Vehicle Detection Based on YOLOv7 for Drone Aerial Visible and Infrared Images

Object detection using drone-captured aerial images holds great significance for real-time traffic detection and control by traffic management authorities. This paper introduces a novel algorithm based on the widely-used YOLOv7 model. The algorithm presents an enhanced evaluation metric, the Normalized Gaussian Wasserstein Distance (NWD), utilizing normalized Wasserstein distance for small object detection. NWD measures object similarity by considering associated Gaussian distributions. To address challenges from complex backgrounds and redundant features due to irrelevant noise in aerial images, the paper introduces the CBAM attention mechanism. This mechanism improves feature expression for vehicle detection, enabling selective focus on target regions. The proposed algorithm is evaluated on the Drone Vehicle dataset and compared with state-of-the-art algorithms. Experimental results demonstrate favorable average precision values in both visible light and infrared images. The improved network model leads to a 1.2% increase in mean average precision (mAP) for visible light image detection and a 0.9% increase for infrared image detection.