A multi-level dynamic programming method for stereo line matching

Abstract

This paper proposes the use of a multi-level dynamic programming method to solve the line matching problem of lateral stereo vision. A Local Similarity Measure between the left and right images is calculated for each line segment pair. At level 1, line segment pairs that have a very high Local Similarity Measure are selected for the matching process, so that the probability of a correct match at this level is very high. A dynamic programming method is used to search for the best match and the matching probability is represented by the Local Similarity Measure. Matched pairs are used to assist the matching process at the next level. By considering the geometric properties between the matched and the unmatched line segments, a Global (Structure) Similarity Measure is calculated for each unmatched line segments pair. An overall Similarity Measure (matching probability) is obtained by using the Local Similarity Measure and the Global Similarity Measure. The algorithm begins the second match. Line segment pairs that have a Local Similarity Measure and a Global Similarity Measure larger than a threshold T are selected for the matching process. T is set to a relatively high value at level 2 and gradually decreased as the level advances. The new matched results are used to modify the Global Similarity Measure and the overall Similarity Measure. These processes are repeated until a predefined level n stop (or a predefined condition) is reached. By using the Global Similarity Measure and a multi-level searching technique, the proposed technique achieves a high success rate (matching accuracy) and a high discover rate when dynamic programming is used for stereo matching.