A visual based method for small target unmanned aerial vehicle recognition

The rapid development and popularity of drones has posed a serious threat to the safety of the airspace in special areas. In order to ensure airspace security, drone detection is particularly necessary. Aiming at the problem of small target drones being difficult to detect, this paper proposes a detection method optimized based on the Yolo V4 algorithm. In the proposed method, the edge detection result of the original image and the original 3-channel image are combined into a 4-channel image firstly. Then, the shallow feature information and deep semantic information are fused to optimize the network structure for improving the utilization of shallow information. At last, by increasing the resolution of the feature map output from the detector head, it can effectively improve the detection ability of the detector head for small targets In addition, to improve the detection efficiency of the network model, depth-separable convolution is used to replace the standard convolution in order to reduce the number of operations in the network. In order to verify the effect of the proposed method, experiments running in Jetson TX2 based on established two data sets including 3000 drone flying images show that the accuracy of the method is similar to that of Yolo V4 tiny, but the detection efficiency of the proposed method is 13.71% higher than the latter.