LiDAR Loop Closure Detection using Semantic Graphs with Graph Attention Networks

• DOI: 10.1007/s10846-025-02223-6
• Publication Date: 2025-01-11T08:29:03Z
• Authors: Liudi Yang; Ruben Mascaro; Ignacio Alzugaray; Sai Manoj Prakhya; Marco Karrer; Ziyuan Liu; Margarita Chli
• Keywords: Closure (psychology)

In this paper, we propose a novel loop closure detection algorithm that uses graph attention neural networks to encode semantic graphs perform place recognition and then use registration estimate the 6 DoF relative pose constraint. Our has two key modules, namely, encoder module comparison module. The employs efficiently spatial, geometric information from of input point cloud. We self-attention mechanism in both node-embedding graph-embedding steps create distinctive vectors. vectors current scan keyframe are compared identify possible closure. Specifically, employing difference showed significant improvement performance, as shown ablation studies. Lastly, implemented takes candidate scans estimates constraint for LiDAR SLAM system. Extensive evaluation on public datasets shows our model is more accurate robust, achieving 13% maximum F1 score SemanticKITTI dataset, when baseline algorithm. For benefit community, open-source complete implementation proposed custom at https://github.com/crepuscularlight/SemanticLoopClosure