Optimized YOLOv7 for Small Target Detection in Aerial Images Captured by Drone

It is challenging to detect small targets in aerial images captured by drones due to variations in target sizes and occlusions arising from the surrounding environment. This study proposes an optimized object detection algorithm based on YOLOv7 to address the above-mentioned challenges. The proposed method comprises the design of a Genetic Kmeans (1IoU) clustering algorithm to obtain customized anchor boxes that more significantly apply to the dataset. Moreover, the SPPFCSPC_group structure is optimized using group convolutions to reduce model parameters. The fusion of Spatial Pyramid Pooling-Fast (SPPF) and Cross Stage Partial (CSP) structures leads to increased detection accuracy and enhanced multi-scale feature fusion network. Furthermore, a Detect Head is incorporated into the classification phase for more accurate position and class predictions. According to experimental findings, the optimized YOLOv7 algorithm performs quite well on the VisDrone2019 dataset in terms of detection accuracy. Compared with the original YOLOv7 algorithm, the optimized version shows a 0.18% increase in the Average Precision (AP), a reduction of 5.7 M model parameters, and a 1.12 Frames Per Second (FPS) improvement in the frame rate. With the above described enhancements in AP and parameter reduction, the precision of small target detection and the real-time detection speed are increased notably. In general, the optimized YOLOv7 algorithm offers superior accuracy and real-time capability, thus making it well-suited for small target detection tasks in real-time drone aerial photography.