Action Recognition Based On SlowFast And Temporal Segment Strategy

As an important way of information dissemination,video plays an increasingly important role in the exchange of information in today's society.At the same time,the development of the Internet and imaging equipment has lowered the cost of video production and transmission.The growth of video data has put tremendous pressure on video supervision.The development of artificial intelligence makes it possible to build an intelligent video supervision system.Therefore,behavior recognition,one of the core technologies of the intelligent video surveillance system,has become a research hotspot.Deep neural networks have become an important tool in the field of image processing due to their excellent feature extraction capabilities.Thanks to its success in the field of image processing,deep learning has also become the mainstream of human behavior recognition algorithms.The Slow Fast network combining 3D convolution and dual stream structure has become one of the better networks at present due to its good performance.This article mainly studies on the basis of Slow Fast network,and the main work done includes:(1)A Slow Fast network with time-domain segmentation structure is proposed.By analyzing the input of the Slow Fast network,it is found that it can only process short-term video information and cannot model the entire video.This paper proposes an algorithm for the fusion of Slow Fast network and time-domain segmentation structure-TS-Slow Fast.This structure can maintain the overall time domain information of the video on the basis of the original computational complexity and obtain a video-level understanding.Experiments show that TS-Slow Fast is better than the original network.In addition,this paper also inputs the inter-frame difference and optical flow as the time stream input information of the TS-Slow Fast network to the network,and compares its recognition accuracy.Finally,the TS-Slow Fast network is compared with some existing algorithms,which proves that the proposed TS-Slow Fast algorithm has a certain improvement in recognition accuracy.(2)An improved TS-Slow Fast network algorithm with MERS is designed.Optical flow is better when used as TS-Slow Fast time flow network input because it provides enough motion information.However,as an artificial feature of optical flow,computing optical flow requires a lot of computing resources and storage resources.MERS combines distillation and privileged learning ideas to directly extract advanced motion-related features from RGB images,avoiding optical flow calculations during testing.Experiments show that this method greatly improves the operation efficiency of the behavior recognition algorithm under the premise of losing a certain recognition accuracy.(3)Designed an intelligent monitoring system based on improved TS-Slow Fast network.The system is mainly composed of video acquisition module,video preprocessing module,feature extraction module,feature recognition module and alarm module.In addition to operations such as denoising and histogram equalization to improve video quality,the video pre-processing module can also segment target video segments in the video.After that,the feature extraction and recognition module completes the recognition process of the video segment.Through the cooperation of multiple modules,an intelligent monitoring system with good recognition function is formed.