SALane: Speed and Accuracy Balanced Detection of Degraded Lane Markings Based on Visual Images

As a fundamental technology in the field of autonomous driving, lane detection plays a critical role in various advanced driver-assistance systems, including lane keeping, lane departure warning, lane change assistance, and forward collision warning. However, due to vehicle load, lane lines are prone to damage, making lane line detection more difficult. This study proposes SALane, an improved row-based ultra-fast and lightweight lane line detection network aiming at balancing ultra-fast speed and accuracy. Specifically, the Inception-v3 network is employed as a feature extractor, where the input image is partitioned into a grid with significantly fewer cells than pixels via anchor points. This design enables efficient localization of lane lines directly on the grid. Furthermore, Otsu's method is utilized to automatically determine an optimal image binarization threshold. Additionally, a hybrid approach combining random sample consensus and least-squares fitting is introduced to enhance the robustness and accuracy of lane line modeling. Cross entropy is also used to attenuate the effect of imbalance between lane lines and background categories, and a simple calculation method of lane marking degradation rate is proposed. On the CULane benchmark, SALane achieves a total accuracy of 79.6% at a speed of 60.2 frames per second, which outperforms the compared algorithms by 10% in accuracy and demonstrates a well-balanced performance between precision and speed. Further, this study explores the effect of the degree of lane marking degradation on the detection results, and the algorithm performance is stable when the damage category is less than 3. However, as the number of damage categories increases to 4 and 5, the performance of the algorithm decreases sharply.