Advanced SeqSLAM using discriminative information for light-rail localization at high frame rate

Abstract

In recent years, vision-based localization technology in driving assistance system has drawn much attention. In this paper, an Advanced SeqSLAM method is proposed to solve the problem of localization due to the high similarity of scenes in high-accuracy scene matching of light rail system. In this method, salient regions with discriminative information are extracted from high-similarity frames of reference sequences by off-line processing, and binary feature descriptors are generated in these regions to improve the speed and precision of scene matching. Compared with the local features, the error of the proposed scene matching method is reduced by 31.43% and the computation time is reduced by 94.22% in the Hong Kong MTR dataset. Compared with the scene tracking algorithm of SeqSLAM, the precision of scene tracking based on proposed binary features in salient regions is increased by 9.84% compared without significant increase of running time in the Nordland dataset. The experimental results show that the proposed method improves the performance of the light rail localization.