Tiny Obstacle Discovery by Occlusion-aware Multilayer Regression.

Edges are the fundamental visual element for discovering tiny obstacles using a monocular camera. Nevertheless, tiny obstacles often have weak and inconsistent edge cues due to various properties such as small size and similar appearance to the free space, making it hard to capture them. To this end, we propose an occlusion-based multilayer approach, which specifies the scene prior as multilayer regions and utilizes these regions in each obstacle discovery module, i.e., edge detection and proposal extraction. Firstly, an obstacle-aware occlusion edge is generated to accurately capture the obstacle contour by fusing the edge cues inside all the multilayer regions, which intensifies the object characteristics of these obstacles. Then, a multistride sliding window strategy is proposed for capturing proposals that enclose the tiny obstacles as completely as possible. Moreover, a novel obstacle-aware regression model is proposed for effectively discovering obstacles. It is formed by a primary-secondary regressor, which can learn two dissimilarities between obstacles and other categories separately, and eventually generate an obstacle-occupied probability map. The experiments are conducted on two datasets to demonstrate the effectiveness of our approach under different scenarios. And the results show that the proposed method can approximately improve accuracy by 19% over FPHT and PHT, and achieves comparable performance to MergeNet. Furthermore, multiple experiments with different variants validate the contribution of our method. The source code is available at https://github.com/XuefengBUPT/TOD_OMR.