Global Alignment of Deep Features for Robot Localization in Changing Environment

摘要

Localization is an elemental requirement for autonomous navigation, simultaneous localization and mapping for mobile robots. As robots can perform long-term and large-scale tasks, finding locations in changing environment is a crucial problem. To resolve the problem, we present a robust localization system under severe appearance changes. The system consists of two stages. First, a robust feature extraction method using a deep convolutional auto-encoder is proposed. Then, global alignment of extracted feature sequences is proposed to find the actual robot's locations. Since the proposed method not only uses the condition-robust features but also considers the actual trajectory of the robot by aligning features sequences, it can show accurate localization performances in changing environments. Experiments were conducted to prove the effective of the proposed method, and the results showed that our method outperformed than existing methods.