Human Behavior Recognition Based On Visual Analysis

With the development of computer vision,human behavior analysis has always been a focus of research.Behavioral classification is susceptible to light and image background factors,and bone data can better avoid this problem.In this context,this thesis mainly carries out attitude recognition for RGB video and extracts bone node data,and conducts behavioral classification based on deep learning algorithm.The following are the main research contents and contributions of this paper:Firstly,the pose estimation problem is studied,and an pose estimation algorithm based on spatial transformation is proposed.In view of the positioning errors and redundant detection problems that often occur in the current attitude estimation algorithms,the addition of space-time transformation network and parallel attitude detection branch further improves the performance of spatial transformation network,which can extract high-quality human body regions and improve the accuracy of attitude estimation.The redundant detection box and redundant attitude are removed by non-maximum suppression.The experimental results show that the proposed algorithm is more accurate and robust than the traditional attitude recognition method,which provides a data base for the following behavior recognition.Secondly,the behavior recognition of human body is studied,and the behavior classification algorithm model based on space-time network is proposed.The traditional graph convolution network is applied to human behavior recognition,and time convolution is introduced to conduct behavior analysis by deep learning.For the current behavior classification method can not make full use of the space structure of the bone data and connection point of the relationship between the problem,this thesis introduces convolution extract to bone data space characteristics,combined with the time convolution,will get 17 key points of gesture recognition of the topology of the human body skeleton model is applied to the space-time figure convolution model behavior classification,achieves the classification accuracy of 90.6%in NTU dataset,experiments prove that the space-time network algorithm model compared with the traditional behavior classification algorithm obtained good classification performance.Thirdly,a behavior classification network model based on two streams is proposed for the behavior classification problem.This network model adds a secondary network on the basis of the original space-time network,which can fuse the data of joint information and bone point information.The experimental results show that the dual-flow network improves the classification performance of the network.The network performance is also optimized,mainly in the addition of BN layer and the introduction of ResNet.Since the graph convolution needs to share weights on different bone nodes,the scale of input data needs to be kept consistent,so BN layer is added to the network model.ResNet layer was added to solve the problem of gradient dispersion in multi-layer space-time cell network model.Experiments show that the classification accuracy of the optimized network model is improved,and the network convergence is accelerated to some extent,which proves the effectiveness of the model.