Abstract:

To accurately detect obstacles during the operation of unmanned mining trucks in open-pit mines, an improved obstacle detection model based on YOLOv8 was proposed. Firstly, STA mechanism was introduced to address the issues of redundant computation and insufficient global modeling in traditional attention mechanisms, thereby enhancing the feature representation capability for small and occluded targets. Secondly, a neck network structure was designed to further improve the localization accuracy of small targets by fusing high-level and low-level features. Then, a compound loss function was constructed to balance the regression accuracy and robustness of bounding boxes, so as to adapt to the diverse shapes of obstacles in open-pit mines. Finally, a self-built dataset was utilized to conduct comparative experiments for verifying the effectiveness of the improved model. The results show that the accuracy, recall rate, and mean Average Precision (mAP) at an Intersection over Union(IoU) threshold of 0.5 of the improved model are 86.2%, 78.9%, and 83.1% respectively, which are 10.3%, 19.4%, and 18.0% higher than those of YOLOv8. Compared with traditional detection algorithms, the improved model exhibits advantages in detection accuracy, recall rate and computational complexity.

Key words: open-pit mine, unmanned mining truck, obstacle detection, super token attention mechanism (STA), YOLOv8