DFALaneNet: A Dynamic Real-Time Lane Detection Network Based on Adaptive Scheduling

초록

Lane detection is an important aspect of autonomous driving. For real-world applications, both accuracy and processing time are very important for this technology. Thus, this study proposed a dynamic real-time deep learning-based lane detection method based on the adaptive scheduling of input frames. An adaptive scheduling network was proposed to separate the input into key and adjacent frames based on the expected confidence. Consequently, an improved LaneNet encoder-decoder model was adopted as the primary image segmentation network, which was used to perform lane detection on a keyframe to ensure high accuracy and robustness. A neural optical flow network was used to perform lane predictions on adjacent frames to improve the speed of the detection network. We evaluated our proposed method on TuSimple, UCLane, and our campus datasets to demonstrate its effectiveness. The experimental results indicated that the proposed method performed well on both challenging public and self-made campus datasets, with substantial improvements compared with certain well-known methods.

키워드