中文摘要：

在现代监控和视频检索系统中,跨摄像头行人跟踪问题仍然是个挑战,其原因是行人再识别时庞大的搜索空间,特别是当有大量的摄像头和行人的时候。针对跨摄像头行人再匹配计算量大,耗时久等问题,提出一种融合样本数理统计和混合高斯分布的时空关系模型。该模型可以有效地预测行人活动,即当一个行人从一个摄像头可视区域离开时,我们能够预测该行人下一次直接进入摄像头可视区域的时间和所在的出入点位置。根据预测的结果,极大地减少了行人再识别的匹配范围,从而提高匹配识别的准确率,再依赖行人的表现模型和轨迹进行行人再识别,最终实现跨摄像头行人持续跟踪的目的。实验结果表明了模型的表现与实际情况比较接近。

英文摘要：

In the modern monitoring and video retrieval systems,it is still a challenge to track pedestrian from multiple camera. The reason is that there is large search space in the process of the pedestrian re-identification,especially when there are a large number of cameras and pedestrians. In view of this,a spatiotemporal relationship model based on the sample statistics and mixture Gaussian distribution for multiple cameras pedestrian tracking is proposed.The model is able to predict the pedestrian activities effectively,which means,when the pedestrian disappears from one camera view,the model is able to predictthe time and place that the pedestrians will appear in another camera view.According to the results of prediction,the matching range of pedestrian re-identification is greatly reduced and the accuracy rate of matching is improved. The purpose of multiple cameras pedestrian tracking is finally realized by reidentifying the pedestrians＇ expression model and track. The experiment result demonstrates that the performance of the model is consistent with the observations in the extensive experiments.