Free Space Estimation on Nonflat Plane Based on V-Disparity

Recent progress in free space estimation provides precise segmentation between the ground and obstacles on flat plane. However, it remains challenging on nonflat plane, especially with the varying latitudinal and longitudinal slope or in the case of multiground plane. This letter presents a well-integrated framework for free space estimation in this challenge. Our approach couples an improved V-disparity with a proposed confidence map in a probabilistic fashion. The improved V-disparity representation adopts the sliding windows paradigm and a new V-disparity filter is designed. The proposed confidence map represents the ownership of pixels like occupancy grids, but is built only using the information of the disparity map. The free space estimation is implemented in the confidence map with the global optimization paradigm of dynamic programming. We demonstrate our superior performance compared to two other methods from the literature on a manually labeled dataset from KITTI's object detection benchmark.