Lane Detection Based on Instance Segmentation of BiSeNet V2 Backbone Network

Abstract

Most lane line detection algorithms still have room for improvement in detection accuracy, speed, and robustness. Meanwhile, these algorithms only test the performance indicators through the test set of the open-source dataset rather than deploying them on actual vehicles and evaluating the performance indicators through road scenarios. Therefore, this paper proposes a lane detection algorithm based on instance segmentation. Firstly, a dual-branch neural network model for lane line image segmentation was designed based on BiSeNet V2. Then the discrete lane line feature points are operated through the clustering model. The corresponding feature points are selected for fitting by combining straight lines and curves to obtain the appropriate fitting parameter equation for the specific visual field area. Finally, the model is trained and verified based on the TuSimple dataset. The algorithm has a noticeable performance improvement under the two evaluation indicators of mIoU and FPS. Meanwhile, the model is integrated into the ROS task platform for intelligent vehicles. The results show that the algorithm’s accuracy and detection speed are increased to about 3.9 and 2.9 times, respectively, that of the improved probabilistic Hough transform algorithm under the two evaluation indicators of lateral distance and the detection time of each image frame.