中文摘要：

针对基于角点检测的视频人数统计方法中一阶静态线性模型随机误差较大的问题,提出了一种基于一阶动态线性模型回归的人数统计方法。该方法认为每帧的角点数与人数的比例关系是一个随时间变化的系数,此系数综合利用了当前帧和前一帧的角点数,在一定程度上避免角点检测突发不稳定对后续结果的影响。首先,提取视频帧的角点;其次,为了消除背景角点对统计结果的影响,算法采用光流法来估计角点的运动矢量,从而滤除静止角点;最后通过一阶动态线性模型回归出人数。在PETS2009数据库的S1.L1.Time_13-59.View_001视频段上的实验结果表明,与一阶静态线性模型相比,平均绝对误差（MAE）减小了27.0%,平均相对误差（MRE）减小了19.6%。

英文摘要：

In order to solve the problem of random errors in first order static linear models based on corner detection, a method based on first order dynamic linear regression model was proposed, based on the idea that the ratio of the number of corners per frame to the number of people is a time-varying coefficient. The corner number of the current frame and the previous frame were both considered to avoid the influence of burst instability of corner detection to a certain extent on the follow-up results. Firstly, the corners of the video frames were extracted. Then, the study employed optical flow to remove the static corners. Lastly, according to the number of corners, the people number was regressed by using a first order dynamic linear regression model. Experimental results on S1. L1. Time_13-59. View_001 vedio from PETS2009 show that compared with the first order static linear model, the Mean Absolute Error（ MAE） was improved by 27. 0% and the Mean Relative Error（ MRE） was improved by 19. 6%.