Multitask Knowledge Distillation Guides End-to-End Lane Detection
Huihui Pan, Xuepeng Chang, Weichao Sun
Abstract
Autonomous driving has witnessed rapid development with the application of artificial intelligence technology in recent years. Lane detection is one of the tasks of environment perception, which affects the planning and decision-making directly, and requires the algorithm to meet both high precision and high efficiency. Most of the existing methods extract pixels belonging to lanes in the image, which should be postprocessed, otherwise it cannot be applied to subsequent tasks like planning. This article proposes an end-to-end lane detection method that utilizes auxiliary supervision and knowledge distillation based teaching-test module to predict the parameters of polynomials of lanes directly. The teaching-test module guides the polynomial regression branch to learn the shape features from the segmentation branch to improve the fitting accuracy under complex road conditions. The proposed method is validated on TuSimple and CULane datasets, and is competitive with state-of-the-art methods in efficiency and accuracy.