由于目标姿态、摄像头角度、光线条件等因素的影响, 行人重识别仍然是一个具有挑战性的问题。目前大多数方法主要注重提高重识别精度, 对实时性考虑较少。因此, 本文提出了一种基于增强聚合通道特征(ACF)的实时行人重识别算法。利用ACF对行人进行检测, 并在此基础上, 结合直方图特征和纹理特征构成增强ACF, 作为行人重识别的特征描述子。利用测度学习方法对重识别模型进行训练。在4个数据集上的实验结果表明, 与传统的重识别特征相比, 提出的特征描述子逼近最好的重识别准确率, 并且具有更快的计算速度。整个行人检测与重识别系统的运行速度达到10 frame·s-1以上, 基本可以满足实时行人重识别的需求。

The pedestrian reidentification is still a challenging problem due to various pedestrian poses, camera viewpoints and illumination conditions etc. Most of the reported works focus on improving the reidentification accuracy without considering the real-time capability. We propose a real-time pedestrian reidentification algorithm based on aggregated channel features (ACF). The ACF is applied to detect the pedestrian candidates, and the extracted ACF features are enhanced with histogram features and texture features and used as a pedestrian reidentification feature descriptor. Finally, based on the enhanced ACF features, we apply the metric learning to train the pedestrian identification model. The experimental results on four datasets show that the proposed feature descriptor obtains the higher recognition accuracy and much faster computation speed compared with the traditional reidentification features. The proposed pedestrian detection and reidentification framework has a running speed of above 10 frame·s-1, and it can basically meet the needs of real-time pedestrian reidentification tasks.