基于双模型融合的监控视频人数统计

People Counting from Surveillance Video Based on Dual Models Fusion

针对垂直角度监控视频下的人数统计任务所遇到的角度特殊、检测精度较低等问题,提出了一种基于双模型融合的监控视频人数统计方法。在基于COCO数据集的YOLOv3预训练模型上,加入了垂直角度的训练样本进行迁移学习。分别训练两个模型,一个检测俯视视频中的行人,另一个检测俯视视频中的人头。将迁移学习训练后的行人检测模型和人头检测模型的检测结果进行融合,并引入基于卡尔曼滤波的跟踪方法得到了运动轨迹,利用运动轨迹信息实现了人数统计。通过多次实验,验证了在垂直角度监控视频中双模型融合方法较单一检测模型和预训练模型能取得更好的检测效果,最终获得更准确的人数统计结果。

In order to relieve the problem of low people counting accuracy due to the angle of the surveillance camera,a method based on dual models fusion of people counting is proposed in this article.In the transfer learning step,vertical view training samples are added into the pre-training model of YOLOv3,which is based on COCO data set and can hardly detect the people in overlook video.Then we train two separate model independently,one can detect the people and the other can detect the head in vertical view surveillance video.After fusing the people detection model and the head detection model,a tracking method based on Kalman filter is added to obtain the people’s trajectories.Finally the trajectories-based people counting is performed.Through experiments,it is verified that the dual model fusion method in the vertical angle surveillance video has better detection results than the single detection model and the pre-training model,and better detection leads to more accurate results of people counting.