GOReloc: Graph-Based Object-Level Relocalization for Visual SLAM

Abstract

This article introduces a novel method for object-level relocalization of\nrobotic systems. It determines the pose of a camera sensor by robustly\nassociating the object detections in the current frame with 3D objects in a\nlightweight object-level map. Object graphs, considering semantic\nuncertainties, are constructed for both the incoming camera frame and the\npre-built map. Objects are represented as graph nodes, and each node employs\nunique semantic descriptors based on our devised graph kernels. We extract a\nsubgraph from the target map graph by identifying potential object associations\nfor each object detection, then refine these associations and pose estimations\nusing a RANSAC-inspired strategy. Experiments on various datasets demonstrate\nthat our method achieves more accurate data association and significantly\nincreases relocalization success rates compared to baseline methods. The\nimplementation of our method is released at\n\\url{https://github.com/yutongwangBIT/GOReloc}.\n