Research And Implementation Of Forward Visual Traffic Safety Field Based On Single Vision Sensor

The ability of environment perception has always been the bottleneck restrict-ing the development of automatic driving technology.The environment percep-tion scheme based on single vision sensor has a high landing advantage,and has been a hot topic in academic and industrial research.Single vision sensor envi-ronment perception mainly includes target detection and segmentation,in which 3D obstacles and lanes are important targets.Acquiring the three-dimensional information of the forward obstacles is conducive to the autonomous avoidance of vehicles,and the perception of the lane is conducive to the route planning of the lane.At the same time,the 3D obstacles can be located in different lanes through lane detection and segmentation,and the threat classification can be carried out according to the lanes,so as to filter out a large number of invalid targets.The current difficulties lie in:(1)how to integrate 3D object detection and instance segmentation into one framework and solve two tasks at the same time;(2)how to improve the speed of obstacle instance segmentation to achieve the speed level of 3D object detection.Aiming at the above problems,this paper proposes a real-time framework,which can segment and detect 3D objects simultaneously through depth dimen-sion,and set lane detection and lane segmentation as parallel tasks to improve the speed of detection and segmentation.The following aspects are the main work of this paper(1)3D target detection is divided into four sub tasks,2D detection,instance level depth estimation,3D position estimation and corner regression.3D detec-tion is carried out indirectly by solving sub tasks,and 3D information of target is output.The experimental results show that after splitting the task,the subtasks can be completed in parallel,which greatly improves the detection speed and ensures the detection accuracy compared with other two-stage and binocular methods.(2)A case segmentation method based on spatial discretization is proposed.The instance task is divided into pixel level depth estimation task and instance level depth estimation task.Finally,pixels are clustered into each instance through post-processing.The experimental results show that the speed of the method has been greatly improved.(3)A method of lane line detection and pixel segmentation is designed,and the method is also designed The binary detection of lane line is combined with the segmentation result of lane pixel instance to determine the clear lane boundary line.Finally,the discontinuous part is fitted i nto continuous l ane l ine,and the different lane obstacles are classified by lane line.