Research On Vehicle And Pedestrian Detection Algorithm In Complex Scenes

The rapid increase in the number of vehicles and pedestrians in recent years has led to traffic jams and safety accidents.To relieve this problem intelligent vehicle systems and assisted driving systems based on autonomous driving technology have become a hot research topic.Real-time detection of vehicles,pedestrians and other objects from contextaware information,especially in complex scenes such as rain,snow and night,is one of the important problems that intelligent vehicle systems need to solve.This thesis focuses on the detection of vehicles,pedestrians and other targets in complex scenes.The complex scenes involved include scenes with many visual distractions such as bad weather(e.g.,rain and snow)and night,scenes where vehicles and pedestrians are occluded,and scenes where vehicles and pedestrians are densely distributed.In addition,vehicles and pedestrians have different characteristics: vehicles are rigid targets with small deformation and similar shapes,while pedestrians are non-rigid targets with varying shapes.The target size varies greatly as the target depth varies.All of these problems make accurate detection of vehicles and pedestrians difficult.In this thesis,we study algorithms for detection of vehicles and pedestrians in these complex scenarios,and the main contents are as follows:Firstly,a vehicle and pedestrian detection method ATDW-YOLOX integrating convolution and attention mechanism is proposed,which is based on CSPNe Xt,and a feature extraction network integrated with self-attention mechanism,convolution operation,and coordinate attention mechanism module is designed to enhance the feature extraction capability of the backbone network;Based on the thoughts of depth-separable convolution and attention mechanism,the feature fusion module ATDW-PAFPN is designed to promote the sufficient fusion of shallow and deep features;CIo U and Varifocal Loss are used in the bounding box regression loss and confidence loss,respectively,to adapt to the vehicle and pedestrian detection in complex scenes.The experiments’ result shows that the method can achieve better detection results when dealing with bad weather and mutual occlusion scenes between targets.Secondly,a vehicle and pedestrian detection method TA-ASFF-YOLOX based on adaptive feature fusion with classification and regression task alignment is proposed,which uses the idea of bi-directional weighted feature fusion to first build the feature fusion module Sim-Bi FPN,then uses the adaptive feature fusion module to enhance the feature fusion to make full use of various useful features,and finally uses the task aligning the detection heads so that the classification and regression tasks are decoupled while maintaining some information interaction to make the detection more accurate.Experiments show that the method can achieve better detection results when dealing with dense distribution of vehicles and pedestrians and excessive scale differences.Finally,the proposed algorithms are experimented on the publicly available dataset BDD100 K,and the experimental results show that the proposed methods are able to provide high detection accuracy and robustness when detecting vehicles and pedestrians in complex scenes.