Optimizing Algorithms for Human Behavior Recognition: Single-Modality Spatiotemporal Convolutional Networks

Behavior recognition constitutes a pivotal area of research within computer vision and serves as an integral technology in video surveillance and related domains, thereby possessing substantial research value. Convolutional neural networks (CNNs) have been proven to possess robust learning capabilities. However, due to the high complexity and variability of human behavior, recognizing human behaviors in videos remains challenging due to the difficulty of capturing useful features over long time series and the redundancy of multi-frame information. To address these issues and improve the utilization of continuous frame information, we have designed a cross-frame multiscale information fusion module. This module allows the backbone network to acquire multiscale temporal features from different frames, thereby reducing information redundancy across multiple frames while minimizing additional computational costs. Furthermore, to mitigate the problem of video temporal feature loss, we introduce an easily integrable attention mechanism module, the Easy-Attention Block. This module facilitates rapid capture of long-range dependencies, enabling the model to focus on the most relative parts of continuous video sequences and suppress irrelevant temporal information. Consequently, the model can extract more discriminative spatial-temporal features, thereby enhancing its efficiency in handling long sequence data. Experimental results express that this model can achieve high recognition accuracy and effectively extract the temporal features of video, suitable for real-time applications in critical infrastructure surveillance, which enhancing safety and reducing potential risks in energy power systems.