视觉传感网络分布式在线数据关联

Distributed Online Data Association in Visual Sensor Networks

数据关联是视觉传感网络监控系统的基本问题之一.本文针对无重叠视域视觉监控网络的多目标跟踪问题提出一种基于多外观模型的视觉传感网络在线分布式数据关联方法,将同一目标在不同摄像机节点上的外观用不同的高斯模型描述,由分布式推理算法综合利用外观与时空观测计算关联变量的后验概率,同时通过近似最大似然估计算法对各传感节点上的外观模型参数进行在线估计.实验结果表明了所提方法的有效性.

One of the fundamental requirements for visual surveillance with smart camera networks is the correct associa- tion of camera＇s observations. In this paper, we present a distributed online approach based on multiple appearance models for multi-object tracking with distributed non-overlapping cameras. Firstly, we use multiple Gaussian models to describe each object^s appearances under different camera nodes. Secondly, we develop a novel distributed online framework, in which the posterior margins of association variables are calculated using appearance and spatio-temporal information by a distributed inference algorithm, and the model parameters are updated online on each camera by approximate maximum likelihood estimation. Experimental results show the validity of the proposed method.