A Novel Vision-Based Framework for Real-Time Lane Detection and Tracking 2019-01-0690
Lane detection is of crucial importance in ADAS because various modules (i.e., LKAS, LDWS, etc.) need robust and precise lane position to locate themselves and traffic participants to plan an optimal routine or making proper driving decisions. While most of the lane detection approaches depend on great amount of pre-processing and various assumptions to get reasonable result, the robustness and efficiency are deteriorated. To address this problem, a novel framework is proposed in this paper to realize the robust and real-time lane detection. This framework consists of two branches, where canny edge detection and Progressive Probabilistic Hough Transform (PPHT) are introduced in the first branch to achieve an efficient detection. To eliminate the dependency of the framework on assumptions such as flatten road, deep learning based encoder-decoder detection branch, which leverages the powerful nonlinear approximation ability of CNN, is introduced to improve the robustness and contribute a precise intermediate result. Since the detection rate of the CNN branch is much slower than the feature-based branch, a coordinating unit is designed. The two branches also backup each other so that the system can be failure-tolerant. Finally, Kalman filter is applied for lane tracking. Experiment result shows that the proposed framework can achieve robust detection result under various driving scenario with more than 100 FPS which provides a good foundation for the development of the ADAS module.