Article Preview
Top1. Introduction
In recent years, with the rapid development of image acquisition and large-scale data storage technologies, large scale camera-nets have produced rich video surveillance information. Video analysis technology can realize the recognition and tracking of designated pedestrians in video, which greatly releases the pressure of manual viewing, and improves the effectiveness of monitoring system. Person re-identification is to identity the target persons in the video sequences from the non-overlapping camera view. Re-identification (re-id) tries to solve the problem of “where the target person was presented in the past” or “where the target person was captured in the surveillance in the future”. Although in a controllable environment, it is a relatively mature technology to recognize persons based on human faces and other biological features. However, because the environment of monitoring video is complicated and flexible, the resolution of the extracted person images is low. So it is difficult to obtain robust face features. Therefore, person re-id often uses the appearance features of the clothes and some significant things carried with target persons.
The re-id problem can be summarized as follows: there are several cameras from non-overlapping views. We obtain a number of person images by preprocessing technology. The images in cameras 2, 3… defined as probe and images in camera 1 defined as gallery. The aim of re-identification is to find out special images containing the target persons which are appointed in probe. The sketch map of re-id as shown in Figure 1.
In the application of the real scenes, the re-id faces some essential problems:
- 1.
The differences of views result in the misalignment of person appearance. In fact, the in-stallation of cameras is usually fixed in practical applications, but the visual angle under various cameras is often different. This diversity can even result in the problem that some features which can be seen in camera A will not be found under camera B. The deviation of appearance position will cause the same person could not be discovered in different videos. In existing public databases, more than 90% of person images exist this difference of perspective.
- 2.
The traditional distance function does not consider the sample characteristics and this problem lead to a weak discriminant. The distance functions such as Euclidean distance, cosine distance and correlation distance have bad discriminant ability because they do not take the distribution characteristics of samples into account. According to the experimental statistics, the recognition rate of the measure function based on sample distribution learning in VIPeR database is about 32% higher than the traditional distance ones.
The key technologies of person re-id include feature representation, feature transformation, distance measurement and sorting. These days, more scientists put their concentrate on two technologies: feature representation and distance learning. Typical algorithms of person re-id now can be ranked as follows:
- 1.
Design a robust feature representation. In order to design a stable and discriminant feature representation, the dimensions of feature vectors have been increasingly high, from thou-sands to tens of thousands.
- 2.
Learn a good distance metric. Learning a robust and discriminant metric function or cross-view subspace can improve re-id performances.
Nevertheless, due to the differences in visual angle and illumination of different monitoring videos, the extracted features of persons change sharply, which may lead to a more similar appearance of distinct persons than the same person in a number of videos. Hence, a stable feature representation is the first step to achieve person re-id. Then, academic circle often calculates feature vectors of two-person images to match them, and require the distance between same persons must less than different ones. Above all, we can conclude that feature representation and metric learning are two key technologies in person re-id.