Research On The Algorithm Of Road Target And Lane Line Detection Based On Deep Learning

Road target detection and lane line detection are the main technologies of intelligent driving system.Moreover,they have extremely important theoretical research value in smart city traffic system and other fields.Due to the many sorts of targets to be detected and complex scenes in the actual road target detection scene,the detection is inclined to the influence of the scene surroundings;In the lane line detection scene,the lane line will be affected by conditions.For example,incompleteness,damage,and bright light can lead to poor test results.This dissertation mainly studies the low detection rate in road target detection and the problem that the environment can easily affect lane line detection.It plays an essential role in the field of theoretical research and engineering value for intelligent drive system.After researching the modern domestic and foreign used detection algorithms for road target detection and lane line detection,the content of the research article is determined.In view of the problem of YOLOv3 network in the actual road target detection scene.there are many missed detections in target detection,and a feasible improvement plan is proposed.Firstly,the K-means++ clustering algorithm is used to replace the K-means clustering algorithm in the original network,and the number of anchors and the aspect ratio of the KITTI dataset are analyzed by clustering,so that the obtained anchor parameters are more applicable;Secondly,for purpose of improving the performance of the algorithm,the existing network output is upgraded,and a 104×104 feature detection layer is added,effectively reducing the phenomenon of feature disappearance;Finally,use the spatial pooling pyramid to perform different block pooling on the feature maps obtained by the 32-fold down-sampling,and extract a feature from each block as a dimension to ensure that the dimensions of the final features are consistent,thereby solving information loss and scale the problem of inconsistency.In terms of the SCNN lane line detection algorithm,an improved SCNN network is proposed to solve the problem of low lane line detection accuracy and recall rate in several scenes such as the strong illumination of the SCNN network in curved the lane line detection methods.The backbone network of this method still uses the SCNN slice convolution method.The difference between improved SCNN and SCNN is showed following: Firstly,After slice convolution,row convolution and column convolution extract feature information.Secondly,the feature information of multi-scale features is extracted.Thirdly,through up-sampling or down-sampling the feature information of different scales to obtain the same-dimensional feature map for matching and fusion;Finally,the required feature map can be obtained,the marking of the lane line is predicted network output test result graph.The improved network algorithm of road target detection YOLOv3 is compared with the original algorithm and several other current target detection algorithms.The results show that the improved YOLOV3 algorithm has greatly detection performance and meet the application requirements.Comparing the four lane line detection algorithms of SCNN network,improved SCNN network,Lane network and VPG network,the performance in the three evaluation indicators of F1,Precision and Recall verified that the improved SCNN algorithm has strong detection performance on the three data sets of Tusimple,Performance of the improved SCNN network in F1,accuracy and recall rate shows that the improved SCNN network has higher evaluation and BDD100 K and CULane.In short,the improved YOLOv3 algorithm and SCNN algorithm have good detection results in road target detection and lane line detection,respectively.