Lane Line Detection And Tracking Based On Key Point Approach

Detection and tracking of lane lines are extremely important research topics in the field of visual perception of autonomous driving.Identifying the lane lines on the current road can provides a navigation path for automatic driving path planning that the vehicle can be driven safely and in control.At the same time,accurate lane line recognition is also one of the important basis for high-precision map.Aiming at the problem that the current lane line recognition algorithms are easily disturbed by environmental negative samples and poor robustness.Basic recognition categories of lane lines cannot provide sufficient navigation information for autonomous driving.In this paper,the road scene model is improved,new lane line categories are added,the lane line recognition algorithm based on U-Net model is improved.Meanwhile,the lane line recognition is realized based on key points to improve recognition performance.Finally,the lane line tracking algorithm is constructed by using Hungarian matching and Kalman filtering.The main research contents and conclusions of this paper are as follows:(1)Road data modeling and collection of dataset about lane lines.For the current lane line datasets,only four types of lane lines are labeled(the left lane line of the current lane,the right lane line of the current lane,the left lane line of the left lane,and the right lane line of the right lane),which cannot meet the requirements of real autonomous driving path planning.According to the requirements of typical scenarios,this paper adds new categories for lane line recognition.At the same time,the lane line data is collected in different scenes,and the cosine acquaintance is used to frame the collected video to ensure the difference between the data samples.(2)Lane line recognition based on U-Net.First,a channel attention module is added to U-Net semantic segmentation model to adaptively integrate multi-level semantic information and enrich the lane line scene level information extracted by the model.Then,the training of the model is completed through data enhancement and back propagation,and the precision and recall rate are used to evaluate the performance of lane line recognition on the test set.This model can identify lane lines more accurately,but this model is susceptible to interference from background negative samples,and it is difficult to meet the real-time requirements when it used in embedded devices.(3)Lane line recognition based on key points.In view of the low precision and high time complexity of lane line recognition based on U-Net,a key point-based lane line recognition scheme is proposed to avoid the imbalance between positive and negative samples and huge size of outputs.First,according to the perspective relationship,the key points of the lane lines are linearly sampled at different intervals on the vertical axis of the picture,and the model predicts the lateral offset values and their probability values corresponding to different types of lane lines on a fixed sampling line.At the same time,according to the slope trend between the key points,the geometric constraint loss function of the lane line is proposed to ensure the correct shape of the lane line.Compared with the lane line recognition model based on improved U-Net,the recognition effect and time complexity of this model are greatly improved,the F1 score was improved by 6%.(4)Lane line tracking based on Hungary matching.In order to obtain stable and usable lane line curve parameters,the parameters of the lane line are tracked in the vehicle’s coordinate system.First,coordinate of the lane line in the vehicle’s coordinate system is obtained through the camera model,and the detected key points are fitted to obtain the curve parameters of the lane line.Then this paper uses the Hungarian matching based on the envelope area of the two curves as the cost matrix to match and update the lane lines,and finally uses the Kalman filter to predict and update the tracked lane line parameters.