TCLaneNet: Task-Conditioned Lane Detection Network Driven by Vibration Information

Lane detection is an important task in autonomous driving. However, it poses great challenges in occlusion and low-light conditions. To deal with these problems, we propose to utilize the vibration signals generated when vehicles pass over the vibration marking lines as supervision for lane occlusion prediction, whose features are then used to adaptively adjust the weights of lane detection network to improve its performance. Specifically, we develop a Task-Conditioned Lane Detection Network (TCLaneNet) that consists of the main lane detection branch and an lane occlusion classification branch. To effectively refine the lane detection features, Conditioned Adaptation Modules (CAMs) are proposed to obtain the conditioned parameters (i.e., channel weights) from the feature embedding of the lane occlusion classification task to adjust the lane detection features. In addition, as lane lines are elongated, while the receptive field of the convolution operation is spatially limited, we propose a Contextual Spatial Information Attention Module (CSIAM) to capture the global shape features of the lanes. Substantial experiments have been carried out on a vibration-based lane detection dataset VBLane that we constructed. The dataset consists of synchronously collected visual images and vibration signals, as well as the labels for the lane lines. Experimental results show that the proposed TCLaneNet achieves an Intersection over Union of 94.29%, which is better than the state-of-the-art lane detection methods. In addition, the TCLaneNet is lightweight (0.84 M) and runs in real-time (43.44 FPS), and the vibration information is not needed in the inference phase.