MSOD - A Small Obstacle Detection Network for Open-Pit Mining Areas

With the development of artificial intelligence technology, autonomous driving has advanced significantly, and open-pit mining areas have emerged as one of the important scenarios for the implementation of autonomous driving in recent years. The perception module forms the foundation of the autonomous driving system for mining trucks, among which the detection of small obstacles such as stones and cones is crucial for ensuring driving safety. Currently, small obstacle detection faces problems such as small size, insufficient features, and complex backgrounds. Meanwhile, a single sensor is vulnerable to sensor failure and difficult to deal with occlusion and multi-source noise interference. To address these challenges, this article designs a novel algorithm framework based on feature fusion of LiDAR and camera, combining strategies such as multi-modal data augmentation, feature layer fusion, and temporal modeling to achieve high-precision detection of small obstacles. The problem of feature scarcity is alleviated through multi-scale feature extraction and cross-modal data augmentation, a feature fusion scheme is constructed based on the cross-attention mechanism, and a temporal self-attention mechanism is introduced to realize dynamic association of small obstacles in consecutive frames. Experimental results demonstrate that the proposed algorithm achieves a mean Average Precision (mAP) of 83.4% and a NuScenes Detection Score (NDS) of 72.5% on the self-built dataset. Compared with advanced methods, MSOD significantly improves detection accuracy for small obstacles.