Research on Violence Action Recognition Method Based on Two-Stream Spatiotemporal Network
Siteng Zhu, Chuanping Hu · 2024
Violence detection through surveillance video is crucial for ensuring public safety, with the reduction of computational resource consumption being a key prerequisite for its widespread application. Given the large volume of raw data and complex data features inherent in violence detection, we propose a lightweight two-stream detection network and a cloud-edge collaborative framework to conserve computational resources while ensuring real-time and accurate detection. The ViBe activity detection algorithm is integrated into the overall detection process as a means for coarse data filtering and data feature enhancement. To simplify the model structure, super-resolution image synthesis is utilized for modeling temporal and spatial features, allowing an image classification network to replace the more complex video classification network. Building on this, a lightweight Vision Transformer network forms the backbone of the two-stream violence detection network. Finally, by leveraging the characteristics of both edge and cloud computing, lightweight models and high-accuracy models are deployed respectively, with their collaboration enhancing overall performance. Experimental results on public datasets demonstrate that the lightweight model can operate in real-time on edge devices and achieve an accuracy of 86.3% on the RWF-2000 dataset. The collaborative approach shows better accuracy performance and lower computational resource consumption compared to solely edge or cloud-based approaches.