Lane detection is of great importance in autonomous driving systems since it provides the instance segmentation of lanes separated from the background image.
Recent deep learning models have been proposed to improve
the lane detection techniques to better address challenging
issues such as lane occlusion, low visibility, extreme weather
conditions, and so on. In this paper, we present a novel
approach, named Direction-aware Network (DANet), to further
advance current techniques for the detection of the curved
lanes in addition to the straight ones. A novel curvature feature
representation, which can efficiently describe most curve lanes
in real-world applications, is exploited to accurately locate curve
lane lines in the images. Besides, three main components are
interwoven to each other in our proposed DANet, including
Primary Convolutional Neural Network (P-CNN), Far-Near
Convolutional Neural Network (FN-CNN) and Bend Network
(BendNet). The first component, P-CNN, encodes the visual
features from the input image into a latent feature map, which
is then fed separately to both FN-CNN and BendNet. FN-CNN
then extracts horizontal and vertical feature maps, whereas
BendNet generates horizontal and vertical orientation weights
to weight the corresponding feature maps, allowing the network
to learn the curvature of the lanes. Finally, the combined feature
map is used to produce a probability map which in turn yields
the desired instance segmentation results. Experiments on the
CULane dataset [25] show that our proposed DANet delivers
outstanding performance on on curved lane detection as well
as comparable results on regular lane detection.
|
