Onboard Vehicle Detection for Autonomous UAV Reconnaissance under Degraded Visual Conditions

Abstract

Autonomous unmanned aerial vehicle (UAV) reconnaissance imposes three simultaneous requirements on onboard target detectors: high accuracy for small targets that frequently appear in medium-altitude imagery, robustness under realistic image degradation, and sufficient compactness for deployment on embedded hardware with strict power and latency constraints. This paper presents UAVR-YOLOv8, an enhanced YOLOv8-based detector designed to address these three requirements within a unified framework. The proposed model incorporates three main components: a P2 detection branch to better exploit shallow features for small-target detection, a Coordinate Attention module to enhance spatial feature representation, and a curriculum augmentation strategy covering motion blur, low-light, fog, sensor noise, and atmospheric haze. The model was trained and evaluated on a selected aerial image dataset consisting of UAV images containing three major target groups: main battle tanks, armored personnel carriers, and military trucks, captured under various terrain and daylight conditions. In addition to conventional evaluation metrics, this study introduces a robustness index to quantify model reliability under degraded imaging conditions, thereby extending the assessment beyond average accuracy on clean images alone. Experimental results on the NVIDIA Jetson Orin Nano show that the proposed model maintains strong detection accuracy on clean images, significantly improves performance retention under the five considered degradation types, and achieves real-time inference within the allowed power budget. The full ablation study further confirms that all three proposed modifications contribute positively, with small but consistent synergistic gains when integrated into a single design. These results indicate that an appropriate combination of proven techniques, together with evaluation under realistic operating conditions, can provide a practical detector for autonomous UAV reconnaissance deployment.