动态云台摄像机无人机检测与跟踪算法

为应对小型无人机的黑飞、滥飞对个人隐私、公共安全造成的威胁，本文采用高清云台摄像机定点巡航的方式对近地动态复杂背景中的无人机进行检测与跟踪，并提出了一种适用于动态云台摄像机的闭环无人机检测与跟踪算法，包含检测与跟踪两种模式。在检测模式下，本文设计了一种基于运动背景补偿的运动目标检测算法来提取分类候选区域，然后利用基于神经网络结构搜索得到的轻量级卷积神经网络对候选区域进行分类识别，可在不缩小高清视频图像的条件下实现无人机检测；在跟踪模式下，本文提出了一种结合卡尔曼滤波的局部搜索区域重定位策略改进了核相关滤波跟踪算法，使之在高清云台伺服追踪过程中仍能对目标进行快速稳定的跟踪；为将检测模式与跟踪模式结合在闭环框架中，本文还提出了一种基于检测概率和跟踪响应图状态的自适应检测与跟踪切换机制。实验表明，本文算法可应用于定点巡航状态的高清云台摄像机，实现近地复杂动态背景中无人机的实时准确检测、识别与快速跟踪。

Drone detection and tracking in dynamic pan-tilt-zoom cameras

To cope with the threat to personal privacy and public safety caused by illegal uses of small unmanned aerial vehicles, we propose a closed-loop UAV detection and tracking algorithm suitable for dynamic PTZ camera, which uses the fixed-point cruise mode of high-definition PTZ camera to detect and track unmanned aerial vehicles in the near-earth dynamic complex background, including two modes of detection and tracking. In the detection mode, this paper designs a moving target detection algorithm based on moving background compensation to extract the classification candidate areas, and then uses the lightweight convolutional neural network based on neural network structure search to classify and identify the candidate areas, which can realize UAV detection without reducing the high-definition video image; In the tracking mode, this paper proposes a local search area relocation strategy combined with Kalman filter to improve the tracking algorithm of kernel correlation filter, so that it can still track the target quickly and stably in the servo tracking process of HD PTZ; In order to combine detection mode and tracking mode in a closed-loop framework, this paper also proposes an adaptive detection and tracking switching mechanism based on detection probability and tracking response graph state. Experiments show that the proposed algorithm can be applied to high-definition PTZ cameras in a fixed-point cruise state, and realize real-time accurate detection, recognition and fast tracking of UAV in complex dynamic background near the ground.