Deeply Supervised Z-Style Residual Network Devotes to Real-Time Environment Perception for Autonomous Driving

The visual environment perception plays an irreplaceable role for autonomous driving. Wider and deeper convolutional neural network (CNN) architectures are explored in most approaches which focus on improving performance. However, for real-world applications, the environment perception need to be accurate enough to ensure safety and fast enough to guarantee prompt control. Toward this goal, we propose a deeply supervised Z-style residual network (DSZRNet) to join object and lane detection via a unified architecture, named DSZRNet, which consists of three main parts: 1) the Z-style structure commits to fuse the feature information of shallow and deep layers for effective and robust representations; 2) the residual net is to gradually improve the location accuracy; and 3) the deeply supervised lane detection branch is grafted on the shared convolutional layers by adding only a little computational burden. Our DSZRNet is simple and can be trained end-to-end. To effectively evaluate the proposed approach, a large-scale China urban traffic (LSCUT) dataset is collected which contains objects that have a large intra-class distance because of the large-scale variance, occlusions, etc, and a variable number of lanes. Besides, the real-world computer vision KITTI dataset is also used for performance evaluation. The experimental results we obtained demonstrate that the DSZRNet gains the best tradeoff between the detection speed and detection accuracy than the other published works, and it is well suited for environment perception for autonomous driving. The download links of the LSCUT dataset will be at: https://github.com/DaiCV/LSCUT.