EWH-YOLO: Efficient Small Unmanned Aerial Vehicle Detection with Weighted Bidirectional Feature Fusion and Hybrid Bounding Box Regression Loss

Abstract

Vision-based unmanned aerial vehicle (UAV) detection has become increasingly important since it is a key technology in aerial collision avoidance systems. However, the detection of small UAVs is still unsatisfactory in practical applications. To address this problem, this paper proposes EWH-YOLO, a novel deep convolutional neural network-based method for small UAV detection. First, an efficient feature extraction network is designed to exact the multi-level features of small UAVs while reducing the network parameters and computational complexity. Second, a weighted bidirectional feature fusion network is proposed to enhance the low-level and high-level features in the output feature maps. Third, a hybrid bounding box regression loss is introduced to evaluate the difference between the predicted bounding box and the ground-truth bounding box during training and improve the detection accuracy. Finally, a new dataset is created on the basis of considering small UAVs to verify the detection performance. Compared with the state-of-the-art methods, the proposed method achieves higher detection accuracy with lower model complexity. The experimental results demonstrate that the proposed detector significantly improves the detection performance of small UAVs.