PQMPV: Parallel and Quantified Dense Stereo Disparity Estimation Based on Multi-Path Viterbi

Abstract

Stereo depth estimation plays a crucial role in enabling robots, autonomous driving and other applications. As an ill-posed problem, both deep learning based and traditional methods have been extensively studied in recent years. Traditional methods mainly focus on feature matching between multiple images to estimate depth, but often result in blank pixels with zero disparity and also require significant computation time. In this paper, we propose a new stereo disparity estimation algorithm based on the dynamic programming method known as Viterbi. Our approach optimize the range of disparity calculation through image pyramid, combined with Viterbi algorithm for cost aggregation, and accelerates several calculation steps through parallel computing techniques. Additionally, we improve data transmission and processing speed by quantifying data from float to integer. Experiments on the KITTI dataset and MiddEval benchmark demonstrate that our algorithm achieves significant improvements in average error rate and processing speed of disparity map, making it a promising solution for stereo depth estimation in practical applications.