基于改进卷积神经网络的非合作无人机检测应用

Non-cooperative Unmanned Aerial Vehicle Detection and Application Based on Convolutional Neural Network

针对“黑飞”无人机侵犯公民隐私、危害个人及公共安全,现有的无人机检测算法难以平衡检测速度和精度且对小目标的检测精度较低等不足,在YOLOv3的基础上进行改进,提出多尺度目标检测网络(multi-scale object detection network,MS-Net)对低空中的无人机进行快速高效地检测,为实现后续的防护压制提供依据。针对锚点框,通过K-均值(K-means)聚类方法得出更准确预测目标区域的位置。在特征提取部分,使用空间金字塔池化(spatial pyramid pooling,SSP)提取更丰富的特征信息,提升分类精度。在检测部分,提出增强压缩和激活(enhanced sequeeze and excitation,ESE)通道注意力机制,在基本不影响检测速度的同时实现更加精确的多尺度目标检测。实验结果表明:该方法在由无人机、风筝、鸟等组成的数据集上的检测速度为51 FPS,平均准确率(mean average precision,mAP)为91.39%,比YOLOv3网络提高了6.42%;特别地,在无人机目标上的平均精度(average precision,AP)提升了7.42%。

Aiming at the“black flying”unmanned aerial vehicle(UAV)violating the privacy of citizens and endangering public safety,and the small objects in low-attitude were difficult to be detected by the existing UAV recognition methods with fast detection speed and high detection accuracy,the multi-scale object detection network(MS-Net)was proposed on the basis of YOLOv3 to perform the accurate and efficient detection for the small UAV in low-attitude,which can provide the theory for subsequent protection and suppression.For anchor boxes,the K-means clustering algorithm was used to get the location of the object areas more precisely.Moreover,spatial pyramid pooling(SSP)was adopted to extract more feature information and improve classification accuracy in the process of feature extraction.The channel attention mechanism named enhanced sequeeze and excitation(ESE)was proposed to achieve more accurate multi-scale object detection without affecting the detection speed in the process of detection.The experimental results show that the detection speed of MS-Net on the data set composed of UAV,kite and bird is 51 FPS,and the average accuracy(mAP)is 91.39%,which is 6.42%higher than that of YOLOv3.In particular,the average precision(AP)on the class of UAV is improved by 7.42%.