Human Behavior Recognition Method Based On Double-stream Deep Convolution Neural Networks

In recent years,benefiting from the rapid development of intelligent city,intelligent security and other fields,human behavior recognition technology using deep learning methods has become a hot research direction of many researchers,whether in the academic field or in the industrial field.Although these deep learning methods for video streams have achieved good recognition rates,they all have complex network design,difficult training,and high-performance hardware support.Therefore,this paper divides the human behavior recognition into two steps,firstly,a set of frame-level skeleton images is extracted by using the built-up human detection model,and then the human behavior classification model is used for recognition.Converting the video stream into set of frame-level skeleton images by "two-step" can reduce the dependence on high-performance hardware,and use set of frame-level skeleton images to classify the behavior,reducing the complexity and training difficulty of the network structure.In short,this method can not only reduce the difficulty of training and the dependence on high-performance hardware,but also realize the correct recognition of seven general human behaviors,and has good robustness and generalization ability.First of all,in order to overcome or improve the interference of complex back-ground on human behavior recognition,a set of human behavior images is transformed into a set of skeleton images,and the detection algorithm is used to construct the human detection model based on the double-stream deep convolution neural networks,Among them,in order to obtain an accurate human skeleton images,a human torso correlation domain method is added.Through the multi-stage,step-by-step refinement of the net-works to predict the correct connection between the keypoints of human body and human torso,the model improves the accuracy of outputting the human skeleton images,and achieves real-time output of human behavior skeleton images.Secondly,this method self-built the dataset of human behavior skeleton images.By analyzing the different characteristics of seven human behaviors in skeleton images,thecorresponding constraints are designed to ensure quality and efficiency of the collected dataset.Then combined with the human detection model to build dataset collection software,Then,the dataset collection software is built based on the human detection model,which can collect data efficiently in different scenes by using multiple cameras.Finally,the behavior categories represented by the set of frame-level skeleton images are integrated into the dataset of human behavior skeleton images.Then,a human behavior classification network is built for the set of skeleton images,which uses the aggregation pooling module and the horizontal pyramid pooling module.The aggregation pooling module aggregates the features of the input multiple skeleton images into one feature vector,which retains the global and local characteristics of the features.The horizontal pyramid module obtains global and local spatial information of input features through different scales,which makes the features more discriminative and improves the recognition of different behavior categories.At last,the experimental results show that the human detection model meets the requirements of subsequent experiments,which can extract accurate skeleton images and ensure real-time performance.The human behavior classification model was tested by self-built dataset of the human skeleton images,and the average recognition accuracy reached 92.6%.In addition,the model was tested using seven kinds of behaviors at five angles,all of which achieved good recognition accuracy.In summary,this method can correctly identify the general seven kinds of human behavior,and has good robustness and generalization ability,which further demonstrates the effectiveness and feasibility of the proposed method.