An Object Detection Algorithm Based on YOLOv8-LSK for Plateau Tibetan Brown Bears

To overcome the performance limitations of existing object detection algorithms in plateau wildlife monitoring scenarios, particularly the accuracy degradation caused by partial occlusions in Tibetan brown bear images. This paper proposes an improved YOLOv8-based detection framework incorporating species-specific visual characteristics. The proposed YOLOv8-LSK architecture integrates a Large Selective Kernel (LSK) attention module into the backbone network, enhancing multi-scale feature extraction through parallel multi-branch depth-wise separable convolutions with varying kernel sizes. For occluded targets requiring extended contextual reasoning, the LSK module implements a dynamic spatial selection mechanism that automatically prioritizes the most discriminative receptive fields. This adaptive attention mechanism enables context-aware feature refinement by selectively aggregating multi-scale spatial information, effectively compensating for missing visual cues in occluded regions. Comprehensive evaluations on a dedicated plateau wildlife dataset demonstrate the framework's effectiveness, with YOLOv8-LSK achieving 87.5% mAP50 and 67.6% mAP50-95, outperforming comparative models including YOLOv5 (83.2%/63.1%), YOLOv6 (84.7%/64.8%), baseline YOLOv8 (85.9%/65.3%), RT-DETR-ResNet18 (82.4%/61.9%), and RT-DETR-ResNet50 (84.1%/63.7%). Ablation studies further validate that the LSK module contributes 2.3% and 1.8% absolute improvements in mAP50 and mAP50-95 respectively compared to the original YOLOv8 architecture.