Small object detection in complex scenes, particularly for Unmanned Aerial Vehicle (UAV) aerial imagery, remains a core challenge in computer vision, primarily due to scarce feature information, vulnerability to background interference, high sensitivity of the traditional intersection over union (IoU) metric to minor positional deviations, and severe background clutter. To address these issues, this paper proposes a small-object detection model based on multi-module collaborative optimization. Built upon the YOLOv8 framework, the model introduces systematic improvements in three aspects: feature fusion, feature enhancement, and loss function design. Specifically, a weighted bidirectional feature pyramid network (BiFPN) is incorporated to provide richer multi-scale contextual information for small objects; a novel global-local collaborative attention mechanism (GLSA) is embedded to enhance the discriminative power of key features; and a bounding box regression loss based on Wasserstein distance is adopted to deliver smoother gradient behavior during training. The proposed model is comprehensively evaluated on a public dataset comprising numerous aerial and ground small objects. Compared with the baseline YOLOv8, our model achieves substantial performance gains: the optimal F1-confidence threshold rises from 0.42 to 0.67, while the mean average precision at IoU = 0.5 (mAP@0.5) increases from 81.4% to 81.8%, yielding a 0.4 percentage point improvement. These results indicate that the predicted confidences are better calibrated with respect to true accuracy, thereby enhancing the reliability of output detections. Meanwhile, the F1-score, which balances precision and recall, improves from 77% to 78%. Ablation studies further confirm that BiFPN and GLSA jointly improve classification and recall for small objects, whereas the Wasserstein Loss specifically optimizes localization accuracy. Together, these three modules constitute a cohesive high-performance detection pipeline. This work provides an effective and reliable solution for real-time small object detection in complex environments.
Long-horizon ground robots answer thousands of navigation queries on persistent 3D maps, making repeated planning a computational burden. We present a topology-grounded framework combining seed-reachable surface extraction, persistent-homology room segmentation, and a compact Transition Graph (TG). Surface extraction removes more than 80% of states in furniture-rich scenes. Room segmentation selects its persistence threshold automatically from map statistics, without using query outcomes. Paired boundary portals, cached Dijkstra-tree paths, and a consistent directed landmark bound reduce online computation. On 100 Habitat-Matterport 3D (HM3D) scenes and 500 queries at 0.1 m resolution, TG with one portal per boundary achieves a 0.054 ms median query time, 2.52× faster than corrected surface ALT (A*, landmarks, and triangle inequality). Five portals reduce median geometric path overhead to 3.27%, with 2.24× speedup. All three configurations return 451/500 positive-length paths, all passing the surface-adjacency check. Directed search costs match independent Dijkstra references in a separate targeted validation. Earlier cross-dataset measurements provide additional context: 452× faster search than 3D2M with 100% versus 66% success, and an approximate Geekbench-normalized cross-platform pipeline speedup exceeding 10× over point cloud tomography (PCT). These results support topology-grounded planning for repeated queries on persistent maps.
