复杂场景中相关滤波跟踪算法的优化

Optimization of Correlation Filter Trackers in Complex Scenes

相关滤波跟踪器近几年在许多视觉跟踪任务中都取得了优异的性能,显示出了较高的精度和帧率。但限制其跟踪性能的问题仍有很多,例如图像表征方式单一,传统尺度估计的方式计算复杂度高,对快速运动、背景杂乱等复杂场景跟踪效果不佳。针对这些问题,提出了一个跟踪框架来分别估计目标位置和尺度的变化,通过相关滤波和颜色直方图模型相结合的方式评估目标位置,再从最可信帧中训练核尺度相关器来预测目标的尺度变化。同时,对模型采用保守的方式进行自适应在线更新,以防止误差累加破坏模型。在跟踪数据集(OTB-2015)中该算法在运动模糊、尺度变化、背景杂乱、快速运动等挑战性场景中表现出良好的优越性。

Correlation filter-based trackers have recently achieved superior performance for a lot of target tracking tasks,showing high precision and frame rates.There continue to be many issues limiting the tracking performance,such as the single method of image representation,repeated calculation of the traditional scale estimation and less robustness to fast motion,background clutter and other complex scenes.For these,we present a framework to estimate the position and scale changes separately where the position is estimated by the integration of correlation filters and colour histograms.And then a kernel scale correlator is learned from the predicted position to predict the target size.Meanwhile,the model is updated online sparingly,alleviating the drift problem caused by the accumulation of small errors.The experiments show that our tracker does well on the object tracking benchmark dataset(OTB-2015),especially in dealing with motion blur,scale transformation,background clutter,fast motion,and other challenges.