FM-YOLO: A Novel Approach to Small Object Detection for Dense Vehicle Scenarios in UAV Images

The increasing adoption of UAVs in traffic management and surveillance brings significant challenges, particularly in the detection of small and densely distributed objects in complex urban environments. To address these, we propose FM-YOLO, an enhanced object detection framework based on YOLOv8-n, designed specifically to tackle the problem of small and dense object detection. FM-YOLO incorporates four key improvements, each aimed at solving the core challenge of detecting small objects in crowded and complex settings. First, the backbone adopts the Mixed Aggregation Network (MANet) for advanced feature extraction, which enhances the model’s ability to capture both fine-grained details and high-level features, thus improving its robustness for small object detection. Second, we introduce Shallow Fusion (SF), a novel method designed to preserve essential low-level details that are crucial for detecting small objects, particularly in cluttered environments. Third, we propose Deep Fusion (DF), a technique that enhances high-level semantic features to improve detection performance for larger objects while maintaining the accuracy of small object detection. Finally, the Spatial Pyramid Pooling with Cross-Stage Partial Connections (SPPFCSPC) module is integrated to achieve efficient multi-scale feature fusion, which further improves the model’s ability to detect objects at varying scales.