Behavior Detection And Recognition Based On Deep Learning

With the advent of intelligent and information society,behavior detection and identification method using deep learning has replaced behavior detection and identification method using traditional features step by step.Behavior features extracted based on deep learning have better generalization performance than traditional manually extracted features.Therefore,behavior detection and recognition technology based on deep learning has received extensive attention from researchers.At present,there are many researches on behavior detection and identification method using deep learning,but they rarely involve the office field.The subject of this paper is based on the extension of Zhejiang Province's key R&D project,which detects and recognizes the behavior in the images collected by the surveillance camera in the office scene.Due to problems such as the distance and angle of the surveillance camera from the target to be detected,the detection of small-scale target objects in the image is an urgent problem to be solved.At the same time,in scenes with complex actual backgrounds,human targets may be occluded,and behavior detection and recognition of occluded human bodies is also a problem that needs to be solved.In view of the questions mentioned above,this article using deep learning technology,human skeleton key point detection technology and target detection technology to finish the study on the detection and identification of behavior in the image.This topic uses the office area of the employees in the cinema as the experimental environment,and put forward a behavior detection and identification method using deep learning to detect and identify the behavior in the office area.The major research contents of this article are as follows:(1)Aiming at the problem of low detection accuracy of small-scale targets,a target detection method based on neural network cascade is proposed.This method uses two Retinanet network models which is cascaded to detect targets,relatively increase the receptive field of small target objects,thereby improving the detection accuracy of small target objects.At the same time,it is proposed to change the conventional convolutional layer in the Retinanet backbone network to a deeply separable convolutional layer to reduce the resource consumption and calculation occupancy caused by network cascade,thereby improving the detection speed of the algorithm.Finally,an improved bounding box loss function LIoU is proposed to further improve the accuracy of the algorithm.Experimental results show that on the collected data set,the average detection accuracy(mAP)of the algorithm for small target objects reaches 56.9%,and the detection time is 0.083s/frame.Under the same detection speed,the average detection accuracy is improved by 4.9% compared with the traditional algorithm.(2)Aiming at the problem of inaccurate positioning of key points caused by occlusion of human targets,a key point positioning algorithm of human skeleton based on improved cascaded pyramid network is proposed.The algorithm adds the attention module to each residual block of the CPN feature extraction network,and assigns different weights according to different parts of the feature map and the importance of different feature maps.At the same time,the two upsampling operations of the original cascaded pyramid network are changed to one to reduce the redundant background features generated during the upsampling process.Experimental results show that the algorithm has an average positioning accuracy of 73.2% for key points under occlusion and complex background conditions,which is 1.1% higher than the original CPN network.The behavior detection and recognition algorithm based on deep learning proposed in this paper can effectively improve the detection accuracy of small-scale target objects,and at the same time,it can also improve the positioning accuracy of key points of occluded human targets.