Drone Aerial Object Detection Based on Improved YOLOv8

Drone object detection has shown significant applicability in multiple fields. However, despite its promise, drone object detection continues to encounter several challenges. These challenges arise primarily due to the limited accuracy of lightweight detection models designed for deployment on devices constrained by hardware resources, compounded by the inherent complexities in aerial imagery captured by drones, such as complex backgrounds and the presence of numerous small objects. To address these limitations, this study introduces an enhanced drone-based aerial object detection model, building upon the YOLOv8n architecture. First, the BiFPN_Concat is utilized to refine the original Concat module within the Neck component of the network. Additionally, a new detection branch is incorporated into the Head section, specifically designed to improve the detection of small objects, augmenting the existing three detection branches. Furthermore, the Focaler-CIoU loss function is integrated to enhance the model's localization accuracy. The proposed lightweight and small-scale models were evaluated on the VisDrone2019 dataset, with the results demonstrating a 4.3% improvement in mAP50 and a 3% improvement in mAP50-95 compared to the baseline YOLOv8n model. These findings indicate that the proposed model significantly enhances both the precision and robustness of drone-based object detection systems.