Research On Vision-Based Environment Perception Algorithm For Autonomous Driving

In the autonomous driving system,environmental perception is the basis and prerequisite for other modules,and it is the prerequisite for autonomous driving vehicles to travel safely.The visual perception technology based on cameras has unique advantages in environmental perception.The key technologies for visual perception mainly include the detection of moving objects such as pedestrians and vehicles,detection of lane lines,and distance estimation based on binocular vision.Among them,the detection of moving objects such as pedestrians and vehicles as well as lane line detection require algorithms with high real-time performance and strong robustness.Therefore,this topic is devoted to the research of correlation algorithms in visual perception.(1)A moving object detection algorithm based on improved YOLOv5 s is proposed.In view of the problems in moving object detection in the field of automatic driving,such as mutual occlusion of objects,serious omission and misdetection of small objects,poor real-time performance and robustness,the YOLOv5 s algorithm is improved reasonably.In the improved model,the detection scale is adjusted,the global attention mechanism is embedded into the residual block in the C3 structure of the backbone network,and the SIo U bounding box regression loss function is used.A number of experiments were carried out on SODA10 M dataset to verify the effectiveness of the improved algorithm.The experimental results show that the detection accuracy of the improved algorithm is greatly improved,and the detection effect is better in complex scenes with occlusion,multi-object,small object and dark environment.The improved model was combined with multi-object tracking algorithm DeepSORT,and the detection effect was verified on dashcam images.The results show that the model has good tracking effect and few ID-switch.(2)A lane detection algorithm based on improved LaneNet was proposed.The lane detection model LaneNet was improved to solve the problems of obstacle occlusion and fuzzy marking in the field of automatic driving.In the improved model,Bise Netv2 semantic segmentation network was used as a shared feature input network for binary segmentation branches and pixel embedded branches.Position attention mechanism was introduced into Bise Netv2 semantic branches,and DBSCAN clustering algorithm was used to complete the lane line pixel clustering task.The validity of the improved algorithm is verified on Tu Simple data set and CULane data set.The experimental results show that the accuracy of the improved model is higher than that of the original model and better than other mainstream models.The improved model can deal with a variety of complex environments and has good robustness and anti-interference.Finally,the improved YOLOv5 s model was integrated with the improved LaneNet through Open CV to achieve the effect of detecting road objects and lane lines simultaneously.(3)A object ranging system based on binocular vision is established.A binocular camera model KS4A418-D was selected to build the binocular stereo vision experiment platform in this paper.After camera calibration and binocular correction,SGBM stereo matching algorithm was used for pixel matching,and ranging experiment was conducted.According to the experimental results,a method was proposed to characterize the distance between objects and binocular camera by depth estimation.Finally,combined with the improved YOLOv5 s algorithm proposed in this paper,road objects were detected and ranging.