上下文感知自适应的相关滤波跟踪算法

Context-Aware Adaptive Correlation Filter Tracking Algorithm

针对相关滤波类目标跟踪算法在目标出现遮挡、目标快速移动、复杂背景等情况下跟踪精度低的问题,提出了一种上下文感知自适应的相关滤波跟踪算法。首先,提取目标周围8个方向的背景样本为相关滤波器提供训练样本,然后利用粒子滤波对目标的运动状态进行估计,预测目标的运动方向。在训练滤波器时,给予目标运动方向上的背景样本更多的权重;接着,引入了一种新的模型更新判别依据APCE,只有当APCE值和响应最大值同时分别以一定比例大于各自的历史平均值时,才对模型进行更新;最后将上述算法与当前一些主流的跟踪算法在基准测试集OTB100上进行实验对比。实验结果表明,所提算法的成功率为0.647,精确度为0.866,与其中最优算法相比,分别提高了4.7%和7.3%。且上述算法具有较强的鲁棒性。

Aiming at the problem of low tracking accuracy of correlation filter target tracking algorithms in the case of target occlusion, fast target moving, complex background, etc., a context-aware adaptive filtering tracking algorithm is proposed. First, the background samples in eight directions around the target were extracted to provide training samples for the correlation filter. Then, particle filtering was used to estimate the motion state of target and predict the motion direction of target. When training the filter, more weight was given to the background samples in the direction of the target’s movement;Then a new model was introduced to update the discrimination basis APCE,the model was updated only when the APCE value and the maximum response value were larger than their historical averages respectively in a certain ratio at the same time. Finally, the algorithm in this paper was compared with some current mainstream tracking algorithms on the benchmark test set OTB-100. The experimental results show that the success rate of the algorithm in this paper is 0. 866 and the accuracy is 0. 647. Compared with the optimal algorithm,it increases by 4. 7 % and 7. 3%, respectively. This algorithm has strong robustness.