Unsupervised obstacle detection for weeding robots: a video transformer-based approach

摘要

Automation of tasks such as mechanical weeding has advanced significantly, especially with autonomous rovers that increasingly rely on AI and RTK GNSS for navigation. However, large multiton machines still face safety challenges, especially in detecting obstacles. Deep learning applied to image analysis is a suitable solution to this challenge. Therefore, VMTAD, an unsupervised videobased approach using a video memory transformer for anomaly detection incorporating temporal information has been developed. VMTAD outperforms previous methods and provides improved obstacle detection that enhances safety and overcomes critical limitations in current robotic systems. The code is available at: https://github.com/TheoBiardeau/VMTAD.