Bi-PPYOLO tiny:一种轻量型的机场无人机检测方法

Bi-PPYOLO tiny: a lightweight airport UAV detection method

机场“黑飞”无人机的检测关系着整个机场的安全问题,机场现有基于雷达手段的无人机探测方法无法正确识别无人机的类型及个数。基于现有雷达探测无人机方法存在的缺点,对PPYOLO tiny目标检测方法在无人机检测中存在的问题进行改进,结合机场“黑飞”无人机的特性和硬件设备部署中模型参数量小的特性,提出了基于Bi-PPYOLO tiny的轻量型无人机检测方法,提出双锥台特征融合结构,并优化检测头部的锚框大小,有效提升了无人机的检测精度。经试验验证,该方法将平均检测精度PmA从68.07%提升至76.71%,模型参数量为4.06 MB,推理速度为32.21帧/s。所提方法有助于轻量型无人机检测方法在光电设备上的部署与实施,与现有机场无人机探测手段共同保障机场安全。

Based on the shortcomings of existing radar-based Unmanned Aerial Vehicles(UAVs) detection methods, this paper proposed a Bi-PPYOLO tiny-based UAVs detection method that combines the characteristics of “black-flying” UAVs at airports and the requirement of small model parameters in the deployment of hardware equipment. In this paper, we proposed a Bifrustum feature fusion structure, which effectively fuses the small object UAV signals present in the multi-channel information to provide a feature map containing more UAVs information for the subsequent head detection UAV. In addition, the different layers of the feature map are passed in different channels to the subsequent head. Shallow feature maps contain more detailed information and are suitable for small UAV detection;deeper feature maps contain more global information and are suitable for large UAV detection;Additionally, this paper optimized the setting of the prediction anchor frame in the detection head. Due to the small size of the UAV target, using a smaller prediction frame improves the accuracy of target detection, so that it can match more accurately with the real position of the object when predicting small targets and optimizing the size of the anchor frame in the detection head, effectively improving the detection accuracy of the UAVs. In the experiment, we used 1 488 UAV pictures for training and testing, and the experimental results verified that the method in this paper improved the mean average precision PmAfrom 68.07% to 76.71%, with a model size of 4.06 MB and an inference speed of 32.21 frames/s. The experimental results show that the size of the model meets the requirements of embedded hardware devices, and the model can be embedded in the detection equipment of airports. The method in this paper will contribute to the deployment and implementation of lightweight UAVs detection methods on the photodetector, which work together with existing UAV detection methods to ensure airport safety.