Agenda

Signal Processing Seminar

GPU-Accelerated Adaptive Unstructured Road Detection Using Close Range Stereo Vision

Bugra Ozutemiz, PhD candidate, Middle-East Technical University

Detection of road regions is not a trivial problem especially in unstructured and/or off-road domains since traversable regions of these environments do not have common properties. Even the properties and appearance of these environments can change on the run. Hence, an algorithm working under unstructured conditions should have a continuous adaptation capability. To achieve this, a novel unstructured road detection algorithm that can continuously learn the road region is proposed in this work. The algorithm gathers close-range stereovision data using a simple roughness threshold and uses this information to estimate the road region in the far field. The proposed approach simplifies over the approaches in the literature by changing offline supervised learning and pose estimation of the vehicle and sensor with a simple heuristic coming from the nature of the problem: roughness (or smoothness) of the terrain. Thanks to the parallel nature of the algorithm, it is also implemented on a GPU with CUDA and a real-time running performance is achieved even on a low-performance graphics card. The experiments show that the algorithm gives excellent road detection results even under fast-changing light conditions and a running frequency of 50 Hz is achieved even in the worst case.

Overview of Signal Processing Seminar