Depth Completion with Edge Optimization for Depth Maps from Lidar

Abstract

Depth completion is an important research field in self-driving which complements depths mapped from point cloud of Lidar. This paper shows a depth completion method combining traditional image processing and image optimization. In our method, a depth map is completed through a series of well-designed morphological dilation and filtering methods, and then is optimized referring to a RGB image and a confidence map. The method is simple, data independent and runs only relying on the CPU. It is evaluated on the challenging KITTI depth completion benchmark [20]. The result performs as good as IP-Basic and better than sparse CNN in depth accuracy. Furthermore, it optimizes the edges of objects in the depth maps, which has a greater help for image segmentation, obstacle perception or other tasks in self-driving.