基于关键点建模与弱监督外观更新的多目标跟踪  

Multi-target Tracking Based on Key-point Modeling and Weakly Supervised Appearance Updating

作　　者：张国平[1] 周改云[1] 马丽[1] 

机构地区：[1]平顶山学院软件学院,河南平顶山467000

出　　处：《计算机工程》2016年第8期261-265,共5页Computer Engineering

基　　金：国家自然科学基金资助项目(61503206);河南省科技厅科技发展计划基金资助项目(132102310516);平顶山学院青年基金资助项目(PDSU-QNJJ-2013010)

摘　　要：针对多目标跟踪在复杂场景中的遮挡、漏检和噪声问题,利用关键点建模和弱监督外观模型更新,提出一种改进的多目标检测与跟踪方法。使用角点检测器获得关键点及其绝对位置,运用背景差分法得到图像的二值映射。根据图像映射将关键点分为显著关键点和微弱关键点,利用显著关键点构造候选模型,并应用弱监督外观模型对目标跟踪框进行更新,从而实现多目标检测。在多个视频集上的实验结果表明,与基于高斯混合概率密度滤波器的跟踪方法、连续前向估计的多目标跟踪方法相比,该方法具有更高的多目标跟踪精度及更快的运行速度。Concerning the occlusion, missing detection and noise in multi-target tracking under complex scene, an improved method of multi-target detection and tracking using key-point modeling and weakly supervised appearance model updating is proposed. Firstly, a corner detector is used to get the key-points and theirs absolute position, and background subtraction is applied to obtain binary mapping. Then, key-points are classified into significant key-points and weak key-points by mapping images. Significant key-points are used to build candidate models. Finally, weakly supervised appearance model is used to update tracking frames and realize multi-target detection. Experimental results on several video sets show that compared with GM-PHD tracking method and multi-target tracking method of continuous forward estimation, the proposed method has higher multi-target tracking accuracy and faster running speed.

关 键 词：多目标跟踪 角点检测 关键点 弱监督外观模型 跟踪精度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]