End-To-End Visual Place Recognition Based on Deep Metric Learning and Self-Adaptively Enhanced Similarity Metric

摘要

Place recognition, which aims at recognizing the previously visited places, is the key component of loop closure in most visual simultaneous localization and mapping systems. Despite significant progress, challenges still remain especially in the longtime mapping and localization as the appearance of an environment may change greatly over time. In this paper, deep metric learning which jointly optimizes feature extraction and similarity metric is utilized to train an end-to-end network specifically for place recognition task to handle the appearance changing over time. A self-adaptively enhanced similarity metric is designed to strength the discrimination ability and calculate the similarity between descriptors of image pairs which are extracted from a convolutional neural network. Experiments on two typical open datasets illustrate the superior performance of our approach and the outstanding robustness when appearance changes.