结合核相关滤波器和深度学习的运动相机中无人机目标检测

针对无人机与相机快速相对运动造成的运动模糊问题，以及小型无人机外观信息缺失和背景复杂造成漏警和虚警问题，提出了一种新的无人机检测-跟踪方法。针对成像尺寸小于32像素×32像素的无人机目标，提出改进的多层特征金字塔的分类和目标框回归器作为目标检测器，克服漏警。利用检测结果初始化基于核相关滤波的目标跟踪器，并持续修正跟踪结果，跟踪结果为剔除检测器虚警提供依据。在跟踪过程中，引入对观测场景纹理自适应的相机运动补偿策略实现目标重定位。多场景下的实验结果表明：提出的方法在对高速运动小目标的检测和跟踪指标上显著优于传统方法，且运动补偿机制的引入进一步增强了方法在极端复杂场景下的鲁棒性。

UAV detection in motion cameras combining kernelized correlation filters and deep learning

To solve the problem of motion blurs caused by the rapid relative movement of the UAV and the camera, and the missed detection and false detection problems resulted from the lack of appearance information and complex background of small drones, a new drone detection-tracking method is proposed. Aiming at UAV targets with imaging sizes less than 32 pixel×32 pixel, an improved multi-layer feature pyramid classification and a target box regressor are proposed as target detectors to overcome missed detection. The detection result is used to initialize the target tracker based on kernelized correlation filters, and continuously modify the tracking result which provides a basis for the elimination of false detection. During the tracking process, a camera motion compensation strategy adaptive to the observed scene texture is introduced to achieve target relocation. Experimental results in multiple scenarios show that the proposed method is significantly better than traditional ones in the detection and tracking of small high-speed moving targets. In addition, the introduction of motion compensation mechanism further enhances the robustness of the method in extremely complex scenarios.