Dynamic Weighting-Based Deep Learning for SLAM Loop Closure Detection

In LiDAR SLAM systems, loop closure detection is a crucial component to ensure localization accuracy and map consistency. OverlapNet, a loop closure detection algorithm based on image overlap rate, has demonstrated strong performance. However, the current loss function of OverlapNet does not account for the uncertainty differences between tasks, which may lead to insufficient adjustment of task weights. To address this, this paper proposes a dynamic weight-based SLAM loop closure detection method, DW-LoopNet. By introducing a dynamic weight mechanism, this method automatically adjusts the relative weights of overlap and yaw angle estimation tasks. Based on real-time evaluation of environmental feature importance, DW-LoopNet dynamically adjusts the weights of different features, ensuring that key features receive higher attention in loop closure detection. Experimental results show that DW-LoopNet significantly outperforms other baseline methods on the KITTI and Ford Campus datasets, providing a more accurate and robust loop closure detection solution for SLAM systems.