YOLOv11-dense: a small UAV detection algorithm for low-altitude security

Abstract

To address the prevalent issue of illegal flights of small unmanned aerial vehicles (UAVs) in low-altitude security scenarios, as well as the critical limitations of general-purpose object detection models—namely insufficient feature extraction capability for small targets and high false positive rates—this paper proposes an improved small UAV detection algorithm based on YOLOv11n, namely A DenseBlock-Enhanced YOLOv11 (YOLOv11-Dense). First, in the backbone feature extraction stage, the original C3K2 blocks are replaced with dense blocks, and the Adaptive Attention Module (AAM) is embedded to substitute the native attention mechanism. The inherent feature reuse property of dense blocks enhances the multi-scale feature propagation capability. Meanwhile, AAM adaptively focuses on the discriminative features of small targets and suppresses interference from complex backgrounds, thereby reducing the detection false positive rate. Second, in the neck feature fusion stage, the Multi-scale Feature Fusion (MSFF) module is introduced to construct a cross-layer bidirectional feature extraction and aggregation path. This module strengthens the fusion of deep semantic features and shallow spatial features, mitigates the problem of small target feature loss in deep networks, and further improves the detection accuracy of small targets. Experimental results on a UAV dataset formatted in the YOLO standard demonstrate that YOLOv11-Dense achieves a mean average precision (mAP@0.5) of 82.5%. Under the hardware environment of NVIDIA GeForce RTX 4050, the inference speed for a single image reaches 6.3 ms, which fully satisfies the real-time detection requirements of edge devices in low-altitude security scenarios.