Application of Face Recognition and Automatic Inspection Technology in Subway Security System

At this stage, the main application of video intelligence analysis in the rail transit industry includes several major categories: behavior analysis, namely congestion detection, detection, reverse detection, etc.; feature recognition, ie, face recognition system; automatic image quality inspection system Automatically analyzes and alarms the image quality of the front camera. As the current application of intelligent analysis in the rail industry is still in its infancy, based on the principle of starting from the facts, the following focuses on the combination of practical engineering cases, focusing on the application of two technologies of face recognition and automatic inspection.

Application of Face Recognition and Automatic Inspection Technology in Subway Security System

First, face recognition system

There are 20 stations in the No. 1 line of XX Rail Transit, with 8 sets of face recognition equipment at the entrance gates of each station, altogether 160 sets. The facial features of the passengers entering and leaving the gates are collected in real time, and the offenders at the back office are also involved. The library compares alarms in real time.

Usually, the front face position is the best working position of the face recognition system, but as long as two eyes can be seen at the same time, the face can be recognized, and the posture change within 35 degrees will not affect the recognition effect. The track entrance gate as a special channel has natural scene advantages. The camera can easily capture the facial features of passengers and provide an effective means for the organs to arrest criminals and criminal suspects.

Face recognition includes face detection, face tracking, and face comparison techniques. Face detection refers to the presence or absence of human faces and the separation of faces in dynamic scenes and complex backgrounds. Face tracking refers to the dynamic target tracking of the detected face. The face comparison is to confirm the identity of the detected face or perform a target search in the face database.

Face detection is divided into reference templates, face rules, sample learning, skin color models, and feature sub-faces. The reference template method first designs one or several standard face templates, then calculates the matching degree between the test sample and the standard template, and determines whether there is a face through computer comparison; the face has a certain structural distribution feature, and the face rule That is, these features are extracted to generate corresponding rules to determine whether the test sample contains human faces; sample learning uses artificial neural network methods in pattern recognition, and classifiers are generated by learning the face sample set and the non-human face sample set; skin color model basis The facial skin color is detected by a relatively concentrated pattern in the color space; the feature sub-face treats all face sets as a human face subspace, and determines whether there is a face based on the distance between the detected sample and its projection in the subspace.

Face tracking generally adopts a model-based method or a combination of motion and model. In addition, skin color model tracking is also a simple and effective method. The face comparison is essentially the sequential comparison of the sampling face and the inventory face to find the best match object. Therefore, the description of the face determines the specific method and performance of face recognition. This system mainly uses the feature vector method to determine the size, position, distance, angle, and other attributes of the facial features of the pupil, the nose, the mouth, etc., and then calculates their geometric features. These feature quantities form the characteristics that describe the face. vector.

The face recognition system of the No. 1 line of the XX Rail Transit Line adopts the “local feature analysis” algorithm, which has high speed, low misidentification, and no learning. It mainly uses the data such as orientation, proportion, and corresponding geometric relations of the human face organs and feature parts to form identification parameters, and compares, judges and confirms with all original parameters in the database.

The analysis of local features originates from the principle of local statistics similar to that of building blocks. It is a calculation method based on the fact that all faces (including complex patterns) can be derived from many structural units that cannot be further simplified. Integrated. These units are formed using sophisticated statistical techniques that represent the entire face. They usually span multiple pixels (in a local area) and represent general facial shapes, but are not facial features in the usual sense. In fact, the facial structure unit is much more than the face part.

However, to form a realistic, accurate face, only a few subsets of the entire available set (12-40 feature units) are needed. To determine identity depends not only on the features of the unit, but also on their geometry (such as their relative location).

In this way, local feature analysis maps individual characteristics into a complex digital representation that enables comparison and recognition.

Second, face recognition steps

1. Create a face file: You can use camera or photo scanning methods to collect face files or directly take photo files, generate facial feature vector database, import existing database;

2. Acquiring a face of a current comparison object, capturing a face with a camera or the like to obtain a photo input, and generating face feature vector data of the comparison object;

3. Compare the face feature vector data of the current face with the existing data in the database;

4. Confirm the identity of the face or list the similarity of similar persons.

The entire process described above is done automatically, continuously, and in real time. And the system only needs ordinary processing equipment.

The workflow of the system is:

1. Automatically search for facial images in the video data stream;

2. When the system captures the full picture that appears;

3. Automatically use multiple types of matching algorithms to determine if there really is a face in that location. These algorithms can accurately detect multiple faces that appear at the same time, and can determine their exact position; once a face is detected, the face's image will be separated from the background. This image will then pass through The special treatment of the series to restore its size, light, expression and posture;

4. Convert this face image to face data within the system, which contains the unique information of this face;

5. Compare the “face data” acquired in real time with the “face data” already in the database;

6, complete the confirmation of a face.

The face recognition method of matching is based on the essential features and shapes of the face. It can eliminate changes in light, skin tone, facial hair, hair style, glasses, expressions, and gestures, and has a strong reliability, making it possible to A person is accurately identified from millions of people.

Image quality automatic inspection system

The city’s Rail Transit Line 1 has built an automatic image quality inspection system to realize the real-time detection of 1,413 standard definition cameras and 522 high-definition camera images. If there are eight conditions such as abnormal video clarity, abnormal video brightness, and video noise. , will automatically generate alarm notification operation and maintenance personnel, timely maintenance.

The common image quality inspection only monitors whether the image is lost (black screen), and there is a lack of deep-level monitoring of whether the equipment is working properly and whether the image content is correct. For all cameras used in this project, video quality diagnostics are used to detect common video failures in the surveillance system for failures due to use (snow, scroll, blur, color cast, freeze, gain imbalance, and The loss of control of the pan/tilt, etc.) Video quality diagnostics to effectively prevent image quality problems caused by hardware and unnecessary losses, providing a solid foundation for the continuous and effective monitoring of video surveillance.

Third, functional design

1, image quality intelligence inspection

Image content analysis alarms, such as: loss of video signal, blue screen, black screen, color bar, video brightness anomaly, video interference, snow, jitter, distortion, color cast, abnormal PTZ motion, picture freeze, etc.

2, cloud mirror state intelligence inspection

By sending a PTZ control instruction, it is judged whether the rate of change of the video picture reaches the standard, thereby judging whether the pan-tilt rotation and the lens zoom can work normally. Since this work may interfere with the normal work of the image monitoring personnel, this inspection function should not be frequently turned on, and Pay attention to the monitoring area that avoids major security guard activities.

3, image intelligence inspection management

According to the pre-defined inspection task, the video switching instruction is initiated automatically, and the image quality and cloud mirror status of each front-end camera are automatically checked. The interval between cameras can be set, the patrol task can be triggered at regular intervals, and an inspection report can be generated automatically.

Identification type

In the specific implementation process, we divided the video faults into eight types: abnormal video clarity, abnormal video brightness, video noise, video snowflakes, video color cast, PTZ motion out of control, picture freeze, and missing video signal.

1. Anomaly detection of video sharpness: Automatic detection of image blurring in the main part of the visual field due to improper focus, lens damage, or foreign object obstruction;

2, video brightness anomaly detection: automatic detection of video due to camera failure, gain control disorders, lighting conditions or artificial malicious cover caused by the screen too dark or too bright;

3, video noise detection: Automatic detection of video images mixed with messy "crossbars", "ripple", or other interference caused by image blur, distortion and superimposed noise and other phenomena;

4. Video snowflake detection: Automatically detect the phenomenon of snowflake or scrolling caused by the disorderly flying spots, thorns, and linear disturbances in the video image;

5, video color cast detection: automatic detection due to poor line contact, external interference or camera failure and other reasons caused by the video screen color cast phenomenon; mainly include a full-screen single color cast or a variety of mixed band strip color cast;

6. PTZ motion detection: Automatically detect whether the front-end head and the lens can move correctly according to the user's instruction, including up and down movements, and zooming in and out of the lens.

7. Screen freeze detection: Automatically detect video freeze due to video transmission scheduling system failure to avoid missing real video images;

8. Loss of video signal detection: Automatic detection of intermittent or persistent video loss due to abnormal front-end camera work, damage, vandalism or video transmission failures, including blue, black, and color bars.

IV. Conclusion

The significance of the face recognition system lies in the control of criminals or suspects who may enter or exit the subway station, enhance the active defense capability of the subway, enhance the ability to fight crimes, and enhance the deterrence of law and order. The automatic image quality inspection system can free the operation and maintenance personnel from tedious work, improve work efficiency, and ensure trouble-free system operation. It can be predicted that in the near future, intelligent analysis technology will certainly play a greater role in the track industry, gradually upgrading the current "passive prevention" to "active prevention."