Vision-Based Lane Detection and Lane-Marking Model Inference: A Three-Step Deep Learning Approach

In many advanced driver-assistance systems (ADAS), lane detection is often necessary. Vision-based lane detection is popular because of its cost efficiency, but it can be easily affected by illumination changes, especially abrupt ones. Moreover, since most camera systems have a very limited angle of view (AOV), a single camera ADAS can only perceive a portion of a highly curved road. This introduces another challenge to ADAS when fitting lane models. In this paper, we propose a method for lane model inference, which uses one of the two lane-markings if there is only one lane-marking can be seen; or even, using lane-marking models from previous moments if there are no lane-markings to be seen at the current moment. In addition, we also propose using deep neural networks (DNN) to reduce noise at feature extraction stage. We use two DNNs in our method: a YOLO network for detecting an removing vehicles from images; a CPN network for detecting road surfaces in order to remove noises that are not on road surfaces. We tested our method on a video in which the roads are mostly curved and the lighting conditions can change very fast. We use the distances between our lane-marking models and the ground truth to evaluate our method. We see some big improvements in scenarios where the scene suddenly becomes very bright and where the road has a very high curvature.