Research On Pedestrian Detection Algorithm Based On Deep Learning

Pedestrian detection has always been a research hotspot in the field of computer vision.In recent years,with the rapid development of deep learning research,pedestrian detection has been widely used in many fields such as autonomous driving,intelligent security and intelligent robots,which has made pedestrian detection more and more concerned in academia and industry.Pedestrian detection identifies all pedestrians in a given image or video frame by correlation detection algorithms and marks their position and size in a rectangular frame.However,in practical application scenarios(such as autopilot and video surveillance),the image to be detected is usually affected by factors such as illumination,camera equipment,obstacles,and multiple pedestrians,resulting in background interference,pedestrian scale being too small,and occlusion,especially small-scale pedestrian and pedestrian occlusion,has been difficult to detect pedestrians bothered research,which seriously reduces the pedestrian detection performance.In this paper,based on the typical problem of pedestrian detection,the Faster R-CNN target detection network is improved,and the deep learning-based pedestrian detection algorithm with higher detection accuracy and more robustness to non-ideal detection environment is discussed.The contents and results of the specific research work are as follows:(1)In order to reduce the existing detection algorithm and the fact that Faster R-CNN generates a large number of missed detections for small-scale pedestrians,a new Faster R-CNN deep learning network detection algorithm is proposed.It adopts an alignment pooling method based on bilinear interpolation to avoid the positional deviation caused by two quantization operations in the pool of interest regions;and proposes a cascade-based multi-layer feature fusion strategy,which will have The shallow feature map with rich detail information and the deep feature map with abstract semantic information are merged to solve the problem of insufficient feature information of small-scale pedestrians in deep feature maps.The experimental results show that,on the small target pedestrian datasets in INRIA and PASCAL VOC2012,the average accuracy rate(AP)is increased by 17.58% and 23.78%,respectively,compared with Faster R-CNN under the condition of comparable detection efficiency.(2)On the basis of the above research,in order to improve the robustness of the algorithm for occlusion pedestrian detection,especially in the crowds of dense crowds,the occlusion between the people(i.e.,pedestrian self-occlusion),by analyzing the error cases in the actual detection,it is found that Faster R-CNN can not accurately locate the overlapping pedestrians under the guidance of the original regression loss function,resulting in the missed detection of self-occlusion pedestrians,therefore,it is considered to further optimize the detection network from the perspective of the loss function.This paper introduces a regression loss function-Repulsion loss function with "repulsive force" on adjacent detection objects,and retrains the detection network.The experimental results show that the average accuracy(AP)is improved by 6.07% and 1.42% on the Occlusion-person occlusion pedestrian dataset and INRIA dataset compared to the improved Faster R-CNN[43] under the condition of reasonable detection efficiency.