysis. The work comprehensively explores action recognition techniques, integrating spatial and temporal information to enhance classification accuracy. The process is implemented using the KTH dataset, an action recognition benchmark dataset that is widely used. The proposed approach employs a VGG16 architecture to extract spatial features from input video frames, capturing appearance-related details. Simultaneously, calculating the optical flow between consecutive frames in the dynamic images is utilized to extract temporal features from optical flow images, representing the video’s motion dynamics. These spatial and temporal features are fused to create a comprehensive action representation. For classification, a Bidirectional Gated Recurrent Unit (BiGRU) is employed, where the fused features are fed into the GRU to model temporal dependencies and capture long-term contextual information. The evaluation results validate the effectiveness of the proposed methodology. Fusing spatial and temporal features and incorporating BiGRU improves classification accuracy compared to individually using spatial or temporal features. These findings highlight the importance of integrating appearance and motion dynamics in action recognition.