自适应时空正则的无人机目标跟踪算法

针对无人机跟踪场景中目标分辨率较低且易受无人机(unmanned aerial vehicle,UAV)飞行姿态、速度变化等因素的影 响而难以对目标进行鲁棒跟踪的问题,提出了一种自适应时空正则的无人机目标跟踪算法以 有效解决上述问题。在时空正则相关滤波器(spatial temporal regularized correlation filter,STRCF)算法基础上引入AutoTrack中的空间正则性代价并利用峰值 旁瓣比和局部响应变化量,在线动态更新时空正则化参数以提升跟踪器的准确性,通过在跟踪 器中嵌入遮挡处理模块解决目标遭遮挡后跟踪漂移的问题。在多个无人机基准数据集上进行 了测试,实验结果表明,与基准算法AutoTrack相比,本文算法具有更高的精确度和更快的 处理速度。其中在DTB70数据集上跟踪精度和速度分别提升了1.5% 和74.4%；在UAVDT 数据集上9个属性的分类对比中,本文算法在尺度变化(scale variation,SV)、目标模糊(object blur,OB)等7个属 性上取得较高的性能,均排在第一位。由此可见本文算法可以更好地满足无人机应用需求。

UAV target tracking algorithm based on adaptive spatial-temporal regularization

Aiming at the problem that the target resolution in unmanned aerial vehicle (UAV) tracking scene is low and it is difficult to track the target robustly due to the influence of UAV flight attitude,speed change and other factors,an adaptive spatial-temporal regularization algorithm for UAV target tracking is proposed.Based on the spatial temporal regularized correlation filter (STRCF) algorithm,the cost of spatial r egularity in AutoTrack is introduced,and the spatial-temporal regularization parameters are dynamically updated online by using the peak side lobe ratio and local response variation ve ctor to improve the accuracy of the tracker.Occlusion processing module is embedded in the tracker to solve the tracking drift problem after the target is occluded.Experimental r esults on several UAV benchmark datasets show that the proposed algorithm has higher accuracy and faster processing speed than the benchmark algorithm Autotrack.The tracking accuracy a nd speed of DTB70data set are improved by 1.5% and 74.4% respectively.In the classifica tion comparison of nine attributes on UAVDT data set,this algorithm achieves high pe rformance on seven attributes,such as scale variation (SV),object blur (OB),etc.It can be seen that this algorithm can better meet the requirements of UAV applicat ions.