低空空域小型無人機目標檢測算法

摘要： 在利用視覺檢測算法進行檢測與定位的過程中，當空域中目標無人機相對較小時，現有的檢測算法容易受到空中其他飛行物、復雜背景和光照強度變化影響導致檢測精度較低。為了解決這一問題，提出了一種基于YOLOv3改進的目標檢測算法。當空域中目標無人機體積相對較小、視覺特征較弱、存在其他干擾時，通過增加主干特征提取網絡對圖像特征提取的層級數，提取多個不同尺度特征層進行跨層連接融合，使多個不同層級的特征層之間的語義信息聯系得更加緊密，讓網絡模型可以學習到不同尺度目標的特征信息，以此增強檢測算法對小目標無人機檢測的精度。最后，利用Drone vs Birds數據集進行實驗測試，所提出的算法可以有效地提高小型無人機目標的檢測精度，檢測速度基本滿足實際要求。

關鍵詞： 低空空域；小型無人機；目標檢測；深度學習；特征融合

中圖分類號： TP39" " " " 文獻標志碼： A

doi：10.3969/j.issn.2095-1248.2023.02.007

Detection algorithm of small UAV target in low altitude airspace

WANG Chuan-yuna， SI Ke-yib

（a.College of Artificial Intelligence，b.College of Computer Science， Shenyang Aerospace University，

Shenyang 110136，China）

Abstract： When the target UAV（Unmanned Aerial Vehicle） in the air domain is relatively small， the existing detection algorithm is prone to be affected by other flying objects in the air， complex background and light intensity changes， resulting in low detection accuracy in the process of using visual detection algorithm for detection and positioning. To solve this problem， an improved target detection algorithm based on YOLOv3 was proposed. When the target UAV in the space domain was relatively small in size， with weak visual features and other disturbances， the main feature extraction network was added to the image feature extraction level， and multiple feature layers of different scales were extracted for cross-layer connection and fusion， so that the semantic information between multiple feature layers of different levels was more closely related. The network model could learn the characteristic information of targets of different scales， so as to enhance the accuracy of detection algorithm for small target UAV detection. Finally， Drone vs Birds dataset was used for experimental testing. The algorithm can effectively improve the detection accuracy of small UAV target ， and the detection speed basically meets the actual requirements.

Key words： lowattitude airspace；small UAV；target detection；deep learning；feature fusion

由于無人機技術的快速發展，無人機的操作難度逐漸降低，靈活性顯著提高，所以無人機在多個領域都得到了廣泛的普及和應用［1-3］。但利用無人機進行恐怖襲擊、侵犯公共安全等事件也隨之不斷發生，這不僅對公民隱私和生命財產安全造成了危害，也對公共安全和國家安全帶來了巨大威脅［4-6］。因此，反無人機技術的研究得到了各界的高度重視［7］。在反無人機技術中，通過基于視覺的檢測方法對無人機進行識別發現是最普遍最常用的［8］?；谝曈X的檢測方法需要采集視頻圖像信息對目標進行有效地識別及發現，其中基于深度學習的檢測算法在圖像處理方面有著優秀的性能和較高的分類準確度［9-11］。……