结合自适应空间权重的改进型时空正则项跟踪算法

An Improved Spatial-Temporal Regularization Term Tracking Algorithm Combining Adaptive Spatial Weights

为解决时空正则项的相关滤波视觉跟踪算法在目标部分遮挡时存在的模型漂移和尺度估计不准确问题,提出了结合自适应空间权重的改进型时空正则项跟踪算法。采用平均特征能量比将无法准确表达目标或过多表达背景信息的特征通道裁剪掉,以提高跟踪精度。在滤波器训练时加入空间权重正则项,利用时间正则项在目标遮挡时被动更新滤波器,使得在空间权重更新时更为准确,以此着重学习目标未被遮挡部分,获取可靠的相关滤波器系数。将滤波器求解划分为2个子问题,分别采用交替方向乘子法进行优化计算,保证算法运算速率。在牛顿迭代法中设置精度阈值,在保证定位精度的同时减少迭代次数。实验结果表明:在OTB-100数据集上所选择的6个视频序列中,所提算法较STRCF算法的平均中心位置误差降低了12.3像素,平均重叠率增加了7%,运算帧率可达19.25帧/s;在OTB2015遮挡视频序列中,所提算法较STRCF算法的成功率曲线下积分面积(SAUC)增加了0.7%,使用深度特征的所提算法较DeepSTRCF和ASRCF算法的SAUC分别提升了3.9%与0.9%。

To solve the problem of model drifting and inaccurate scale estimation for partially occluded object in spatial-temporal regularized correlation filter visual tracking algorithm,an improved spatial-temporal regularization term tracking algorithm combining adaptive spatial weights is proposed.For cutting out the feature channel that unables to express the background information more accurately,the average feature energy ratio channel is chosen to improve the tracking accuracy.Then the spatial weight regularization term is added to the filter training,and the time regularization term is used to passively update the filter when the target is occluded,so that the spatial weight update gets more accurate.Focusing on learning the non-occluded part of the target,the reliable correlation filter coefficient is obtained.The filter solution is divided into two subproblems,the alternating direction multiplier method is used for optimization calculation to ensure the algorithm operation rate.The accuracy threshold is set in Newton iteration method to reduce the number of iterations as ensuring the positioning accuracy.In the selected video sequences of OTB-100 dataset,the proposed algorithm reduces the center position error by 12.3 pixels,the average overlap rate increases by 7%,and the frames per second reaches up to 19.25,compared with those of the algorithm STRCF.And the integral area under the success curve of STRCF in the occlusion video sequence of OTB2015 dataset is increased by 0.7%,while the integral area of DeepSTRCF and ASRCF is increased by 3.9%and 0.9%,respectively,by the deep feature version of the proposed algorithm.