From Face Recognition to People Recognition to See the New Development of Re-ID Technology 【Full Text】

At the current stage of security exhibition network technology development , face recognition in standard scenes has been able to achieve very mature applications, but in some non-standard scenes, the application effect of face still needs to be improved. Because face information is relatively simple and easily affected by external environmental interference, the recognition results have been affected. Therefore, in recent years, for cross-camera and cross-space target person retrieval, pedestrian re-recognition technology Re-ID is receiving more and more attention attention.

Person Re-Identification (Person Re-Identification, Re-ID for short) is also known as cross-mirror tracking technology. It uses computer vision technology to determine whether a specific pedestrian exists in an image or video sequence. It refers to whether it is worn, posture, hairstyle, etc. The AI ​​vision technology that can identify the same person again in a certain scene and depict the individual's trajectory in this way. In short, it uses computer vision technology to retrieve the same target person under different cameras.

Re-ID is different from face recognition. Face recognition is used to authenticate a person's identity, and Re-ID is to match images of the same person under different camera devices to generate a cross-camera spatiotemporal trajectory. A typical application based on pedestrian ReID is to search for pictures with pictures.

Over the past few years, with the increasing demand for pedestrian re-identification technology in public security, transportation, retail and other industries, many manufacturers have also published the results of relevant data set tests to highlight the company's Re-ID technology research. Results. Among them, Market 1501, DukeMTMC-reID, and CUHK03, three data set tests that measure Re-ID technology are also commonly used pedestrian re-identification data set tests in the industry. In the field of Re-ID technology research, Rank-1Accuracy and Mean Average Precision (mAP) are the core indicators to measure the algorithm level.

The security knowledge network editor sorted out the test results of some domestic enterprises in the three major public data sets in the past two years. As shown in the figure below, under the continuous iteration of algorithm technology, taking the Market 1501 data set as an example, the first hit rate and the average accuracy They have successively surpassed the level values ​​of 97% and 94%, pushing the pedestrian re-identification algorithm level to usher in a new breakthrough. The test results of some domestic enterprises on the three major public data sets in the past two years

In addition to some companies in the field of intelligent security and artificial intelligence, the list of Ali and Tencent Youtu also made the industry pay more attention to the Re-ID technology, laying a preliminary foundation for the commercialization of this technology. .

Pedestrian re-identification Re-ID technology research several major difficulties:

At this stage, Re-ID's technical research still faces many practical problems and technical difficulties. These problems mainly include the difficulty of data acquisition, the difficulty of algorithm training, and the input-output ratio of some practical perspectives.

Data acquisition is difficult: compared with face data, pedestrian data in Re-ID is seriously scarce. The pedestrian mainstream data set (Market1501) only has 1000-3000 pedestrian IDs, and the public data set ID of the face has exceeded 1 million. , The scale of the enterprise's private ID may be larger. The main reason for this phenomenon is that pedestrian data sets need to be collected from the same person and appear under multiple cameras at the same time for a period of time. Such severe conditions limit the construction of pedestrian data sets. Because of the lack of data, higher requirements are placed on the algorithm research of cross-mirror tracking technology.

Algorithm training is difficult: the scarcity of data itself is a big problem. On this basis, existing video surveillance is also restricted by factors such as imaging quality and resolution, which can also lead to blurred image information. Of course, there are also factors such as large differences in camera shooting angles, changes in indoor and outdoor environments, changes in pedestrian clothing accessories, large differences in seasonal dressing styles, and differences in light during the day and night. This allows cross-camera, cross-region, and cross-time Re-ID Analysis becomes more difficult, and the actual problem to be solved is very complicated. In addition, in many security systems, the code stream bandwidth of the video surveillance probe and the density of the camera deployment will also limit the accuracy of the Re-ID algorithm.

At the same time, users have greater concerns about the input-output ratio. To apply on products, land on projects, and improve accuracy by only a few percentage points, but the amount of calculation, memory overhead, and storage overhead have increased significantly, and customers are definitely not willing to pay. In the case of inaccurate accuracy, applying technology to products requires a lot of effort to do application innovations, such as computing speed and memory overhead.

In response to the above-mentioned problems related to data collection and algorithm training, some companies in the industry have already launched some technological breakthroughs:

Ali: Through the mining of local information, we focus on solving the problem that the pedestrian's apparent posture changes drastically during the recognition process and is not easy to align. On the one hand, parts with strong semantic information are obtained through the human body, and they are used to find distinctive areas in them. On the other hand, a pyramid-based horizontal partitioning strategy is used to obtain identifiable information for pedestrian fixed areas. In the training, a combination of two strategies is used to achieve the alignment of pedestrian pictures, thereby achieving more matching recognition.

ZTE: Innovatively propose a multi-module multi-granularity joint feature extraction network, which effectively solves the complex situations such as side faces, occlusion, and missing caused by camera shooting angle, color difference, light intensity, etc., and significantly improves the network feature matching performance .

Pengsi Technology: Using global features to measure the weight of each frame of pictures often loses a lot of important information. A segmentation and recombination strategy is used to reconstruct specific local features into multiple video sequences for learning, thereby greatly reducing the impact of local feature loss on the final features; secondly, a new two-way graph attention mechanism module is proposed. Combining graph convolutional neural network and SEnet, the channel domain mode selection learning is performed on the entire sequence. At the same time, the attention area learning in the spatial domain is conducted through the two-way network. Due to the characteristics of graph convolutional networks, the attention features of each frame of pictures are the result of learning and combining with other frames, which greatly improves the representativeness of the features; finally, the sequence fusion is performed using the similarity between frames. In this way, the data The intra-class similarity has been greatly improved. After training with the ternary loss function, the similarity between classes is reduced, and the re-identification effect is improved.

Dahua: It applies image data enhancement methods, which mainly include strategies such as random blurring and random interception, which can effectively simulate complex situations such as human occlusion, blurring and incompleteness in various environments; second, for multi-branch component networks For the feature granularity difference problem, a progressive part network model PPM (Progressive Part Model) is adopted. In addition to the basic convolutional network part shared between branches, there is also a cascading semantic relationship; In the network, through the overlapping sampling operation, each component branch is promoted to extract more significant feature information, and the improved loss function is used to learn the feature embedding space based on spherical constraints.

Qianshitong: Based on years of actual combat experience, it proposes a card-vision linkage technology and warfare. Combining the characteristics and advantages of face recognition and cross-mirror tracking (Re-ID), a small number of face mounts plus a large amount of ordinary monitoring The deployment of the probe can not only lock the identity of the suspect, but also reproduce the trajectory of the suspect. This innovative comprehensive technical warfare method can cover the monitoring range with a high probability, and has strong practical value.

Singularity Cloud: Focus on overcoming two core technologies of pedestrian re-identification algorithm: block-based random discarding subspace feature enhancement attention mechanism with low utilization rate, multiple loss fusion enhancements, making the original Re-ID 93% accuracy rate Get promoted.

Tencent Youtu: Alternate training of multi-task framework, block-based pyramid model. Through the successful application of these two core technologies, Tencent Youtu has achieved at least a 6.34% improvement in the index from the original baseline model on all three mainstream databases.

Yuncong Technology: The combination of global features and multi-granular local features not only captures the unifying features but also details, such as clothing LOGO, backpack ornaments, etc., increases the retrieval and identification of feature elements and improves accuracy.

It can be said that Re-ID pedestrian re-recognition technology raises the cognitive level of artificial intelligence from "face recognition" to a new stage of "person recognition", which can be played under the application scenario of person target retrieval across space, time, and region. More important than face recognition.

For example, typical AI tracing applications such as lost person retrieval and suspect tracking. Another more typical application scenario is the heat flow statistics of passenger flow in shopping malls and supermarkets, which can help the mall to improve the store display and shopping experience through real-time dynamic tracking of user trajectories (under the premise of privacy protection). With the maturity of related technologies, it is believed that the future Re-ID technology will improve the traditional "find people" and "find things" models in more scenarios, bringing about a significant improvement in operating efficiency.