A Multi-Scale UAV Cluster Detection and Counting Framework Based on Multi-Source Information Fusion and Enhanced YOLOv9

Abstract

To address the safety hazards posed by unauthorized "black flight" of unmanned aerial vehicle (UAV) clusters, as well as the limitations of existing detection methods in tiny target recognition and complex environment adaptability, this paper proposes a real-time detection and counting method based on a feature fusion framework. The core contribution of this work is a novel multi-level fusion architecture that aggregates multi-scale visual features to enhance detection robustness under challenging conditions. First, we construct a multi scene UAV data set containing 7,113 images, and design a dedicated feature fusion module based on improved YOLOv9 backbone. The Haar Wavelet Down-sampling (HWD) module is introduced to fuse high-frequency edge details with low-frequency approximation features, preserving critical texture information that is essential for tiny target detection. The Programmable Gradient Information (PGI) mechanism is adopted as an auxiliary reversible branch, which propagates reliable deep-layer gradient information to shallow layers to effectively mitigate the information bottleneck problem. Second, a decision-level fusion strategy is developed to aggregate confidence scores and localization outputs from multi-scale predictions. The strategy is further optimized with a tailored Non-Maximum Suppression (NMS) algorithm to resolve occlusion issues in dense UAV clusters. Evaluations on the self-built dense UAV dataset show that the proposed method achieves a detection accuracy of 85% with a per-frame processing time of 0.10–0.15 s. Tests across diverse scenarios validate that the method can accurately locate and identify UAV targets. By achieving a favorable balance between detection accuracy and real-time performance, the proposed method effectively alleviates the missed detection and false detection problems in dense UAV cluster detection tasks.