Research On Lane Detection And Tracking Algorithm Based On Structured Road Segmentation

With the development of driverless technology,it is very important to improve the safety for practical applications.Specially,lane detection is an important foundation of unmanned driving technology.In this thesis,some research about the lane detection with the road segmentation in the structured road scene is done as follows.Firstly,the road segmentation algorithm is studied.DeepLabv3+ network is modified as follows.VGG is selected as the feature extraction basic network via removing the fully-connected layer,an ASPP structure is added to extract the multi-scale feature information after the pool5 layer,and a skip connection structure is used to fuse both shallow and high layer information.The improved algorithm is tested on the KITTI dataset with an average accuracy of 92.37% and a single-image segmentation time of 83.39 ms.Secondly,the lane detection algorithm is studied.The SCNN network structure to capture spatial relationships is modified by replacing the classification sub-network of each lane line with a post-processing algorithm to simplify the network model.In addition,the weighted least squares method is used to fit the lane line candidate points from the SCNN network.The test results on the CULane dataset show that the comprehensive index F1 about the accuracy and recall rate of the improved algorithm reaches 71.3%,0.9% higher than that of SCNN network.Then,a multi-task learning algorithm combined lane detection with structured road segmentation is studied.The parameter hard-sharing mechanism is adopted with the improved DeepLab network as road segmentation sub-network and the improved SCNN network as lane detection sub-network whose the hole convolution layer is replaced by ASPP structure.Both the road segmentation and the lane detection subnetwork share the same feature extraction network and are connected by a coding structure.The experimental results on the self-labeled dataset of CULane dataset show that the multi-task learning network increases the average accuracy of the road segmentation network by about 1% and the F1 index of the lane detection network by 0.4%.Finally,the road and lane tracking algorithm is studied.An improved tracking algorithm combining Camshift special in color feature tracking with Kalman good at motion state tracking is proposed.And the change of relative position between lane and road is used to judge a successful tracking.The results on the actual captured test video show that the improved algorithm accuracy of road and lane tracking is 92.6% and 89.0%,respectively and the average speed 47.5 FPS.In summary,this paper researches the lane detection and tracking task combined with road segmentation in unmanned driving,and proposes a solution based on multi-task learning deep learning with an improved tracking algorithm.The experimental results prove that the improved lane detection and tracking algorithm has a good accuracy and real-time performance.