Research On All-Weather Pedestrian Detection Technology Based On Deep Learning

Pedestrian safety is the most important part of unmanned driving.Therefore,pedestrian detection technology in real-world scenarios is particularly important.However,in the real scene,the pedestrian target has various situations such as truncation,occlusion,excessive illumination,insufficient or rain,snow and fog,which will have a great impact on the accuracy of pedestrian target detection.Therefore,in various scenarios,pedestrian target detection can have a good accuracy and speed is a problem that still needs to be solved.This paper studies pedestrian target detection in various scenarios based on deep learning methods.The main research contents of this paper are as follows:(1)For pedestrian target detection,two commonly used target detectio n methods are studied.They are a single-stage target detection algorithm and a two-stage target detection algorithm.Among them,the two-stage target detection algorithm mainly analyzes the RCNN series of algorithms,and the single-stage target detection algorithm mainly analyzes the SSD algorithm and the YOLO series of algorithms.And choose YOLOv5 algorithm as the main research method in this paper.(2)For scenes with sufficient or insufficient daily light,an improved pedestrian detection algorithm based on YOLOv5 is designed.The Conv Ne Xt network is selected to modify the backbone network of the YOLOv5 algorithm,and the CA attention mechanism is introduced to improve the effect of pedestrian target feature extraction.And the part suppressed by the fi nal prediction frame is improved to improve the accuracy and recall rate of pedestrian detection in daily scenes.(3)For other different scenarios such as night,fog,rain,snow and other meteorological environments,pedestrian detection algorithms in com plex scenarios are designed.It is still based on the YOLOv5 algorithm and optimized according to different scenarios.First,the Bi FPN structure is used to replace the PANet structure of the Neck part of YOLOv5 to improve the feature fusion ability of the model and the overall detection accuracy of the algorithm.At the same time,different data enhancement methods are used for images of different scenes.The method of histogram equalization is adopted for nighttime scenes,and the image adaptive enhancement module is introduced for weather such as rain,snow and fog to improve the pedestrian characteristics of different scenes.And through experiments,the accuracy and real-time performance of the improved YOLOv5 algorithm in complex scenarios are verified.