Obstacle detection using sparse stereovision and clustering techniques

We present a novel technique for localisation of scene elements through sparse stereovision, targeted at obstacle detection. Applications are autonomous driving or robotics. Given a sparse 3D map computed from low-cost features and with many matching errors, we present a technique that can achieve localisation in a real-time context of all potential obstacles in front of the camera pair. We use v-disparity histograms for identifying relevant depth values, and extract from the 3D map successive subsets of points that correspond to these depth values. We apply a clustering step that provides the corresponding elements localisation. These clusters are then used to build a set of potential obstacles, considered as high level primitives. Experimental results on real images are provided.