Drone Detection Based on Multi-scale Feature Fusion

With the widespread application of drones in commercial and industrial fields, drone detection has received increasing attention in public safety and others. However, due to significant differences in target scales, complex flight backgrounds, and the existence of interfering targets, drone detection based on vision technology remains a challenging task nowadays. Convolutional neural networks(CNN) can learn target features with strong expressive ability. However, its robustness for detecting large, medium, and small targets is not stable due to the lack of low-level feature semantic information and the insufficiency of high-level feature details. For this challenge, we propose a drone detection network that extracts features of the target from multiple receptive fields via Res2net and realizes hierarchical multi-scale feature fusion with a novel mixed feature pyramid structure. We also build a drone detection dataset to evaluate our approach. Our method outperforms RetinaNet on our dataset, and our model's mean average precision (mAP) is more than 93%.