Method and Analysis of Violent Behavior Recognition Based on Multimodal Information Fusion

Abstract

This article provides a systematic review and analysis of methods for identifying violent behavior based on multimodal information fusion. The research focuses on two main themes: "Feature Extraction Network - Multimodal Fusion Strategy". In a unified experimental framework, the article compared the differences in parameter quantity, computational efficiency, and detection accuracy among 3D CNN, ConvLSTM, GhostNet, and other networks, and analyzed the performance of early fusion, late fusion, and attention fusion strategies on public datasets such as RWF-2000. By quantitatively comparing the interaction between feature extraction networks and fusion strategies, this study aims to fill in the current review that lacks engineering implementation guidance. Research has found that existing reviews often focus on a single technological path, or only list network structures, or compare fusion strategies, lacking a systematic induction of the "network fusion" coupling effect within the same framework. This makes it difficult for researchers to make quick decisions based on computing power, real-time performance, and privacy constraints during actual deployment. This article provides a reference for researchers to choose appropriate feature extraction networks and fusion strategies in practical applications by analyzing several representative works in a unified coordinate system.