Security video analysis

The intelligent video system is composed of a video analysis server located at the front end or the back end. It analyzes the video images captured by the surveillance cameras and can separate the status of people, vehicles or objects in the images from any background, identify and analyze them. track. Comparing the behavior patterns of the tracked objects and the pre-set security rules. If any violations are found, an alarm notification will be made immediately. At the same time, information records or displays will be used by the platform.

Intelligent analysis features:

At present, the application of intelligent video analysis system in the direction of video surveillance is mainly to identify, classify, and track moving targets. The following rules and functions can be set:

1, squall line detection

For people, cars through the specific direction of movement of the monitoring line; its applications such as: cordon, one-way gate flow, fence climbing ... and so on;

2, alert area

For the monitoring of people and vehicles entering or leaving a specific control area; their application examples: the working area of ​​the parking apron, the terminal station, and the back office of the business site...etc.

3, strolling

For people and vehicles that have been unreasonably stranded for a long time, warning messages are issued to remind security management personnel to check and pay attention.

4. Theft

The preservation of specific important assets; for example: valuable paintings or decorations, equipment and equipment, vehicles or equipment in the station, etc.

5. Remnants

For the potential disposal of explosives, flammables, biochemical contaminants ....

6, group events

Focus on crowds

Comparison of smart monitoring and traditional monitoring

Comparison of traditional alarm system and intelligent analysis alarm performance

System uses network structure

1. Simple mode: The video source comes from the analog monitor head, which is more suitable for small single monitoring system

2. Networking mode: The video comes from the encoded network data. It is suitable for large-scale network monitoring. It has unique advantages in large-scale monitoring systems. It does not need to make any line changes, and it can select any video to be analyzed. The operation is extremely convenient.

Intelligent Video Analysis System Overview

Intelligent Video (IV, Intelligent Video) originates from computer vision (CV, Computer Vision) technology. Computer vision technology is one of the branches of artificial intelligence (AI) research. It can establish a mapping relationship between images and image descriptions, so that the computer can understand the content of video images through digital image processing and analysis.

The intelligent monitoring technology mentioned in video surveillance mainly refers to: “Using intelligent algorithms, automatic content analysis of the input video image, extracting the key, effective information that we are interested in the current monitoring image.” If the camera Think of human eyes, and intelligent video systems or devices can be seen as the human brain. Intelligent video analysis technology uses computer's powerful data processing functions to perform high-speed analysis on massive data in video frames, filtering out information that users do not care about, and only providing useful key information for the monitor. Intelligent video surveillance is based on digital and networked video surveillance, but it is different from general network video surveillance. It is a higher-end video surveillance application.

Intelligent video analysis system is an intelligent video analysis product involving image processing, pattern recognition, artificial intelligence and other fields. It can analyze the abnormal conditions such as warning zone entry, left or lost items, retrograde, abnormal population density in the video area, and can automatically identify the target type and track the moving target appearing in the video area. Mark and draw the target trajectory, and send alarm information in time. Can simultaneously monitor multiple targets in the same scene, can be flexibly set according to the characteristics of the defense target; it can adapt to different environmental changes, including lighting, four seasons, day and night, sunny rain, etc., and can be a good anti-camera jitter. It changed the status of previous “passive” surveillance of video, not only providing video images, but also capable of actively analyzing video information intelligently, identifying and distinguishing objects, and customizable event types. Once abnormal situations or emergencies are detected An alarm is issued in a timely manner and its application in the field of security will inevitably help overcome the limitations of human fatigue and thus more effectively assist security personnel in dealing with emergencies.

The development of intelligent video analytics

Intelligent video surveillance technology is a new video surveillance technology based on image processing and pattern recognition. In short, it is to find the moving objects in the image, track and analyze them, find “abnormal” behavior in a timely manner, trigger an alarm and take other measures to intervene. Video Analytics integrates multidisciplinary research results. Mainly include image processing, tracking technology, pattern recognition, software engineering, digital signal processing (DSP) and other fields. With the increase of computer processing capabilities, in the 1990s, the processing of images has gradually become a research hotspot. Among them, Carnegie Mellon (CMU) University completed the intelligent image monitoring system on campus in 1999, which is a representative research project. At that time, smart image monitoring technology was still in the main stage of laboratory research.

After the 9/11 attacks in 2001, the United States greatly strengthened its investment in security research. Many research institutions and researchers have joined the research and development of security technology. Intelligent video analysis is one of the highlights. Judging from the number of research papers, there was a clear peak period from 2002 to 2005. This is consistent with the large investment in research funding during this period. At present, research papers in this area of ​​research gradually shift to the problems and directions of subdivision. This does not mean that intelligent video surveillance has become a problem that has been solved. On the contrary, even the best business systems today are far from people’s expectations for such technologies. There is no consensus on how to solve the problem. It actually reflects the reduction in the theoretical work of originality. The progress of this technology may rely more on the company’s own scientific research and development forces in the future.

Intelligent Video Analysis System Features

Currently, intelligent video analysis systems on the market generally have the following functions:

1. Image acquisition/interface

2. Moving object detection

3. Multi-object tracking

4. Behavioral characteristics analysis

5. Set alarm conditions

6. Alarm linkage

1, image acquisition / interface

The vast majority of intelligent video analysis algorithms are based on non-compressed image formats such as RGB or YUV, so image signals are sent directly to the video analysis unit without being compressed. Almost all video analysis systems come with image capture capabilities. , The analog image signal is usually input through BNC.

The image signal in the existing image monitoring system is usually in the form of a compressed image stream, such as MPEG4, H.264, MJPEG, and the like. IP cameras usually also output the compressed image stream directly. Motion detection directly from the compressed image stream is a biased research direction and has not yet been accepted by mainstream manufacturers. Of course, the image stream can also be decompressed and restored to the original image format before analysis. Common compression formats are not lossless compression. Compared to the original image before compression, the decompressed image loses some information. However, due to the characteristics of the compression algorithm, the lost information is usually a high-frequency noise signal, so the impact on motion detection is small. Of course, the premise is that the compressed stream has enough bandwidth. If the compression ratio is too high, the image will have a "mosaic" effect, making video analysis more difficult.

Due to the very high processor requirements for real-time image processing, the resolution of an image for video analysis is usually less than that for display or transmission only. The size of the resolution affects the detection distance and sensitivity to moving objects. Some products adapt to the processor's processing power by reducing the frame rate of processing. Too low a frame rate can affect the reliability of the tracking algorithm. It may cause misjudgment of the nature of motion of the moving object.

2, moving object detection

In simple terms, motion detection is the discovery of moving objects in an image. Moving objects can be simply defined as changing parts of the image. Some primary motion detection algorithms are based on these concepts, such as the motion detection features of earlier DVR products. They usually do not have tracking capabilities. The false alarm rate of such methods is too high to be suitable for real-time alarm systems.

Not all changes in the image are moving objects that we are interested in. For example, changes introduced by the camera itself include pixel noise, overall brightness changes caused by the camera's automatic iris control circuit, high- and low-frequency periodic noise signals introduced in image transmission, and mutations caused by infrared camera cycle calibration. Changes introduced by the external environment include rapid changes in the ground light in cloudy weather, shadows of moving objects, waves or sparkling phenomena on the surface, swinging of branches on land, halo caused by headlights at night, and rain and snow conditions. . In addition, the camera easily shakes on a windy day, especially a high light pole. The image changes caused by these phenomena above should usually be filtered out. They can be solved by algorithms or other technical means.

From the perspective of the algorithm, it can be simply divided into two categories. One is to establish a background model and find the moving object by comparing it with the background model. The other is through the “optical flow” method to discover moving objects by finding the effect of moving objects on the optical flow field. The other is a method that combines the two or a combination of the two. The background model method is more complete in the extraction of moving objects, which is beneficial to the tracking, classification and future retrieval of objects in the next step. But it requires the camera to be fixed. If there is no effective stabilization algorithm, false alarms are easily generated in the case of camera shake. The optical flow method requires low camera stability. Even if the camera is mounted on a pan/tilt head, or on other sports platforms such as airplanes, motion detection can be performed. However, since the optical flow method is based on the detection method of the derivative. It is more susceptible to image noise. So it is not suitable for detecting small objects and the detection distance is short.

3, multi-object tracking

The biggest difference between existing video analysis algorithms and early motion detection is whether to track moving objects. Moving object detection and tracking is the basis of video analysis. If these two aspects are made solid, it is possible to analyze the behavioral characteristics of objects, and it is also possible to quickly develop new functional modules for specific applications. Tracing is essentially stringing together the same objects found on each frame in chronological order. This area itself is a relatively independent and active research area. The main research direction is to effectively track under complex circumstances such as multiple moving objects, multiple cameras, moving objects blocking each other, disappearing and recurring. For example, tracking a person at a crowded subway platform; tracking a soldier in the grass wearing a camouflage and crawling in a certain direction. Although the soldier's position could not be identified with the naked eye in each frame, the system discovered him after accumulating a certain number of frames. The above examples mainly remain in the laboratory demonstration stage. However, they represent the direction of the development of tracking algorithms.

In practical monitoring applications, especially for some intrusion alarm application cases, the requirements for tracking algorithms are relatively low. Existing commercial systems do not have satisfactory results in tracking "convergence" of moving objects and other complex application scenarios. However, with reference to the pace of technological development in the past, this aspect will soon be perfected.

4, behavioral characteristics analysis

The behavioral feature analysis is to look for an event that meets a predetermined behavioral characteristic from the image. Typical applications on the market today include:

(1) Classification: Judging that a moving object is a person, a vehicle, a ship, or an airplane.

(2) Stopping or suddenly accelerating: For example, the vehicle is anchored in a tunnel or on a highway; scenes such as robbery in the street and escape.

(3) 徘徊: For example, people observed outside the sensitive area. Pedestrians and vehicles that pass normally are not alerted.

(4) Remainings: For example, places left at airports, oil depots, etc.

(5) Loss of items: For example, protection of valuable exhibits of museums. When the exhibits disappear, the system will immediately call the police.

(6) Population statistics: For example, statistics on the number of people entering supermarkets and other places. In combination with sales data, the average consumption curve for the day is plotted.

(7) Population density: For example, when there are too many people gathered, they will call the police. Or the crowd suddenly disperses, and if there is an abnormal situation, an alarm is issued.

(8) Persons fall to the ground: For example, when a person suddenly changes from standing upright to lying down.

In general, intelligent video analytics can do many things. So video analysts and end users need to communicate effectively. Since intelligent video analysis is still a relatively new technology, the circle that knows this technology in the country currently only expands to the integrator level. Therefore, many application scenarios suitable for video analysis technology have yet to be developed in the market. But one thing is clear: companies must master core technologies and have independent research and development capabilities. The market of intelligent video analysis is composed of many subdivided small markets, and new applications are constantly emerging. In the foreseeable future, this will be an obvious feature of this market.

5, set the alarm conditions

The introduction of "smart" in video surveillance has greatly enriched the monitoring content and improved the monitoring flexibility. Users can alert on a specific behavior. For example, simply speaking, an alarm occurs when a moving object crosses a certain boundary. More restrictive conditions can also be used, such as alerting personnel entering a zone between 7 pm and 7 am, and not alerting outbound personnel and incoming and outgoing vehicles. Since the alarm conditions are set via software, changing the alarm strategy is usually very easy. For example, there are a group of valuable goods stored in a warehouse for only one day. Virtual boundaries can be set around the screen warehouses on that day. Just as Sun Wukong draws a circle with gold hoop bars to protect them, it quickly establishes protective measures. . Users can also set different strategies based on the specific security requirements of different facilities. For example, the monitoring content during the day and night is different. The monitoring content and monitoring intensity are also different on weekdays and weekends. The system automatically switches to avoid the arbitrariness of personnel monitoring. The currently available alarm elements include areas, time periods, object types, dimensions, direction of movement, speed, behavioral characteristics, and much more.

6, alarm linkage

After discovering anomalies in the intelligent video analytics system, there are usually three things that need to be done:

1) Verification of the authenticity of the alarm: A detailed investigation of the close of the alarm event by another PTZ camera. Due to the need of monitoring range, the alarm camera usually has a large monitoring range and is often a fixed camera. Another PTZ camera can cooperate with one or more fixed alarm cameras to respond to alarm events automatically or manually. Alarm video clips are usually stored on the hard disk at the same time.

2). Timely notification and reminding of monitoring personnel: Commonly used real-time prompting methods include voice prompting monitoring personnel, such as “discover people at the warehouse doorway”; popping up an alarm image on the screen; marking the alarming object with an identification box on the image; displaying the alarm The trajectory of the object before it. There are also non-real-time technical means, such as the notification of the responsible person in the form of emails or text messages, and a screenshot of the touch alarm. Currently, real-time monitoring of images through mobile phones has become more mature. The popularity of 3G mobile phones in China will surely make mobile phones become a responsive platform. In addition to DVR backups, intelligent video analytics systems can store alarm video clips for quick retrieval by surveillance personnel.

3). Triggering other external response means: To avoid frequent alarms, some systems can be linked to the horn to alert intruders that they have been monitored. Usually they will leave quickly after they are aware of the discovery. The main form and characteristics of intelligent video analysis system

Current intelligent video analysis products are mainly based on general-purpose CPUs such as Intel (servers, industrial computers) or DSPs. Some products are integrated with DVRs, and some products are made into independent modules, which provide an interface and development SDK for integrators to use. The most integrated product has been integrated with the camera and outputs the intelligent analysis results directly.

Server-based (IPC)-based systems are usually suitable for deployment in the back-end of surveillance systems. Because its architecture is relatively open, it can be easily integrated with existing monitoring systems. In addition, the server's CPU processing power is higher than that of the DSP, and more sophisticated algorithms can be used. Multi-core is the development direction of Intel CPU, which is very suitable for the needs and development trend of multi-channel image processing. It helps to reduce system costs. Intel is also launching a new product every two years faster than Texas Instruments. Server-based system performance can be easily upgraded with Intel product updates. In some high-end intelligent video surveillance systems, more servers are used.