基于多尺度帧率的视频行人再识别方法

Video-based Person Re-identification Method Based on Multi-scale Frame Rate

伴随着安防监控领域科学技术的发展和国家对安防监控领域的重视,我国已经成了世界上最安全的国家之一。无重叠视域行人再识别旨在解决通过不同视域的摄像机来识别匹配相同行人,研究对行政机关和司法机关维护社会公共安全具有重大的意义。基于深度学习的行人再识别因数据集偏小而依赖于样本数量,设计了一种基于多尺度帧率的视频行人再识别方法,通过自适应可分离卷积插帧技术生成帧间信息,增加了目标样本和运动间的细节变化特征;并且在训练中对输入进行了帧率尺度变化,提高了时-空注意力网络对行人步伐大小、周期快慢的鲁棒性。实验结果表明,提出的方法在一定程度上可以提高公开数据集的识别率,在PRID2011和i LIDS-VID数据集上进行实验,RANK1分别达到了79%和64%。

With the development of science and technology in the field of security monitoring and the emphasis on national securityin government policy, China has become one of the safest countries in the world. The non-overlapped view person re-identificationaims to solve the problem of identifying and matching the same pedestrian through cameras in different visual fields, and it is ofgreat significance to study the public secuity for administrative agencies and judicial organs. The pedestrian re-recognition basedon deep learning relies on the number of samples because of the small dataset. A multi-scale frame rate video pedestrian recogni-tion method is designed to generate inter-frame information through adaptive separable convolutional interpolation frame technolo-gy. The characteristics of the changes between the target sample and the movement and the frame rate scale change of the input dur-ing training, Improve the ASTPN network＇s robustness to the size of the pace and the speed of the cycle. Experimental results showthat the proposed method can improve the recognition rate of public datasets to a certain extent. Experiments were conducted onPRID2011 and i LIDS-VID datasets. RANK1 reached 78% and 64% respectively.