自适应更新上下文相关滤波补充学习跟踪算法

Real-time Tracking Complementary Learners Tracking with Adaptive Update Context Correlation Filter

为了使跟踪算法在遇到背景混乱、遮挡等问题时,能具有更好的鲁棒性,提出了一种自适应更新的上下文感知相关滤波补充学习跟踪算法(adaptive context-aware correlation tracking,ACCT)。首先,利用上下文感知相关滤波计算出响应图;接着,在更新阶段,加入了峰旁比(peak-to-sidelobe,PSR)置信度检测机制,自适应地更新相关滤波器模型;最后,与颜色直方图分类器进行线性融合,得到最终的响应值,响应值最大的位置即跟踪结果。实验上,将所提方法与先进的实时跟踪算法进行对比,结果表明,所提算法在跟踪标准测试集OTB50和TC128上,相比基准算法成功率有显著提高;并且能很好地解决背景混乱、目标被遮挡等问题。

In order to make the tracking algorithm more robust,when object suffer from background clutters and occlusions,an adaptive context-aware correlation tracking algorithm(ACCT)is proposed.Firstly,the response map is calculated by context-aware correlation filter.Then,in the update step,the reliability detection mechanism of peak-to-sidelobe(PSR)is added,and the correlation filter model is adaptive updated.Finally,the final response value is generated through combining the color response and the context-aware correlation filter response.The position of maximum response is the final tracking result.Experimentally,the results show that the proposed tracker favorable performance against several state-of-the-art tracker on the OTB50 and TC128 datasets and,it can well solve the problems such as background clutters,object occlusion.