Research On Human Abnormal Behavior Recognition Method Based On ResNet

The rapid development of computer vision has created necessary conditions for automatic analysis of video sequence images.Conducting human behavior recognition based on massive videos to ensure the surrounding environment safe has become a hot research topic.Abnormal human behavior refers to the behavior that goes against people's common sense or may have dangerous consequences.At present,there are great challenges in dataset and modeling for human behavior recognition.In this paper,the ResNet network model is used to conduct experiments.Channel excitation and motion excitation are added to the time conversion module respectively to improve,and the improved network is used to identify abnormal human behaviors such as shooting,fencing,impact,boxing and smoking.The main research contents of this paper are as follows:In this paper,a time shift module is added on the basis of ResNet network,which can conduct spatial modeling and capture the time relation.The cost of calculation is equivalent to that of the time shift module without addition,but the network cannot conduct definite time modeling for actions.In order to solve this problem,this paper improves the original time conversion module by adding channel excitation and motion excitation respectively.Through the above improvements,after repeated training and testing on HMDB51 and UCF101 data sets,the recognition accuracy has been improved to a certain extent.The recognition accuracy has reached 72.092% and 94.74% when channel excitation is added,and 70.980% and 94.502% when motion excitation is added.Finally,training and testing are carried out on the human abnormal behavior recognition dataset.The recognition accuracy of adding channel excitation to time conversion model is 96.687%,and that of adding motion excitation to time conversion model is 96.386%.The high recognition accuracy proves the feasibility of the improved behavior recognition method.