一种无人机航拍目标的长期跟踪算法

A long-term tracking algorithm for UAV aerial photography

针对无人机在航拍目标跟踪的复杂场景过程中,运动目标可能会被遮挡或不确定跟踪,导致视觉模型出现逐渐损坏、漂移和不可逆转失败等问题,提出了一种无人机航拍目标的长期跟踪算法。首先,进行互补分类器多特征自适应融合设计,在贝叶斯分类器中采用颜色直方图特征,在相关滤波器中采用方向梯度直方图、灰度以及颜色名特征;结合多种特征的优点,构建目标鲁棒性的外观以适应复杂场景。然后,在相关滤波器中加入自适应时空正则化项。在空间正则化参数中引入局部变化,实现学习时限制像素可信度较低的滤波器;而在时间正则化中,根据全局响应自适应地调整滤波器的学习,并用初始滤波器约束更新范围,这样在缓解边界效应的同时,有效防止滤波器退化。最后,在上述基础上,加入重检测模块,使得跟踪过程更加准确。实验结果表明,本文算法可以适应无人机航拍的复杂场景,缓解边界效应和防止滤波器退化。与同类主流算法相比,在目标经历严重遮挡、移出视野等情况后,仍然满足实时性需要,获得了较好的跟踪效果。

In view of the problem that the moving target may be occluded or tracked under uncertain conditions during the aerial target tracking of UAV in complex scenes,resulting in the gradual damage,drift,and irreversible failure of the visual model,a long-term tracking algorithm for UAV was proposed.Firstly,the complementary classifier was designed for multi-feature adaptive fusion.The color histogram feature was used in Bayesian classifier,and the directional gradient histogram,grayscale,and color name features were used in the correlation filter.In combination with the advantages of multiple features,the robust appearance of the target was built to adapt to the complex scenes.Secondly,an adaptive spatial-temporal regularization term was added to the correlation filter.Local changes were introduced into the spatial regularization parameters for the implementation of the filter with low pixel credibility during learning.In temporal regularization,the learning of the filter was adaptively adjusted according to the global response,and the initial filter was used to constrain the update range,which effectively prevented filter degradation while mitigating boundary effects.Finally,a re-detection module was added to make sure the accuracy of the tracking process.Experimental results show that the proposed algorithm could adapt to the complex scenes of UAV aerial photography,alleviate boundary effects,and prevent filter degradation.In comparison with similar mainstream algorithms,the proposed algorithm could still meet the real-time requirements and achieve better tracking effect even when the target experienced serious occlusion or moved out of view.