Heterogeneous LiDAR Early Fusion and Learned Re-Ranking Strategy for Robust Long-Term Place Recognition in Unstructured Environments
Researchers have developed MinkUNeXt-VINE++, a novel method for robust long-term place recognition in unstructured environments, particularly for autonomous systems in agricultural fields. This approach utilizes early fusion of data from heterogeneous LiDAR sensors, specifically the Livox Mid-360 and Velodyne VLP-16, to create a more comprehensive environmental representation. Additionally, a learned re-ranking strategy is employed during inference to improve accuracy in repetitive environments like vineyards. Evaluations on the TEMPO-VINE dataset showed a significant performance increase, with a 20% improvement in Recall@1 compared to single-sensor methods and a 30% improvement when re-ranking was included. The code for this method has been made publicly available. AI
IMPACT This research could improve the reliability and safety of autonomous systems operating in complex, unstructured environments.