Road Target Detection For Autonomous Driving In Complex Traffic Scenario

With the increasing of vehicle ownership,the traffic problems are becoming more and more obvious.As an important technology to solve the development problems of the traditional automobile industry,autonomous driving has received more and more attention and research from the state and enterprises.The accurate detection of road targets in traffic environment is the premise of completing the task of autonomous driving.At present,with the rapid development of vision-based target detection algorithm,road target detection in general traffic scenarios is no longer a problem.However,in complex traffic scenarios,dense targets are occluded with each other and there are a certain number of small targets,the target missed detection is serious,and the detection accuracy is difficult to meet the requirements.Therefore,based on deep learning target detection technology,this paper carries out research on road target detection algorithm in complex traffic scenarios oriented to autonomous driving.The specific research contents are as follows:(1)Framework design of target detection algorithm based on comparative analysisBased on deep learning target detection technology,three target detection algorithms,Faster Regions with CNN features(Faster R-CNN),Single Shot Multi Box Detector(SSD)and You Only Look Once Version Four(YOLOv4),are studied and implemented in this paper.A complex traffic scenario data set CTS is constructed to solve the problem of the shortage of complex traffic scenario data set,and algorithm comparison experiments are carried out on the CTS data set using the commonly used evaluation indexes.The results show that YOLOv4 target detection algorithm is superior to the other two algorithms in terms of precision and detection speed.The detection precision m AP is 78.46%,and the detection speed is up to32.78 frames per second.While ensuring accuracy,it can also meet the real-time requirements of autonomous driving.Therefore,YOLOv4 algorithm is selected as the basic framework of this paper,but there is a certain degree of missing detection problem for occluded targets and small targets in complex traffic scenarios,which needs further research.(2)Improve non-maximum suppression algorithm to solve the target occlusion problemAiming at the problem that the algorithm has a low positioning accuracy for the occlusion target,this paper proposes a new non-maximum suppression algorithm,Soft-DIo U-NMS,based on the regression loss function of CIo U Loss,combining the attenuation strategy of Soft-NMS and the DIo U evaluation index of DIo U-NMS.The experimental results show that the m AP of the improved YOLOv4 algorithm on CTS is80.39%,and the detection speed is 31.52 frames per second,which further improves the detection accuracy while ensuring the real-time performance;In addition,according to the detection results of different data sets,the detection ability of the improved algorithm is improved on the simple data set VOC2007,while the detection ability of the improved algorithm is not significantly decreased on the complex data sets COCO2017 and KITTI,which verifies the good generalization ability of the improved YOLOv4 algorithm.(3)Introduce Focal Loss to solve small target detection problemAiming at the problem of low proportion of small targets in data set,this paper first uses mosaic data enhancement method to increase the number of small samples to enrich the data set,and then improves K-means clustering algorithm to generate more accurate prior boxes;Aiming at the problem that small targets aggravate the sample imbalance,the algorithm uses Focal Loss instead of cross-entropy loss to participate in the calculation of model loss,which makes the model more inclined to learn from positive samples and samples that are difficult to classify,which improve the detection accuracy of small targets.The experimental results show that the recall rate and precision rate of different categories of the improved YOLOv4 algorithm are improved,and the problem of missing detection and false detection caused by small targets is effectively alleviated;The improved YOLOv4 algorithm has different degrees of the m AP improvement in CTS,VOC2007,COCO2017 and KITTI data sets,which verifies that the improved YOLOV4 algorithm can achieve stable target detection function on different data sets and has good comprehensive detection ability of road targets in complex traffic scenarios.Focusing on the problem of low detection accuracy of road targets in complex traffic scenarios of autonomous driving,this paper studies the detection of occlusion targets and small targets in complex traffic scenarios.The breakthrough points of the research work are as follows:(1)Aiming at the problem of low positioning accuracy of occluded targets,an improved non-maximum suppression algorithm Soft-DIo U-NMS is proposed to realize the effective detection of occluded targets;(2)Aiming at the unbalanced sample of one-stage algorithm,Focal Loss is introduced to participate in the loss calculation of the model,which effectively alleviated the missed and false detection problems of small targets.