A Lightweight Network with Lane Feature Enhancement for Multilane Drivable Area Detection

Abstract

Detecting drivable areas on multilane roads efficiently and accurately is still a tricky problem for autonomous driving. To better address this issue, we present an encoder-decoder network with a lightweight backbone and a lane feature enhancement module in this paper. Specifically, the lightweight backbone built with D-factorized convolutions helps to improve the speed of extracting drivable area features and reduce the number of parameters. The lane feature enhancement is realized by the non-local block at the high-level semantic stage, enhancing the features of the drivable areas such as lane line, direct lane, and alternative lane according to the similarity with the surrounding pixels. By compressing the decoder, the running speed of the proposed model is further improved without losing accuracy. A series of comparative experiments on the BDD100K dataset demonstrated that the proposed model has a better trade-off between speed and accuracy for multilane drivable area detection than the existing state-of-the-art models.