SkyScan: Optimized YOLOv5n Architecture with SAHI for Aerial Surveillance

SkyScan is a UAV-based traffic monitoring and surveillance system designed to overcome the limitations of traditional methods, such as blind spots, limited scalability, and environmental constraints. This work enhances real-time object detection by proposing a modified YOLOv5n architecture, tailored for the VisDrone dataset. Key improvements include the addition of a fourth detection scale (P6/64), integration of a Convolutional Block Attention Module (CBAM), and deeper C3 layers to boost feature extraction and fusion. To further refine detection accuracy, especially for small and overlapping objects, SkyScan incorporates the Slicing Aided Hyper Inference (SAHI) technique in the post-processing stage. Optimization techniques such as quantization and resource efficient inference are also employed. Comparative evaluation with YOLO variants (YOLOv5n, YOLOv7n, YOLOv12n) demonstrates that the enhanced model achieves a notable increase in mAP@0.5, while F1 Score reaching 0.35 compared to 0.30 from the baseline YOLOv5n. SkyScan ensures real-time object detection, tracking, and density estimation, offering a scalable and efficient solution for smart city surveillance with improved detection reliability.