Research On Traffic Target Recognition Algorithm Based On Small Target Detection

With the continuous improvement of industrial digitization,informationization and intelligence,especially the in-depth implementation of the "14th Five-Year Plan" for the development of digital economy,automatic driving technology has become the main theme of intelligent transportation development.Traffic road target recognition,as one of the important technologies in automatic driving,has been widely applied to highway chokepoints,violation detection points,parking lot entrances and exits,and other fields.This study takes the small target detection algorithm as the breakthrough,takes the traffic road target as the research object,and uses the Atrous Convolution strategy,super-resolution reconstruction technology,backbone network lightweighting,combined with YOLO(You Only Look Once)series of algorithms to build the traffic road target recognition model,respectively,for lane lines,vehicles and license plates,traffic signs and other targets to Detection,to provide theoretical and technical support for solving the problem of small target missing and wrong detection.The main innovative work and research contents of the thesis are as follows.(1)Aiming at the problem of inaccurate and unstable lane line detection,the original YOLOv3 algorithm was improved by the Dilated Convolution strategy and lane line detection was carried out.The network felt field was improved by setting appropriate dilation rate parameters,and lane line information was extracted effectively to avoid missing and error detection.Finally,Kronecker convolution is used to improve the Dilated Convolution strategy,and then the optimal parameter value is found.Experimental results show that the Improved Algorithm of YOLOV3-KCID(Kronecker Convolution Improved Dropout)can detect lane lines with a detection accuracy of 94.59% and a detection speed of 33ms/ page,which can meet the needs of safety and real-time in automatic driving.(2)Aiming at the problem that the license plate target is easy to appear fuzzy and difficult to be recognized when shooting in complex environment,the residual network with strong extraction ability of license plate features is selected to achieve target classification,so that the original network has a certain multi-scale representation ability.The improved YOLOv3-Res algorithm is used to detect license plates in complex environment.The results show that the detection accuracy of the algorithm is improved by 13% compared with the original network,and the Super-Resolution(SR)reconstruction effectively improves the detection accuracy.(3)Aiming at the problem that YOLOv5 s algorithm model is difficult to complete multi-target recognition task in small target detection,Adam is selected as the optimizer.The Density-Based Spatial Clustering of Applications with Noise(DBSCAN)is used to improve the network.Training was conducted on a TT100 K dataset containing up to 221 traffic sign categories of information.The experimental results show that the m AP(Mean Average Precision)of YOLOv5s-Mobile Netv2 algorithm is improved by 12.9%,and the multi-class small target recognition task is completed well.The innovation of this paper mainly includes two points: first,YOLOv3 algorithm is improved by Residual module,and Deep Recursive Residual Network(DRRN)is used in the pre-processing stage,so as to improve the clarity of fuzzy image and enhance the license plate feature information;Secondly,Mobile Net V2 network is used to replace the backbone network CSPDark Net53 in YOLOv5 s algorithm.The training results show that the number of network parameters is reduced by 65.6%,and the calculation amount of model is reduced by 59.1%,which effectively reduces the complexity of the algorithm.