Research On Human Behavior Recognition Method Based On Vision

Human behavior recognition has always been one of the most challenging problems in computer vision,which has been widely concerned by researchers in industry and academia.However,human actions usually take a certain amount of time to complete.In the past,most of the methods were based on single image recognition and classification,which was difficult to extract the time-domain information in the video,resulting in unsatisfactory model classification results.In the open data set of behavior recognition,although the quantity and quality of videos of various kinds of behaviors are guaranteed,human posture and background environment are still irregular and cannot be described linearly.Therefore,it is difficult to model video features through traditional manual feature extraction methods.Due to the different shooting methods among different open data sets,the video feature manifestations are quite different.Compared with the artificial classification with high precision,high accuracy and complex reasoning ability,the network model with fixed structure usually has weak generalization ability among different data sets.Different from the common open data sets,the data generated by human daily activities are usually disturbed by a lot of noise,and the distribution form of the data is uneven,which has a serious impact on network recognition.Therefore,based on the above problems,this paper focuses on the modeling ability of the model to the video medium and long time sequence information,as well as the antiinterference ability of various types of noise in the data samples.The main research contents of this paper are as follows:(1)I made a scientific analysis of SlowFast backbone network,improved the structure of the basic network through theoretical reasoning and practical experience,and studied the imbalance of original data and the diversity of feature forms to achieve the information complementarity between samples,and enhanced the model’s processing ability of the data with occlusion,uneven lighting and fuzzy video frames.Enhance the ability of the deep Web to retain more key features.(2)Build an embedded DCN-Attention module based on deformable convolution and attention mechanism,which can enhance the modeling ability of the network for irregular targets and effectively improve the network performance.(3)Fusing the DCN-Attention module with the improved network,a Slow Fastv2 network with better performance in training was proposed.Aiming at the problem of unbalanced distribution of training set data,the Focal Loss function in the binary task was extended to this task through the formula.The experiment demonstrated that the Slow Fastv2 network greatly improved its robustness and reliability in the video classification task.