Research On Pedestrian Detection Technology Based On Deep Learning

Pedestrian detection is a technology that uses computer vision related technologies to determine whether there are pedestrians in an image or video and to provide precise position.This technology can be combined with pedestrian tracking and pedestrian re-identification to detect pedestrians,identify pedestrians and track pedestrians.It is widely used in artificial intelligence systems,intelligent driving assistance systems,and intelligent robots.Because the appearance of pedestrians is easily affected by scale,occlusion,posture,etc.,the detection accuracy is greatly compromised.Therefore,how to improve the detection accuracy of pedestrians in complex environmental conditions while meeting the requirements of detection speed is a hot topic with great research value.In this paper,based on the deep learning pedestrian detection method,researches are carried out in two aspects: small-scale pedestrian detection and occlusion pedestrian detection.The specific work and research contents are as follows:1.Introduced the one and two stage detection algorithms,and selected the YOLOv3 detection algorithm and the Faster RCNN detection algorithm to compare on the Caltech pedestrian data set.In order to better meet the requirements of detection speed,select the far subset to test.YOLOv3 is used as a small-scale pedestrian detection algorithm due to the better performance on far subset.In order to achieve higher occluded pedestrian detection accuracy,the two stage detection algorithm Faster RCNN with higher accuracy in the occluded pedestrian subset is selected as the occluded pedestrian detection algorithm.2.Aiming at small-scale pedestrian detection,it is proposed to use the Multi-Feature Fusion module to enhance the feature expression of model for small-scale pedestrians.This module mainly consists of two parts.The first part uses the enhancement of receptive field module to extract features from the lower-level feature map to enhance the receptive field of the shallow feature map.The second part uses an adaptively spatial feature fusion module to replace the traditional dimensional splicing operation,adaptively learns the spatial weight of each scale feature maps,makes full use of the feature information of each scale,and enhances the detection results of small-scale pedestrians.At the same time,the kmeans++ clustering measurement parameters are adjusted to obtain anchors that better match the pedestrian scale of the dataset.Tested on the Caltech pedestrian dataset and the Person subset in VOC 2007 dataset,the results show that the proposed MFFE module has an effect on the improvement of pedestrian detection performance,and good detection results have been obtained on different datasets.3.Aiming at the detection of occluded pedestrians,an improved attention mechanism is proposed,which combines channel attention and spatial attention to improve the network’s attention to the unoccluded parts of pedestrians.Our method improved the regression strategy of RPN and enhanced the positioning accuracy of occluded pedestrians.Our method selected NMS algorithm based on DIo U to retain more accurate candidate frames,and improved the accuracy of occluded pedestrian detection.Tested on the City Persons dataset and Wider Person dataset,the results show that the detection accuracy of occluded pedestrians is higher,which prove the effectiveness of our method.4.Using Python to build a pedestrian detection system based on deep learning,programming to achieve the detection of pedestrian data collected in actual scenes,mainly including single-image pedestrian detection and video pedestrian detection.