Parallel Computation for Stereovision Obstacle Detection of Autonomous Vehicles Using GPU
Due to the parallelism on general-purpose computation, the graphics processing unit (GPU) is applied in autonomous vehicles and a stereovision obstacle detection system is developed. We first perform census transform on the rectified stereo image pair; then employ the epipolar constraint to facilitate the visual correspondence matching. Based on the dense disparity map, the 3D coordinate of each pixel is calculated, according to which obstacles are identified. Furthermore, several techniques are listed for exploring the specific functionalities of GPU to boost the overall performance. A prototype system is finally implemented and integrated in the onboard PC of autonomous vehicles. Experimental results validate the real-time accuracy under various illuminations and road conditions. Since the low level image-processings are run by GPU in parallel, the proposed design not only merits high speed and efficiency, but also frees up CPU so as to focus on the decision and control.
graphics processing unit parallelism disparity obstacle detection autonomous vehicle
Zhi-yu Xu Jie Zhang
School of Electronics and Information Engineering, Tongji University,Shanghai, 201804, China
国际会议
无锡
英文
176-184
2010-09-17（万方平台首次上网日期，不代表论文的发表时间）