Research On Lane Detection Algorithm Based On Semantic Segmentation And Inverse Perspective Mapping

Autonomous driving is the direction of future development in both transportation sector and car industry,while lane detection is a key task for autonomous driving.In a practical automated driving system,the lane detection module is required to detect the position,the color and the type of lanes.Meanwhile,trajectory planning and decision making for autonomous vehicles are usually based on environmental grid maps.Therefore,the results of lane detection should be represented as grid maps.However,traditional lane detection methods usually detect the lane positions only,and do not recognize the color and the type of each lane.Besides,their detection results are usually represented in pixel coordinates,which suffer from perspective distortion and can't represent the actual lane positions on the ground plane.To this end,a research on lane detection algorithms is carried out based on semantic segmentation and inverse perspective mapping.This dissertation solves lane detection problems based on semantic segmentation,segmenting different lanes on an image into different regions and matching the label of each region with the type of the coresponding lane.This dissertation proposed a lane semantic segmentation neural network based on edge feature merging and skip connections.Firstly,a convolutional neural network for semantic segmentation is constructed basing on the mainstream encoder-decoder framework.Next,considered that the edge features are important in lane detection,an edge feature extracting subnetwork is parallel connected to the encoder,enhancing lane features by merging original feature maps with edge feature maps layer by layer.Finally,skip connections from the encoder to the decoder are implemented,and thus make a better use of hierarchical semantic information,merging feature maps with the same size in the procedure of upsampling.Besides,the annotations of a traditional lane dataset are modified,thus it can be applied to training and testing semantic segmentation neural networks.Experiments are carried out on the modified dataset,proving the structures of edge feature merging and skip connections can significantly improve the performance of the base network on lane semantic segmentation tasks.The proposed network is able to detect the color and the type of lanes as well as their positions.Under the condition of having enough computational resources,the proposed convolutional neural network can achieve realtime detection.Based on the results of lane semantic segmentation,lane grid maps are generated based on inverse perspective mapping.Because this transformation relies on the intrinsic and extrinsic parameters of the camera,the camera is calibrated and the accuracy of the parameters are verified.A region of interest is specified in each semantic segmentation result,according to the range of the grid map.Using inverse perspective mapping,lane grid maps are generated as the outputs of the lane detection module.