Research On Pedestrian And Cyclist Detection Methods In Complex Driving Scenes

Pedestrian and cyclist detection is an integral part of Advanced Driver Assistance System.By detecting pedestrians and cyclists and providing prejudgments to cars can greatly reduce the incidence of traffic accidents and reduce the level of casualties,which has a profound impact on human life and society.Existing algorithms mostly deal with pedestrian detection and cyclist detection separately.Few algorithms detect pedestrians and cyclists at the same time.In actual application,it will result in waste of hardware resources and waste of time.Therefore,this paper aims to propose an algorithm for simultaneous detection of pedestrians and cyclists.Pedestrian and cyclist detection is an application task of object detection task.Although in some fields,classic universal object detection algorithms such as YOLO can perform object detection quickly and accurately,when performing pedestrian and cyclist detection,specific analysis is required for specific problems.In order to provide the system with a clear and wide driving field of view,the images used for pedestrian and cyclist detection usually have high resolution.In a real driving environment,the road environment is complex and changeable,with large differences in object individual and size.Traditional universal detection algorithms are difficult to adapt to these problems when performing pedestrian and cyclist detection.In view of the above problems,this paper establishes a pedestrian and cyclist detection framework,and studies related technologies involved in pedestrian and cyclist detection.The detection framework can detect pedestrians and cyclists at the same time,and distinguish two types of targets on the basis of distinguishing foreground and background.The pedestrian and cyclist detection framework proposed in this paper includes three parts:region of interest extraction algorithm,detection algorithm,and optimize processing.In order to extract the region of interest from high-resolution images,reduce the object detection range,and solve the problem of missed detection caused by large image resolution and low object resolution,this article analyzes and introduces two algorithms:Aggregate Channel Feature(ACF)and Local Decorrelated Channel Feature(LDCF).On this basis of these two algorithms,the article proposes ACF-RP region of interest extraction algorithm and LDCF-RP region of interest extraction algorithm.These two algorithms use ACF detector or LDCF detector to perform preliminary extraction of candidate regions,and analyze and process the preliminary extracted candidate regions to obtain the final region of interest for detecting.The algorithm structure is analyzed theoretically,and the effectiveness of the algorithm is proved by experiments.Experiments show that the proposed algorithm can reduce the detection range while ensuring that the target is not missed as much as possible.In order to solve the problems of pedestrians and cyclists with various poses,lighting and occlusion,the detection network in this paper is improved based on YOLOv3.In order to merge detection results at different resolutions and solve the problem of large size spans between pedestrians and cyclists,this paper uses a multi-branch detection network PM-YOLO.The detection network contains multiple branches,each branch corresponding to a different resolution,and fusion optimization is performed on each branch result to reduce the target missed detection rate.To further optimize the detection results,after the detection results of each branch network are fused,the results are mapped,and the non-maximum suppression algorithm is used to optimize the detection results.This paper performs verification experiments on the proposed algorithm.The experiments show that the proposed pedestrian and cyclist detection algorithm can simultaneously detect pedestrians and cyclists in complex driving scenarios.It proves the effectiveness of the algorithm.