CUDA implementation of belief propagation for stereo vision

Abstract

Measuring distance to obstacles is an important process for intelligent vehicles (IV). With accurate measurement, IV can make appropriate maneuver to avoid such obstacles. To obtain highly accurate result, we used a Markov random field model-based global energy minimization algorithm called belief propagation (BP). However, BP has high computational complexity which makes it difficult for real-time processing. To solve this issue, we took massively parallel approach using Compute Unified Device Architecture (CUDA). In this paper, we first provide profiling result to find the performance bottleneck of BP. Next, we explain CUDA-specific optimization techniques to enhance the performance. We propose a new parallelization technique to speed up the message computation, which takes up the longest time in BP. The experimental result shows that we were able to obtain accurate distance estimation result in real time.

Enabling Technologies for Urban Smart Mobility: Recent Trends, Opportunities and Challenges

“…Vision-based artificial intelligence in autonomous vehicles allows a set of features including the detection of obstacles as well as mechanisms to avoid them [ 147 ]. In [ 148 ], the authors propose a way to increase the processing power to estimate the distance to obstacles using Compute Unified Device Architecture (CUDA) along with Believe Propagation Algorithm (BPA). The use of AI for in-vehicle vision systems to power Advanced Driving Assistant Systems (ADAS) is explored in [ 149 ].…”