LiDAR Imaging-based Attentive Perception

In this paper we present a novel approach on attentive robotic perception by developing a saliency model on LiDAR imaging. Modern LIDAR sensors provide access to a multitude of structured images, namely intensity, reflectivity, ambient, and range. These images can in turn be fused and a saliency model can be developed in a manner analogous to the human attentive system but tailored to the uniqueness of LiDAR perception. The derived LiDAR-based saliency model exploits a bottom-up approach according to which the reflectivity, intensity, range, and ambient images are compared to each other and through a sequence of image processing steps multiple conspicuity maps are acquired which in turn give rise to a unified saliency map that efficiently encodes which objects are most important and worth further analysis. This attentive perception system on LiDAR imaging gives rise to efficient obstacle detection for robotic systems that provides 360deg coverage and allows millisecond-order execution times on embedded processors. The derived model is experimentally evaluated on a set of datasets from ground and flying robots.