Decoupled Keypoint Regression for Real-Time Free-Space Boundary Detection Using Surround-View Fisheye Cameras

Free space detection is one of the fundamental tasks for automatic parking assist. In this task, the state-of-the-art network identifies the precise surround-view drivable space for real-time vehicle control. In this paper, we propose a network based on the surround-view fisheye image (FSVODNet) inputs to detect the concerned static and dynamic object for subsequent free-space construction. Which is different from the common vision-based method taking the around view monitoring (AVM) image as input, our approach cuts out the stitching step and gains a remarkable computational cost reduction for post-process. Experiments on our parking scene dataset demonstrate the outstanding localization accuracy of our network. It could run in realtime and support the whole automatic parking assist system.