Abstract: Pedestrian re identification, as a hot topic in machine vision, is widely used in fields such as public security and safety. In recent years, many excellent solutions in deep learning have been developed to address issues such as insufficient lighting, low image pixels, and constantly changing pedestrian poses in unobstructed scenes with sufficient lighting. However, in practical applications, pedestrians are often obstructed by other pedestrians and objects, leading to problems such as missing important comparison parts and inability to align features during matching. Therefore, compared with general pedestrian re identification, pedestrian re identification under occlusion conditions is more challenging. Firstly, the characteristics of different datasets were compared and the evaluation criteria for datasets were introduced; Secondly, from the perspective of feature learning, commonly used occlusion datasets and processing methods under occlusion conditions were summarized, and the characteristics and differences of five feature learning methods were compared in detail; Again, some representative methods were analyzed from the perspective of improving the benchmark model, and the advantages and disadvantages of various feature learning methods were compared; Furthermore, by evaluating indicators, the performance of existing feature learning methods on the dataset is systematically analyzed; Finally, the problems and future development directions of pedestrian re identification under occlusion conditions were discussed, as well as its application in public security work.

Key words: Pedestrian Re-Identification, Deep Learning, Occlusion, Feature Learning, Convolutional Neural Network

摘要： 行人重识别作为机器视觉的一个热点问题，被广泛应用于治安、安防等领域。针对光照不充足、图片像素低、行人姿势不断变换等问题，这些年在深度学习中很多优秀的解决方案一般在光照充足的无遮挡场景下的，然而在实际场景应用中，往往伴随着行人被其他行人和物体遮挡，导致进行匹配时会面临着重要比对部位缺失、特征无法对齐等问题。因此，与一般的行人重识别相比，遮挡条件下的行人重识别更加具有挑战性。首先，对比了不同数据集的特点和介绍了数据集评价标准；其次，从特征学习的角度总结了常用的遮挡数据集及遮挡状态下的处理方法，详细比较五种特征学习方法特点与区别；再次，从基准模型的改进方向分析了一些代表性的方法，并对比各种特征学习方法下的优缺点；继次，通过评价指标，系统分析现有特征学习的方法在数据集上的性能表现；最后，讨论了遮挡条件下的行人重识别的面临的问题和未来的发展方向，以及在公安工作上的应用。

关键词: 行人再识别, 深度学习, 遮挡, 特征学习, 卷积神经网络