| Autonomous vehicles are an important direction for future vehicle development,and their technologies cover a wide range of areas,including key technologies in multiple fields such as vehicles,control,communications,and computers.In the urban road environment,the application of visual inspection,precise positioning and other methods can well meet the environmental perception needs of autonomous vehicles.However,in the crowded urban road conditions of vehicles,lanes are blocked by congested traffic,and vehicle positioning will also The multi-path effect of urban roads seriously affects its positioning accuracy.In the case of urban structured roads where vehicles are crowded,the visual sensor field of view is blocked.Conventional methods cannot meet the safe driving requirements of autonomous vehicles.Based on the vision detection technology,this paper estimates the lane geometry of the autonomous vehicle through the pose of other vehicles in front Information,and analyze the estimated results to provide reliable information for the subsequent driving path planning of the automatic driving system.The main research work is as follows:1.Front vehicle detectionThe Faster R-CNN algorithm and the YOLO v3 algorithm had be used to detect the vehicle in front,and the KITTI data set had be used to train these two target detection models.Install and calibrate the camera,collect images in the campus area,verify the two algorithms,analyze and compare them,select the algorithm whose detection accuracy and processing speed meet the requirements of subsequent 3D pose estimation,and provide 2D pose estimation The bounding box constraints also provide an experimental basis for the full text.2.Vehicle pose estimation based on 3D visionBased on computer vision technology,the pose information of surrounding vehicles had be detected,and size,center point,and heading angle are selected as the description parameters of surrounding vehicles and used as evaluation indicators.First,returned the 3D attributes of the preceding vehicle through the KITTI left-eye 3D data set;then used the previously detected2 D bounding box of the preceding vehicle to constrain the 3D bounding box to estimate the 3D bounding box of the surrounding vehicles to generate a 3D pose detection model,In order to obtain the pose parameters of the vehicle in front;finally built a crowded road scene through Prescan,and used the above detection method to detect the pose information of the vehicle.3.Road geometry information estimation model based on the pose information of the preceding vehicleTaken the detected pose information of the preceding vehicle as input,the geometric information of the road ahead of the autonomous vehicle had be estimated.The front view of the image had be transformed into a top view through inverse perspective transformation,thereby eliminating the influence of perspective transformation on distance calculation.Used the DBSCAN clustering algorithm to cluster the position information of the detected surrounding vehicles,and then analyze and select the polynomial fitting method to fit the position of the vehicle to obtain the vehicle trajectory fitting curve on the current road,thereby obtaining the front road Geometric information.4.Road curvature estimation model based on unscented Kalman filterWhen the front view had be blocked by the vehicle in front,only the 3D pose information of the first vehicle in front of the vehicle had be calculated.Use Prescan software to build a simulation environment,imitate a real crowded scene,place interfering vehicles in front of the vehicle,and set the curvature information of the road.By modeling the relationship between the vehicle,the preceding vehicle,and the road,the motion model and observation model of the system are obtained.Collect the own vehicle information of the vehicle,such as the speed of the vehicle,the yaw angular velocity,etc.,combine the relative information between the vehicle and the preceding vehicle,such as relative speed,relative angular velocity,etc.,and then use the Kalman filter framework to perform the collected information Filter processing to calculate the curvature parameters of the current lane.This paper verifies the effectiveness of the 3D pose detection algorithm through real-vehicle experiments.Crowded roads had be simulated using a simulation environment.When the front view had be not blocked or partially blocked,the road geometry information had be estimated by polynomial fitting method to obtain the current road fitting curve,and R-square could be maintained at [0.9,1] in the interval,the fitting result has a strong correlation with the real road curve;in the case that the front view of the autonomous vehicle had be blocked by the preceding vehicle,the road curvature estimation model based on the unscented Kalman filter also had the same road curvature as the real road.The curvature matches.The experimental results shown the effectiveness of this method.The research on this subject can provide the correct driving direction for autonomous vehicles on congested roads,which is conducive to future autonomous driving path planning and other related work,and also provides a reference for future research on similar subjects. |
