Detecting the Presence of Stationary Objects from Sparse Stereo Disparity Space

This paper presents a novel approach to detecting the presence of objects in a scene from the 3D sparse disparity space obtained by stereo matching. The use of stereo imaging makes the proposed method particularly useful for detecting stationary objects without the need of learning the appearance patterns on an object or the background. Our approach is based on the fact that sparse image features on an object exhibit cluster structures in the 3D disparity space and this reveals the presence of the object. Hence, we propose to use spectral clustering for grouping matched Scale Invariant Feature Transform (SIFT) interest points in the disparity space and to automatically determine the number of groups and their positions. For grouping matched edge points in the disparity space, a Gaussian mixture model is proposed for its computational efficiency. Experimental results show that our proposed methods can accomplish the task well.