Research On Pedestrian Gait Segmentation Algorithm Based On Deep Neural Network

In recent years,video surveillance has been widely used in transportation,military,urban construction,security and other fields,and its importance has become increasingly important.Pedestrian gait recognition has become an important research direction in the field of video surveillance.Extracting the pedestrian area from the image video of pedestrian gait is an important part of pedestrian gait recognition and one of the most demanding computer vision tasks.At present,there are few studies on pedestrian gait segmentation.Most of the data sets used in computer vision tasks such as pedestrian recognition and gait recognition are manually labeled data or segmentation results of instance segmentation.Among them,the hand-marked pedestrian gait note requires a lot of effort,and the existing semantic/instance segmentation method is not accurate enough for the pedestrian segmentation result of the pedestrian foot segmentation.There are two main problems in the application of the current instance segmentation method in pedestrian gait: First,when a pedestrian stands,the "0" shape between the legs is difficult to segment;the second is the contour between the legs when walking.There is a large distance between the actual leg edges,resulting in a segmentation of the legs that is not fine enough.From the perspective of gait recognition,fine leg segmentation is a prerequisite for accurate recognition.In view of the above problems,this paper is devoted to improving the fineness of pedestrian gait segmentation and realizing the automatic segmentation of gait segmentation.The network with better segmentation effect is selected as the benchmark,and two innovations are improved: we continue the work of Mask Scoring RCNN,and compare the segmentation results as our baseline.On this basis,we first divide the network.The last block of the backbone network ResNet is truncated,replaced by the proposed cavern convolution residual block(Atrous res),to obtain more pedestrian information for subsequent network training by improving the receptive field of the deep network.In addition,this paper also designs an edge detection module,which promotes the proximity of the gait edge contour to the true edge of the gait by the edge detection module,thereby improving the fitting degree of the gait edge and improving the fineness of the leg segmentation.In the experimental part,this paper mainly uses the AP value as the evaluation index.The experimental results prove the effectiveness of the work of this paper: the effect of the proposed Atrous res module is improved by 0.4% in the effect of pedestrian gait segmentation.Adding the designed edge detection module,the effect of pedestrian gait segmentation increased by 5.4%.The gait segmentation algorithm proposed in this paper improves the accuracy of gait recognition,which is 2.23% higher than the existing segmentation algorithm.