Optimizing Census-based Semi Global Matching by genetic algorithms

Abstract

Recent years have shown a great progress in self-driving vehicles and stereovision has proven to be a key aspect towards this goal. Semi-Global Matching (SGM) algorithm is among the best stereo solutions, capable of producing reliable results at reasonable cost. Census transform is generally preferred as a cost metric due to its robustness and invariance to lighting conditions. This paper proposes an original methodology for finding both the optimal Census mask and the best values for the penalties P1 and P2 in SGM by using genetic algorithms (GA). The obtained census masks are thoroughly analyzed and the best ones can be combined in a weighted center-symmetric census to increase the performance of SGM. Kitti test cases show that our GA-based censuses as well as our novel weighted center-symmetric census outperform dense, sparse and center-symmetric counterparts for Census only and SGM.