CUDA-based SeqSLAM for real-time place recognition

Abstract

Vehicle localization is a fundamental issue in autonomous navigation that has been extensively studied by the Robotics community. An important paradigm for vehicle localization is based on visual place recognition which relies on learning a database, then consecutively trying to find matchings between this database and the actual visual input. An increasing interest has been directed to visual place recognition in varying conditions like day and night cycles and seasonal changes. A major approach dealing with such challenges is Sequence SLAM (SeqSLAM) based on matching a sequence of images to the database instead of a single image. This algorithm allows global pose recovery at the expense of a higher computational time. To solve this problem with a certain amount of speedup, we propose in this work, a CUDA-based solution for real-time place recognition with SeqSLAM. We design a mapping of SeqSLAM to CUDA architecture and we describe, in detail, our hardware-specific implementation considerations as well as the parallelization methods. Performance analysis against existing CPU implementation is also given, showing a speedup to six times faster than the CPU for common sized databases. More speedup could be obtained when dealing with bigger databases.