Stochastic Process Modeling for Multiple Human Tracking Using Stereo Video Camera
Stochastic Process Modeling for Multiple Human Tracking Using Stereo Video Camera
复制标题
使用立体摄像机进行多人跟踪的随机过程建模
DOI：
10.1117/12.2020425
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
Wataru
中科院分区：
文献类型：
--
作者：
Fuse;Takashi and Nakanishi;Wataru
近年来，对公共空间中行人个体行为的微观理解变得越来越重要。来自各种传感器的观测数据有所增加。与此同时，一些人类行为的仿真模型也取得了一定的进展。本文综合各种观测数据和仿真结果，提出了一种复杂场景下多人跟踪的方法。其核心思想是将多人跟踪问题视为随机过程建模。采用资料同化技术作为随机过程模拟。资料同化技术包括观测、预报和滤波。对于建模，状态向量被定义为椭圆体及其坐标，其是人的位置和形状。观察向量也被定义为来自立体视频摄像机的观察，即颜色和距离信息。然后利用离散选择模型，建立了一个描述系统状态向量动态的系统模型。离散选择模型决定了每个行人随机的下一步，并处理行人之间的相互作用。观测模型也制定了过滤步骤。颜色的可能性是基于颜色直方图匹配建模，并通过比较椭球模型和观察到的3D数据之间的范围计算之一。将该方法应用于某车站检票口采集的数据，验证了该方法的有效性。我们与其他模型的结果进行了比较，并显示了集成的行为模型的跟踪方法的优势。
Recently microscopic understanding of individual pedestrian behavior in public space is becoming significant. Observation data from diverse sensors have increased. Meanwhile some simulation models of human behavior have made progress. This paper proposes a method of multiple human tracking under the complex situations by integrating the various observation data and the simulation. The key concept is that the multiple human tracking can be regarded as stochastic process modeling. A data assimilation technique is employed as the stochastic process modeling. The data assimilation technique consists of observations, forecasting and filtering. For the modeling, a state vector is defined as an ellipsoid and its coordinates, which are human positions and shapes. An observation vector is also defined as observations from stereo video camera, namely color and range information. Then a system model which represents dynamics of the state vectors is formulated by using discrete choice model. The discrete choice model decides the next step of each pedestrian stochastically and deals with interaction between pedestrians. An observation model is also formulated for the filtering step. The likelihood of color is modeled based on color histogram matching, and one of range is calculated by comparing between the ellipsoidal model and observed 3D data. The proposed method is applied to the data acquired at the ticket gate of a station and the high performance of the method is confirmed. We compare the results with other models and show the advantage of integrating the behavior model to the tracking method.