Understanding the surrounding environment of the vehicle is still one of thechallenges for autonomous driving. This paper addresses 360-degree road scenesemantic segmentation using surround view cameras, which are widely equipped inexisting production cars. First, in order to address large distortion problemin the fisheye images, Restricted Deformable Convolution (RDC) is proposed forsemantic segmentation, which can effectively model geometric transformations bylearning the shapes of convolutional filters conditioned on the input featuremap. Second, in order to obtain a large-scale training set of surround viewimages, a novel method called zoom augmentation is proposed to transformconventional images to fisheye images. Finally, an RDC based semanticsegmentation model is built. The model is trained for real-world surround viewimages through a multi-task learning architecture by combining real-worldimages with transformed images. Experiments demonstrate the effectiveness ofthe RDC to handle images with large distortions, and the proposed approachshows a good performance using surround view cameras with the help of thetransformed images.
展开▼
