稀疏约束与时间一致的背景感知相关滤波目标跟踪

Sparse Constrained and Time-consistent Background-aware Correlation Filtered Target Tracking

背景感知滤波算法通过循环移位采集真实负样本,有效解决了边界效应.但在复杂场景例如遮挡、快速移动、背景干扰等,其较大的采样区域导致过多背景在杂波干扰,从而影响跟踪效果.针对这一问题,本文首先提取灰度HOG特征与颜色CN特征来提高目标外观模型,在基准目标函数基础上引入L1稀疏正则约束形成弹性网络以自适应筛选关键特征,增强滤波器在复杂背景下的判别能力.同时针对BACF在跟踪过程中目标快速变化,本文引入时间正则项提高滤波器抑制畸变的能力.最后,本文提出了一种独立的尺度滤波器算法,准确提供目标尺度大小.实验仿真结果表明,在公开数据集OTB-2013和OTB-2015上,本文算法较基准算法有很大提升,能够较好应对不同复杂场景下的跟踪难题.

The background-aware filtering algorithm collects real negative samples by cyclic shift,which effectively solves the boundary effect.However,in complex scenes such as occlusion,fast movement and background interference,its larger sampling area leads to too much background in clutter interference,which affects the tracking effect.To address this problem,this paper first extracts grayscale HOG features and color CN features to improve the target appearance model,and introduces L1 sparse canonical constraints on the basis of the baseline objective function to form a resilient network for adaptive screening of key features to enhance the discriminative ability of the filter in complex backgrounds.Meanwhile,for the rapid change of the target in the tracking process of BACF,this paper introduces a temporal regularity term to improve the filter′s ability to suppress distortion.Finally,an independent scale filter algorithm is proposed in this paper to provide the target scale size accurately.The experimental simulation results show that the algorithm in this paper is greatly improved over the benchmark algorithm on the public datasets OTB-2013 and OTB-2015,and can better cope with the tracking challenges in different complex scenarios.