SUBSPACE SEGMENTATION ALGORITHM FOR PASSABLE REGIONS DETECTION BASED ON SUPERPIXEL

The problem of passable regions detection on the images is an important problem in automatic driving vehicle, machine learning, and computer vision area. In this paper, this problem is converted into a subspace segmentation problem without labeled samples. To obtain the more helpful and effective clustering result, a novel affinity graph algorithm is proposed which could adaptively balance the sparsity and grouping effect by trace lasso. Moreover, we enforce the coefficient vector non-negative in order to obtain a parts-based representation. Our empirical study shows notable performance improvement of the proposed algorithm comparing with the state-of-the-art algorithms on real word problems. The two main technical contributions of the proposed approach are that a new fusion feature is designed for superpixel and a novel subspace segment graph contrived by trace lasso with non-negative constraint rules is proposed to express the relationship of superpixels in outdoor images, which could cluster the passable regions better than other likely graphs. The proposed method has been implemented, and experiments with three group unstructured environment images demonstrate that it is effective at detecting passable regions in challenging conditions and illuminations.