Visual marking-based AGV localization in parking lot

Abstract

The combination of RTK and UWB is widely used for AGV positioning. In the case of large obstacle occlusion, GNSS-based localization methods may fail to obtain the accurate position of AGV. In this paper, we establish a new landmark dataset and put forward an auxiliary localization method based on landmark detection to improve the position accuracy, where a multi-task learning model is applied for both keypoint prediction and bounding box regression. We further propose the Keypoint Recovery Module (KRM) as a model-agnostic plug-in, to mitigate the challenge of missing rate. By this, the proposed approach is trained and validated on our proposed landmark dataset. Comparative experimental results show that the multi-task architecture in conjunction with KRM greatly enhances the accuracy of landmark detection, surpassing traditional methods.