Development of a graphics processing unit accelerated stereo vision system for depth estimation

摘要

Stereo vision is an important computer vision technique which estimates depth information from two images - similar to how human vision works. However, the computational intensity of stereo vision deters implementation in real-time systems. Recently, massively parallel processors called Graphics Processing Units or GPUs are gaining traction as an acceleration platform to offload compute-intensive tasks to. This paper presents the development of a GPU-accelerated stereo vision system called DepthStream. The DepthStream algorithm uses a combined absolute difference and census cost, cross-based cost aggregation and bitwise fast region voting to achieve error rates up to 3.7× lower than Block-Matching based algorithms in mainstream libraries. Leveraging the parallel processing power that GPUs offer enables the DepthStream algorithm to achieve processing times up to 3.12× faster than Semi-Global Matching based implementations in the same libraries. The study also includes measurements of the accuracy of the estimated distances of the developed stereo vision system. Results show that the estimated depths of the DepthStream stereo vision system are within 4% of the true value for distances less than 6m. Statistical analysis also indicate that there is no significant difference between the stereo vision systems estimated distances and the actual value.