数目可变多目标的实时跟踪

摘要

提出一种单目固定场景下,基于贝叶斯框架的数目可变的多目标实时跟踪方法.通过两个阶段实现对目标的有效跟踪:第一阶段为自动初始化,通过背景建模对视频序列进行检测,并实时提取目标的空间与颜色分布特征;第二阶段为运用粒子滤波器对目标进行跟踪与标定,通过目标间的特征匹配对对应矩阵进行实时更新,判断目标的数量变化情况及其发生概率.本文通过对道路监控视频序列中的车辆进行了跟踪仿真实验,验证了该方法的有效性和可靠性.%We propose an approach based on Bayesian framework for real-time tracking variable number of objects using fixed camera. The approach is performed at detection level and tracking level. At the detection level, a background-building arithmetic is used to extract the spatial and color distribution of objects in a complex circumstance. At the tracking level, we used particle fil-ter to track and to label objects. To analyze the occurrences and probabilities of events as continu-ation, birth and death, we update the correspondence matrix by matching features of objects. We experiment the proposed approach on video sequences and verify the effectiveness and availability of the method.