RESEARCH ON YOLOv5-BASED VISUAL SLAM OPTIMISATION METHOD IN FARM DEPOT ENVIRONMENT

Abstract

Conventional simultaneous localization and mapping (SLAM) systems for agricultural robots rely heavily on static rigidity assumptions, which makes it susceptible to the influence of dynamic target feature points in the environment thus leading to poor localization accuracy and robustness of the system. To address the above issues, this paper proposes a method that utilizes a target detection algorithm to identify and eliminate dynamic target feature points in a farm depot. The method initially employs the YOLOv5 target detection algorithm to recognize dynamic targets in the captured warehouse environment images. The detected targets are then integrated into the feature extraction process at the front end of the visual SLAM. Next, dynamic feature points belonging to the dynamic target part are eliminated from the extracted image feature points using the LK optical flow method. Finally, the remaining feature points are used for location matching, map construction and localization. The final test on the TUM dataset shows that the enhanced vision SLAM system improves the localization accuracy by 91.47% compared to ORB-SLAM2 in highly dynamic scenes. This improvement increases the accuracy and robustness of the system and outperforms some of the best SLAM algorithms while maintaining high real-time performance. These features make it more valuable for mobile devices.