Research On Partially Occluded Vehicle Detection Methods In Intelligent Transportation System

Nowadays,with the increasing number of vehicles,people pay more and more attention to the driving safety of vehicles.With the continuous development of science and technology,intelligent vehicles will become more and more popular.Due to the popularity of vehicle cameras,more and more road images are generated constantly.Image processing algorithms,especially vehicle detection algorithms,have become an important part of intelligent driving.There are two kinds of vehicle detection algorithms based on computer vision,one is based on traditional image processing,the other is based on deep learning.Due to the improvement of computing hardware performance,the increasing amount of data,and the emergence of more accurate detection algorithms,vehicle detection methods based on deep learning have been more accurate than vehicle detection based on traditional image processing.Simultaneously,some vehicle detection methods based on deep learning have achieved real-time requirements.However,in complex traffic environments,especially in occluded conditions,it is still difficult to effectively detect the vehicles with methods based on deep learning.This article summarizes the detection methods and deep learning basics put forward by the predecessors,and carries out research on vehicle detection under occlusion conditions.As for the vehicle detection problem,in order to solve the problems of low detection accuracy and high miss rate under occlusion conditions,this paper proposes a Gaussian reweighted approach for occluded vehicle detection based on a single-stage detector SSD with high real-time and accuracy.This article first analyzes the training process of the SSD in detail,and designs a loss function based on Gaussian weighting for the mismatch between features and labels in the SSD under occlusion conditions.Also,our method gives different weights for different anchor boxes for training samples in loss function.During the training process,data augmentation is used to increase the number of training samples.At the same time,the backbone network VGG pre-trained on Image Net is used to train the SSD.The experimental results show that the method proposed in this paper can improve detection accuracy of SSD vehicle detector under occlusion conditions.In addition,this paper also researches the Non-Maximum Suppression algorithm under occlusion,and proposes a Non-Maximum Suppression algorithm based on spatial adaptive threshold.Non-Maximum Suppression is the last step in vehicle detection.Different areas in the image have different occlusion conditions,and the occlusion conditions in different directions of the vehicle are also different.Non-Maximum Suppression method with a high threshold is easy to cause false positives for non-occluded areas,and the Non-Maximum Suppression method with a low threshold is easy to cause missed detection for occluded areas.To solve this problem,this paper proposes a Non-Maximum Suppression method based on spatial adaptive thresholds,which changes the traditional greedy Non-Maximum Suppression algorithm with a fixed threshold and considers the positional relationship between occluded vehicles.Our method adjust the threshold of Non-Maximum Suppression adaptively in different regions and different directions.Experimental results show that the method proposed in this paper obtains more accurate vehicle detection results and reduces the miss rate of detection.