基于霍夫蕨的实时对象跟踪方法

Real-Time Object Tracking Based on Hough Ferns

基于霍夫变换方法难以在保持高检测精度的同时满足跟踪的实时性,且难以适应初始训练样例十分有限的情况,为解决上述问题,提出一种基于霍夫蕨的对象跟踪方法.该方法以随机蕨作为基础检测结构,将对象的局部表观作为学习数据,在其每个叶节点中计算并保存霍夫空间中属于目标对象的投票概率,在运行时通过在线学习该检测器和对象模型,适应对象表观的变化.结合对TLD跟踪框架的改进,实现了无约束环境下长时间的可视跟踪.在Babenko视频序列集上的实验结果表明,提出的对象跟踪方法在普通PC上的平均运行速率为3帧/s,平均准确率为87.1%,总体上优于现有的跟踪方法.

In order to deal with the tough problem of providing high accuracy and meanwhile achieving real-time tracking using Hough-based approaches under very limited samples for training, a Hough ferns based method was proposed for object tracking. This method uses the random ferns as the basic detector. It samples the local appearances of object as training set, and computes and saves the Hough votes for each leaf-node. The detector and object model were learned online at runtime to adapt to the variation of object and the TLD （tracking-learning-detection） was improved to achieve long-term visual tracking in unconstrained environment. Experimental results on Babenko sequences demonstrate that the average running speed of the tracker based on the proposed approach on a normal PC is 3fps and the average accuracy rate is 87.1% , showing its better tracking performance than several state-of-theart methods.