UAV-based object detection model for smart surveillance using deep neural network

Unmanned aerial vehicles (UAVs) have become indispensable in both civilian and military domains, enabling applications such as smart surveillance, environmental monitoring, and search-and-rescue operations. However, effective object detection in UAV imagery remains challenging due to the small size of targets, high object density, frequent occlusions, and complex backgrounds resulting from varying altitudes and viewpoints. Existing algorithms, such as You Only Look Once (YOLO) v5, exhibit limited accuracy in detecting targets in UAV images. To address these challenges, this study proposes an enhanced YOLOv5-based detection model. The model incorporates an optimized detection module with three prediction heads for multi-scale bounding box predictions. Additionally, self-attention mechanisms and a Convolutional Block Attention Module (CBAM) are integrated to focus on salient regions and mitigate the impact of occlusions. Furthermore, we introduce a ConvELU layer, which replaces the default SiLU activation with the Exponential Linear Unit (ELU). This modified ConvELU layer is applied to the backbone, neck, and head components, effectively improving the model's feature extraction capabilities. Experimental results on the VisDrone dataset demonstrate that the proposed model achieves a precision of 95.1%, a recall of 86.3%, and a mean Average Precision (mAP) of 91.6%, outperforming standard YOLOv5 and other state-of-the-art detectors.