Research On Multi-lane Detection Based On Multi-task Learning

The rapid development of intelligent cars improves the security and efficiency of human travelling,while the stable and reliable road detection with multi-lane perception contributes to a safe driving.Factors such as the accuracy of lane detection and the algorithm robustness to various environment make the lane-level detection full of challenges.Currently,the popular road detection algorithms,such like road segmentation and lane markings detection,are still facing some severe issues.Hence,this paper combined two of the most popular road detection algorithms,and proposed a frame of multi-lane detection based on road segmentation and lane detection,namely,multi-task learning.Optimization was performed aiming at the shortcomings of each task.In the end,two results are fused to derive the accurate and stable multi-lane detection result.This approach is of great value for the research and development of intelligent cars.For the road segmentation task,the network is always lack of consideration about features of road.In this paper,a road segmentation based on prior-knowledge is proposed.It is mainly an image segmentation network of encoder-decoder structure.Focus on the space and edge gradient,the corresponding features are extracted as the prior-knowledge: the road data distribution of images in a large scale dataset as the space prior-knowledge;calculating the 8directions of gradient in 8 neighborhood as the edge prior-knowledge of the primitive space.They are fused with the network to enrich the information.Also,this paper introduced the road distribution probability based on the cross-entropy loss function.In order to improve the effect of road segmentation,larger punishments are given to the missive detection in a probable road area and false detection in a not probable road area.Experiments were made on the Cityscape data,which proved the proposed approach is of high accuracy and low omission rate.For the lane detection task,the robustness of traditional methods is bad,and deep learning methods based on binary segmentation can not tail the different lane instance.In this paper,a detection algorithm based on instance segmentation is proposed.Firstly,lane markings binary segmentation is performed.Then,pixels belong to the lane marking are mapped to the feature space by a convolutional layer.During the training,the Discriminative loss function divides pixels of different lanes,and clusters pixels that belong to the same lanes.Meanshift algorithm is implemented to cluster the feature pixels in the feature space,in order to get different element instances.Finally,the results of lane instance can be derived by remapping the feature point back to the image space.As for the lane fitting,an improved Ransac fitting algorithm is proposed.It converts the source image into a plan view using inverse perspective transform,and performs the lane instance fitting in the plan view.Experiments are accomplished on the TuSimple and CULane data to prove that this lane detection algorithm is robust in complex traffic environment and light changing scene.For the multi-lane detection,this paper introduced a multi-task learning approach based on the road segmentation and lane marking detection.Two tasks are executed at the same time.They share the same encoder of the network which is used for extracting the feature map,but own their decoders independently.To derive the multi-lane results,this paper proposed a fusion algorithm that stably obtain the ego-lane,the left lane and the right lane.Experiments showed that this algorithm segments the multi-lane area effectively,and has adapting perception ability to road with various numbers of lanes.