An Improved Loop Closure Detection for RatSLAM

RatSLAM is a brain-inspired simultaneous localization and mapping(SLAM) system used to construct experience maps for environments, requiring odometry data and visual scenes in the path traveled. But it does not perform well in complex indoor surroundings due to the rough visual processing approach. In this paper, we propose a modified RatSLAM, a feature-based monocular SLAM system that operates in real-time. The learned arrangements of three patch codes(LATCH) and multi-index hashing(MIH) algorithm are introduced to evaluate the similarity between the current scene and previously seen scenes. Then Bayesian filtering and sliding window are employed to filter out the most similar visual template and complete the loop closure detection(LCD). Experimental results show that the improved system achieves higher accuracy on place recognition and better experience map building performance in indoor environments.