Can the killer technology behind face recognition with masks be commercialized on a large scale? Industry News Silver Wing

Since the outbreak of the epidemic this year, non-contact face recognition temperature measurement systems have been widely used, especially in crowded areas, which not only improves screening efficiency, but also facilitates the location and traceability of persons with abnormal body temperature.

However, although face recognition technology can effectively track confirmed cases, suspected cases, close contacts, and people in epidemic areas, the recognition efficiency is greatly reduced due to the wearing of masks and even goggles. As a result, many companies use ReID (pedestrian re-identification) technology to assist in locating and confirming the trajectory of people. In the case of wearing a mask, ReID technology can rely on the overall posture of the person for retrieval, by analyzing the wear and posture, locking the target person, and shortening the investigation time from several days to several seconds.

ReID has shown its skills in the field of security

In fact, long before the application of the epidemic, ReID has shown high application value in the field of public security.

Although our country’s face recognition cameras have spread all over the city, there are still many possible “failure” scenarios: in the face of suspects with strong anti-reconnaissance capabilities, the existing video surveillance system is difficult to help: the camera has limited coverage angle and low resolution , Insufficient capture of facial information, and difficulty in tracking suspects' tracks have become typical problems in the application of face recognition cameras in the public security field. ReID technology came into being under the huge demand of special population retrieval.

ReID technology, also called pedestrian re-identification, is a technology that uses computer vision technology to determine whether there is a specific pedestrian in an image or video sequence. That is, through a monitored pedestrian image, the pedestrian image under the cross-device can be retrieved. It is designed to make up for the visual limitations of fixed cameras, and can be combined with pedestrian detection/pedestrian tracking technology, widely used in intelligent video surveillance, intelligent security and other systems, and can provide a powerful supplement to face recognition technology in police combat.

The algorithm record is constantly refreshed, but still needs to be improved

ReID technology has received more and more attention in the industry. Whether it is an established security company, AI unicorn, or Internet giant BAT, they have deployed and accumulated this technology through algorithms, data and other aspects. Exploit greater market potential in order to lay a solid foundation for greater intelligent layout in the future.

In the past two or three years, news of companies such as Alibaba, ZTE, Yuncong, Hager Star, Pensi Technology, and Yitu, breaking the world record of the ReID data set has been frequently reported.

On the track of ReID, there are three recognized authoritative mainstream public data sets, namely CUHK03, DUKE-MTMC and Market1501. Almost all ReID competitions will be tested in these three data sets. On this basis, the ReID technical capabilities of all manufacturers are clear at a glance.

Face recognition

Table: Comparison of algorithm performance indicators of domestic mainstream ReID vendors

On this basis, the demand side can intuitively estimate the practical value of the algorithm. However, it should be pointed out that the high hit rate of Rank1 only means that the algorithm can accurately find the easiest to identify or match among many images, and it does not reflect the true ability of the model, especially the performance of dealing with complex scenes. Therefore, the mAP value needs to be combined when evaluating the performance of the ReID algorithm, which reflects the comprehensive retrieval performance of the system. The higher the mAP value, the better the practicability of the system, which can be checked both fully and accurately, and can better cope with multiple occlusions, low light, and blurred images.

But so far, even Tencent Youtu, which has the highest mAP value in Duke-MTMC, has only reached 91.1%. Compared with face recognition, there is obviously still a lot of room for improvement. Specific to commercial implementation, there are still many difficulties that need to be broken through.

Three major problems with ReID landing

Recognition of pedestrians is not easier than face recognition.

First of all, the facial features and facial shape of the face are relatively fixed, but the posture of pedestrians varies greatly during different actions, which adds a lot of difficulty to accurate recognition. Secondly, judge pedestrians based on their posture and clothes. If many people wear the same clothes, the difficulty of identification is further increased. In addition, occlusion, light, and low camera resolution are all practical problems to be solved by ReID technology.

In addition to the practical problems in the above practical application scenarios, like other technologies in the AI field, ReID's dependence on data is also limiting its further applications. Although the security industry is generating massive amounts of data all the time, the marked data in public databases is very scarce, resulting in very small data sets currently available for ReID. Compared with face recognition data sets with millions or even tens of millions and diverse identity information, ReID technology with smaller data sets still needs continuous improvement.

In addition, the ReID task in the actual scene not only requires data sets and algorithms, but also requires a large computing power and low power consumption chip that can be deployed in the front-end camera to provide support. Due to the limited space in the front-end equipment, coupled with the constraints of power consumption, cost and other factors, the intelligent front-end will be limited by hardware computing resources, and can only run relatively simple algorithms that require high real-time performance; on the other hand, the algorithm The speech is very fast, and there are certain challenges in the later operation and maintenance. Although the front-end intelligence of video surveillance has become a must for AI chip manufacturers and established security giants, due to the above factors, AI chips still have greater challenges in large-scale popularization and application. Just like face recognition, the current mainstream method is still the cloud-side integrated method, that is, front-end capture, analysis, and retrieval are performed on the back-end, which are relatively compromised in terms of chttp://www.silverwing.com.hk/ost and power consumption.
Can the epidemic accelerate the pace of ReID commercialization?

After the outbreak, Megvii launched a solution of "human body recognition + portrait recognition + infrared/visible light dual sensing". It mainly uses portrait clustering and human ReID technology to integrate infrared temperature measurement bayonet cameras, face bayonet cameras and public security The body temperature, face image and human body image data collected from the bayonet camera are fused and linked to realize real-time tracking and positioning of personnel trajectories.

Tencent Youtu combined the human body recognition under the mask to assist the community in personnel management and investigation. During the epidemic, most people who go out will wear masks, and the success rate of face recognition technology for people wearing masks will decrease. For front-line workers in the community, the failure of face recognition technology to confirm the identity of people wearing masks will greatly increase their workload of investigation and registration. Tencent Youtu and Tencent Hainer use the combination of human body characteristics and face recognition to confirm the entry and exit of people wearing masks that cannot be traced under traditional facial recognition methods, thereby improving the efficiency of community workers to register outsiders.

Just like the strategies of the two major manufacturers mentioned above, most of the current ReID implementations of each company bind face recognition and pedestrian recognition. After ID recognition and authentication through the face, it is matched to the human body through the existing binding, and then the public security The monitored data is stringed together to realize the identification and tracking of fusion linkage.

Because of this, ReID is currently more of an icing on the cake for face recognition.

However, this does not mean that it has no new space. For example, during the epidemic, taking the high-speed rail, you still need to remove the mask for face recognition when you swipe your ID card to enter the station. If ReID is integrated, you can take pictures of your clothes and faces on the day of travel, at home or in a venue with less crowds. When entering the station, there is no need to remove the mask for authentication. This is currently a possible application scenario.

In addition, ReID also helps reduce costs in certain scenarios. Because face recognition requires high camera resolution, at least 1080P or above is required, which results in high hardware costs for face capture machines. And ReID does not have high requirements for pixels, and the traditional dome camera captures. Therefore, in certain closed places, such as large supermarkets, face recognition can be bound with ReID, which can reduce the deployment of face capture machines, thereby saving a certain amount of cost. Based on the distribution of face capture machines, vehicle micro-bayonet cameras and traditional security surveillance dome cameras in a commercial block in Beijing, it is about 5-10% of faces, 10% of vehicle micro-bayonet cameras, and 80-85% of dome cameras. . In commercial blocks in second, third, and fourth-tier cities, due to cost constraints, the use rate of face capture machines may be lower, while traditional dome cameras are more common. And this has the application space of ReID.

The new retail scene is currently recognized in the industry as a landing scene. Through the cross-mirror tracking technology provided by ReID, the relationship data between "people" and "fields" can be collected and reproduced in a visual manner. It is convenient for businesses to have a stronger perception of user portraits and user behaviors, so that they can make more accurate business decisions. In addition, intelligent tracing in public places is also a more concrete application scenario for ReID.

In order to further expand the commercial scope of ReID, mainstream manufacturers in the industry have already participated in the R&D center of the relevant technology matrix, involving engineering, algorithms, databases, hardware, product forms and many other aspects. Among them, the limited database is one of the current outstanding problems, because only a large enough database can provide support in actual combat, which has become the direction of the industry's concerted efforts. In the past two months, Megvii Research Institute and Jingdong AI Research Institute have successively open-sourced the PyTorch-based ReID open source library, which is of positive significance for research and engineering deployment in related fields.

website：http://www.silverwing.com.hk/news_view_2010_153.html

Company

Industry

Product Update..

Product Inform..

AE

Amotech

Central

Harvatek

Pocons

JoYin

Damai

iA

ECS

ITW

SANICO

Sangshin

EVE

Carling

RN2

ABCO

C² Battery

Amphenol

Contact us

Can the killer technology behind face recognition with masks be commercialized on a large scale?

Related News