Using Temporal Information in Deep Learning Architectures to Improve Lane Detection Under Adverse Situations

Abstract

One of the fundamental challenges in the field of autonomous driving is the ability to detect dynamic objects, such as vehicles or pedestrians, and statics ones, such as lanes, in the surroundings of the vehicle. The accurate perception of the environment under the long tale of driving scenarios is crucial for a safe decision making and motion planning.

Mainly, lane detection approaches still function on single-frame basis and do not exploit the (high) temporal correlation of the signals representing the perceived environment. Single-frame detection networks might work well under circumstances where the lanes are perfectly visible, but show a lack of performance under certain situations, like occlusions, shadows, rain, snow, lane degradation, etc. To address the aforementioned problem, this work proves how adding temporal information for lane binary segmentation improves substantially the performance of single-frame architecture under challenging and adverse situations.