Time to Lane Crossing Estimation Using Deep Learning-Based Instance Segmentation Scheme

The main goal of this paper is to develop a deep-learning-based instance segmentation of lane detection for time to lane crossing (TLC) estimation. As is well known, the TLC strategy can build lane departure warning systems (LDWs) with enhanced predictive capabilities which play the role of security early warning in order to prevent dangerous driving situations. LDWs quite relies on the accurate detection of the road markings. Instance segmentation is a task to detect specific objects in an image and create a mask around the object. We use an evolved fully convolutional networks (FCNs) based instance segmentation to yield the high-quality segmentation maps with finer details. It is an effective approach to providing us the high accurate lane detection for LDWs. Time to lane crossing (TLC) is a kind of lane departure indicator in LDWs. For TLC computation, both straight and curved vehicle paths are considered. Thus, the road curvature is an important variable in computations. In this paper, a lane mark fitting algorithm based on inverse perspective mapping (IPM) and sliding window technologies are proposed to compute the road curvature. TLC is duration available for the driver before any lane boundary crossing. The experimental results demonstrate that the proposed deep learning-based instance segmentation can significantly detect the lane. The experiment results illustrate that the proposed TLC estimation strategy can significantly enhance the lane marks recognition accuracy and then provide drivers with timely warnings to avoid unintentional lane departure.