Research On Recognition Algorithm Of Human Abnormal Behavior Based On Video

Video-based human abnormal behavior recognition is one of the key technologies of intelligent video surveillance system,and it is also an important research topic in the field of artificial intelligence and computer vision.The human body abnormal behavior recognition algorithm can reduce the dependence of the video surveillance system on manpower,and greatly improve the handling efficiency of the accident,and control the loss to a small range or even completely avoid it.Therefore,the human body abnormal behavior recognition has important research value.This paper takes risk as the entry point for the study of abnormal human behavior,and conducts research on the recognition of abnormal human behavior in videos from the algorithm level to improve the accuracy of the abnormal behavior recognition algorithm.The main research contents are as follows :(1)Researched the optical flow extraction of abnormal behavior in video and accelerated optimization.Aiming at the problem of difficult extraction of optical flow in abnormal behavior videos,this paper uses the CUDA platform to optimize the optical flow algorithm for GPU acceleration.Firstly,study the principle of CUDA programming,Then analyze the calculation process of optical flow method for parallel processing,use the of parallel reduction to improve the algorithm,and use the shared memory and the CUDA flow model to achieve the optimization of the calculation process.Experimental results show that the optical flow GPU acceleration algorithm based on CUDA can greatly increase the speed of optical flow extraction.For images with a resolution of 240×320,compared to CPU(i5-7300hq)calculations,GPU(GTX 1050)calculations can obtain 45.4 Speedup ratio,optimization has further obtained 48.4 Speedup ratio.(2)Researched the multi-scale global feature fusion abnormal behavior recognition based on temporal segment networks.Aiming at the problem of difficulty in global modeling of abnormal behavior recognition videos,this paper adopts the idea of Temporal Segment Networks,proposes global feature aggregation and multi-scale feature fusion,and realizes a more effective abnormal behavior recognition algorithm.First,the spatial stream network and the time stream network are used to extract the local features of the video segment,and then the local features are aggregated to form a global feature expression of the video,and a multi-scale convolution kernel is designed to further extract different scales features and fusion,and finally a classifier is used for classification and recognition.The experimental results on the UCF101 data set show that this method can effectively express the global features of the video,and the accuracy is improved by 1.33% compared with the original time domain segmentation model..(3)Researched the improved abnormal behavior recognition combined with spatial and channel attention.Aiming at the problem of insufficient effectiveness of local feature extraction by abnormal behavior recognition algorithm,a variety of improvements have been made to the abnormal behavior recognition algorithm.In the local feature extraction stage,the spatial channel attention module is added to capture the key information in the feature stream and improve the effectiveness of feature aggregation;In the scale feature fusion module,the volume integration solution is used to enrich the network level,improve the feature learning ability,and reduce the calculation cost;Dropout is used in training to reduce the risk of overfitting.The accuracy of the improved algorithm is further improved,reaching 93.94%,which is more accurate than other mainstream methods.