空时上下文模型下基于多种特征融合的监控目标跟踪

Object Tracking Algorithm Based on Multi-feature Fusion in Spatio-temporal Context Model

随着视频处理技术和网络技术的发展,视频监控应用逐渐渗透到了人们日常活动中的方方面面,如何设计实现精度高、鲁棒性好的目标跟踪技术仍然是当今研究的热点及难点;在工程应用实践的基础上,提出一多特征融合与自适应模型更新的空时上下文目标跟踪算法,通过将丰富多样的多特征信息整合到空时上下文模型中;由于多特征具有互补特性,可以克服单一特征对目标区域描述不足的缺陷,提升算法的抗干扰能力;同时,也提出了一种自适应学习因子策略,增强了模型的泛化能力;选取的特征集是鲁棒的,包括了颜色、梯度、方向、点特征等总共19个特征,其中子块大小是11×11,高斯核方差为2,损失项正则参数为0.005,其余参数设置与STC保持一致;大量的仿真实验结果表明所提出的改进算法在跟踪中心误差指标上比现有的KCF,MFC和STC跟踪算法分别提高了5.4%,2.1%和3.6%,对复杂的跟踪场景具有更强的鲁棒性与抗干扰能力。

With the development of video processing and network technology,video surveillance applications gradually penetrated into every aspect of people＇s daily activities.How to design a object tracking technique with high precision and robustness is still a hotspot and difficulty in current research.an improved spatio-temporal context tracking algorithm based on multi-feature fusion and adaptive model updating is proposed.Based on the spatio-tempora context tracking algorithm,our proposed algorithm integrates multi-feature informations into the spatio-tempora context model.Since the complementary characteristics of multiple features,it is possible to overcome the disadvantages of the single feature and improve the anti-jamming ability.In addition,this paper also proposes an adaptive learning factor strategy to enhance the generalization ability of the model.The selected feature set is robust,including color,gradient,direction,point feature,and so on.There are 19 features in total,where the size of the sub block is 11×11,the Gauss kernel variance is 2,the parameter of loss term is0.005,and the other parameters are consistent with the STC.A large number of simulation results show that the tracking performance of our proposed algorithm outperforms the existing KCF,MFC and STC tracking algorithm,and has stronger robustness and anti-jamming capability for complex scenes.