Research On Road Traffic Marking Extraction Algorithm Integrating Image And Point Cloud

In recent years,with the development of intelligent transportation and the continuous improvement of autonomous driving technology,the production and production process of high-precision map in the industry has become more and more standardized.Because of this,high-precision map has become an essential condition for intelligent transportation,automatic driving and other technologies.The highprecision map can be regarded as a comprehensive map composed of dynamic data layer and static data layer.The dynamic data layer includes real-time road condition layer and traffic event layer,and the static data layer includes data update layer,traffic facilities layer,lane layer and road layer.Road traffic line marking is an important element of high precision map.It is precisely because of its correct and complete expression on high precision map that drivers or autonomous vehicles can accurately identify and drive safely.Road traffic marking can be roughly divided into linear road traffic marking and planar road traffic marking.At present,the recognition and segmentation of planar traffic marking is based on the difference of reflection intensity between planar traffic marking and road surface in high-thread lidar point cloud.However,such high-thread lidar is expensive and has a large amount of point cloud data,which may affect the production efficiency of high-precision map and the subsequent update efficiency of map due to the small number of equipment and long time of data processing.Therefore,this paper uses the laser radar with low cost and low thread to obtain the point cloud data,segment the road surface point cloud for the road with or without curb,and then fuse it with the image processed by the improved mask r-cnn algorithm in this paper,and use the fused image to extract the coordinate information of road traffic markings.(1)In the part of image segmentation,this paper improves the mask r-cnn neural network framework to enable it to recognize linear and planar road traffic markings.Lane line detection using neural network has become increasingly mature,such as RCNN,SPP-NET,Lane Net,etc.,which can complete lane line detection with high accuracy and short time compared with the traditional algorithm,but this neural network can only detect and identify linear road traffic lines in the image.In order to improve the detection and recognition ability of the neural network,this paper improved the Mask R-CNN neural network framework,added the fusion attention module,and adjusted the screening algorithm of the prediction box to enhance the target recognition ability of the network.Next,some images were selected from well-known public data sets to form experimental data sets and tested in VGG-SS,Deep Lab3+,traditional Mask R-CNN network and this paper’s network.Finally,Precision,Recall and F1 Score are used to compare and evaluate the experimental results.The results prove that the improved neural network structure in this paper can recognize and extract road traffic lines under different road conditions.In this paper,for different types of roads,road surface cloud extraction algorithm is designed for roads with or without kerbs.Firstly,the road point cloud integrated radius filtering algorithm and point coordinate information are used to preprocess the point cloud data and eliminate the remote noise points.After pretreatment,the road point cloud was extracted according to whether there was obvious kerb separation.For the road point cloud with kerb,the elevation mutation characteristic of point cloud at kerb is used to calculate the projection distance from point cloud to plane model to express the undation degree of road surface.The appropriate threshold is determined to judge the abnormal point cloud block and the kerb point is found according to the Zaxis histogram to segment the road surface.For roads without kerbs,this paper establishes fan-shaped grids according to the characteristics of lidar point clouds,and filters the ring-shaped areas according to scan lines.The regularity that each scan line of road point cloud is approximately round is mainly used to calculate the Euclidean distance between points and the center in different regions and set the threshold.Those beyond the threshold range are considered as non-road points and eliminated.Finally,the point cloud is segmented to obtain the complete road point cloud.Experimental results demonstrate the effectiveness of the proposed method.For the part of road traffic marking extraction based on the fusion of image and point cloud,this paper collected road images and point cloud data of the experimental site by a mobile measuring vehicle equipped with lidar and monocular camera.For the image data acquired by the camera,the improved neural network is used to process the image,and the recognition results are annotated to the original image.For the point cloud data obtained by lidar,the road point cloud of each frame can be obtained according to the above point cloud processing algorithm.Then,the camera and lidar are calibrated to make the point cloud projected onto the image,and the point cloud is cut according to the target information of the image.Finally,through the corresponding matrix transformation,the 3d lidar coordinates of the road traffic marking line on the experimental road are obtained.