Post-calibration Fusion And Obstacle Ranging Based On Laser Point Clouds And Images

Lidar and visual images have gained more and more applications in environmental perception tasks.Breakthroughs have been especially made in intelligence devices such as unmanned vehicles.The use of laser point clouds,images collected by laser radar and visual sensors and their fusion data for obstacle detection,distance measurement,and surrounding environment reconstruction are the important basis for the environmental perception and the correct decision making of unmanned vehicles.However,the existing laser point cloud and image fusion algorithms require manual calibration of the sensors,the obstacle distance measurement based on binocular vision is slow and less accurate,and the obstacle recognition rate is low when using laser point cloud ranging.For these reasons,this thesis has carried out the following three aspects of research:(1)To solve the problem that the existing algorithms require manual calibration for fusing laser point clouds and images,this thesis proposes a post-calibration fusion algorithm based on the reference objects for the point laser point cloud and the image.The algorithm first generates an ideal projection matrix based on the pinhole camera model and the positional relationship between the laser radar and the camera.Then,taking advantage of the point cloud characteristic that the edge can be easily detected,the projection matrix is corrected using the reference object with the longer vertical edge.The horizontal coordinate offset is calculated using the pixels of the laser point edge and the image vertical edge,and then added to the projection matrix to get the corrected matrix.An accurate color point cloud can be obtained.The algorithm reduces the complexity of the laser point cloud and image fusion.According to the driving information of the unmanned vehicle collected by the inertial measurement unit,this thesis further fuses multi-frame color point clouds to obtain a three-dimensional color map.The experimental results based on the KITTI public datasets verify the effectiveness of the algorithm.(2)Due to the slow speed and low precision of obstacle distance measurement using binocular vision,this thesis proposes an asymmetric framework with the front vehicle as the obstacle,which processes the left and right images differently.In the right image,the region of interest(ROI)of the vehicle is detected by classical methods,including lane ROI setting,lane detection and vehicle identification.For the left image,this thesis proposes to use a priori measurement distance and stereoscopic imaging characteristics to predict the vehicle ROI,thereby saving operating time.In addition,in order to improve the distance measurement accuracy,this thesis propose to filter the feature points in the vehicle ROI with the disparity values,and select the feature points whose disparity values are close to the median to calculate the distance.Experimental results based on the KITTI public datasets show that the algorithm is 1.57 times faster than the symmetric method,and the relative distance error is reduced by 11.87%.(3)To avoid the low obstacle recognition rate in laser point cloud,this thesis proposes a distance measurement algorithm based on laser point cloud and image fusion.The algorithm first uses the post-calibration fusion algorithm proposed in this thesis to establish a mapping relationship between the laser point cloud and the image.Then,according to the position and distance of the obstacle detected by the binocular vision,the laser point cloud is screened to determine the points near the obstacle.Then,the K-means algorithm is used to cluster the filtered laser points,and finally the distance information is calculated using the clustered obstacle laser points.The experimental results based on the KITTI public datasets show that the proposed algorithm has less error and higher stability.