Reliable Multilane Detection and Classification by Utilizing CNN as a Regression Network

Abstract

Reliable lane detection is crucial functionality for autonomous driving. Additionally positional information of ego lanes and side lanes is pivotal for critical tasks like overtaking assistants and path planning. In this work we present a CNN based regression approach for detecting multiple lanes as well as positionally classifying them. Present deep learning approaches for lane detection are inherently CNN semantic segmentation networks, which concentrate on classifying each pixel correctly and require post processing operations to infer lane information. We identify that such segmentation approach is not effective for detecting thin and elongated lane boundaries, which occupy relatively few pixels in the scene and is often occluded by vehicles. We pose the lane detection and classification problem as CNN regression task, which relaxes per pixel classification requirement to a few points along lane boundary. Our networks has better accuracy than the recent CNN based segmentation solution, and does not require any post processing or tracking operations. Particularly we observe improved robustness in occlusions and amidst shadows due to over bridge and trees. We have validated the network on our test vehicle using Nvidia’s PX2 platform, where we observe a promising performance of 25 FPS.