Research On Road Scene Perception Algorithm Based On Keypoints

In recent years,autonomous driving technology has received widespread attention as an important landing scenario in the field of artificial intelligence.And road scene perception technology,as the basic module and core module of the autonomous driving system,has become a research hotspot.The great success of key point detection technology in the field of computer vision also provides a new research direction for the road scene perception of autonomous driving.Therefore,based on the key point detection network,it is of great significance to develop an algorithm with high accuracy,strong robustness and high real-time performance to detect lane and vehicle.First of all,in order to solve the problem that the lane detection algorithm based on semantic segmentation has a poor real-time performance,a lane detection algorithm based on key points is proposed in this paper.We use the Hourglass network as the basis to improve its down-sampling module,size-maintaining module and up-sampling module,and the three output branches of the network are constrained by designing three loss functions: confidence loss,offset loss and feature loss.The lane post-processing mainly includes the use of RANSAC algorithm for curve fitting and the use of point-by-point matching post-processing algorithm to solve the problem of lane misdivision.For the situation where the lane is blocked and worn,we use dilated convolution with different dilation rate to improve the downsampling module in the network structure design.And through experiments on the Tusimple dataset,it is proved that compared with the algorithm that does not use dilated convolution to improve the network structure,the accuracy of the improved algorithm is increased by1.17%.In the detection scene where lane is blocked and worn,the accuracy of the improved algorithm has increased by 5.56% and 4.21% respectively.Secondly,aiming at the needs of vehicle detection task,We design a vehicle detection algorithm based on improved CenterNet by studying the object detection network.First,the residual module of the feature extraction network Hourglass-104 is improved,and the original basic convolution is replaced by the depthwise separable convolution.Experiments show that the amount of network calculation of the improved algorithm is greatly reduced,and the detection speed FPS is increased by 5.8.Then the attention module CBAM is added to a specific location of the network to improve the detection accuracy of the algorithm when the vehicle is occluded or the object is small.Finally,in order to balance the complexity and performance of the model,the CBAM module is optimized using one-dimensional convolution.Experiments show that the algorithm with the improved CBAM module can increase the average accuracy by 4.3% when the detection speed is equivalent.Finally,we train and test the road scene perception algorithm designed in this paper.First we introduce the experimental environment construction,datasets and the evaluation indicators used in this paper.Subsequently,We accomplish a number of comparative experiments on the lane detection algorithm and the vehicle detection algorithm,and analyze and evaluate the detection effect.Finally,the lane detection algorithm has a detection accuracy of 95.75% on the Tusimple dataset,and an F1 score of 70.8 on the CULane dataset.It can achieve the FPS of 43 on the NVDIA Ge Force GTX 1070 Ti,which exceeds the requirement that the lane detection algorithm detects more than 25 frames per second when the vehicle speed is between 80 km/h and 120km/h.The vehicle detection algorithm based on the improved CenterNet achieved an average accuracy of 87.9% on the KITTI dataset.Compared with the original CenterNet algorithm,the accuracy is increased by 3.3%,and the detection accuracy of small targets is increased by 39.4%.And its FPS increased by 5.7 to 15.9.The experimental results prove the effectiveness and feasibility of our algorithm in autonomous driving task.