Ðóññêèé   Óêðà¿íñüêà
DonNTU   Masters portal

Abstract

When writing this essay master's work is not yet completed. Final Completion: June 2019. Full text of the work and materials on the topic can be obtained from the author or from his scientific adviser after the specified date.

Introduction

Recently, the term “video surveillance” has firmly entered our lexicon and has become an integral part of the modern security system. Intelligent video surveillance systems are installed in virtually every large company, but public places, crowded places — metro, shopping malls, public parks, hospitals — are still not equipped for the most part with similar security systems.

If we talk about intelligent video surveillance systems, then it is worth understanding that this is usually a hardware and software complex used for automated information gathering from streaming video. In their work, these systems rely on various algorithms for image recognition, systematization and processing of the obtained data. Varieties of hardware and software systems of intelligent video surveillance systems and their characteristics are shown in Table 1.1.


Table 1.1. Types of hardware-software systems regarding the method of processing information in them
Type Characteristic
Server type Analytical data processing is carried out centrally on a video server or PC. The hardware component is a CPU or GPU. The main advantage of the server system of intelligent video surveillance in the software used, which allows you to add additional modules and video processing algorithms to the existing shell, as well as to combine existing ones. The main disadvantage is the need for constant transmission of high-resolution video from cameras to a video server, which significantly loads communication channels.
Built-in intelligent algorithms Used directly in security cameras. A partially or fully processed image with analysis results (metadata) is transmitted to the DVR or server. This method significantly reduces (10-100 times) the load on information transmission channels. However, video cameras have a limited set of analytical functions, and their cost considerably exceeds conventional devices.
Distributed video processing The primary analysis of information that does not require complex algorithms can be performed on video cameras. For example, the detection of an object. And more serious intellectual processing, requiring CPU load, is performed using the server capacity.

1. Analytical problem statement

1.1. Problem and relevance

The trend of introducing intelligent video surveillance systems is noticeable both in the CIS countries and abroad. This is largely due to the problem of terrorism, acute in the twenty-first century, and the increase in crime rates in major cities of the world. Manifestations of terrorism entail massive human casualties, the destruction of spiritual, material and cultural values; the threat of terrorism has increased the urgency of maintaining security in crowded places around the world.

Mass terrorist acts have led to the need to create a unified security system in crowded places. It includes: the information system, CCTV, the system of inspection and physical security, the center of control and management of all technological processes.

1.2. The object and purpose of the study. Main tasks

The economic and organizational effect, as well as the increased level of security from the introduction of intelligent video surveillance systems, is well noticeable not only in large networks with a wide territorial distribution, but also in small systems. Examples of the use of intelligent video surveillance systems are given in Table 1.2.

Table 1.2 The use of intelligent video surveillance systems
Scope of application Opportunities
Transportation tasks
  • license plate recognition;
  • automatic passenger counting;
  • detection of the left object in the forbidden zone;
  • definition of a foreign object on the rails.
Urban Security Systems
  • face recognition for wanted criminals;
  • detection of fights and other illegal actions;
  • identification of suspicious activity sites.
Objects of the closed or mode type
  • perimeter control;
  • duplication of fire alarm system functions (visual detection of the source of ignition in the early stages);
  • control of the staff.
Catering organizations, trade and banking institutions, car washes, service stations, hairdressing, etc.
  • automatic counting and classification of customers;
  • analysis of the composition and length of the queue.

Depending on the purpose of using a video surveillance system, intelligent video signal processing can perform one or several functions, such as: object detection, tracking, classification and statistical analysis, recognition, detection of alarm situations.

Recently, intelligent video surveillance systems are widely used and analytical functions - forecasting, intelligent additional video compression, event ranking, selective deletion/editing of private data or blocking the recording of private zones.

The subject of the research is video information from surveillance cameras in crowded places. The goal is to increase the effectiveness of security system management by increasing the probability of preventing illegal actions. The task of designing such a subsystem involves the development of the following subtasks:

  1. Getting video from surveillance cameras.
  2. Pre-processing video.
  3. Recognition and classification of objects.
  4. Analysis and alert staff.
  5. Intellectual compression.

Pre-processing the video sequence involves creating conditions that increase the efficiency and quality of the selection and recognition of the desired objects. The preprocessing methods depend on the research tasks, are quite diverse and can include, for example, highlighting the most informative fragments, increasing them, increasing the contrast resolution, improving the quality of images, etc.

Recognition and classification of images is the task of identifying an object or determining any of its properties from its image.

Designing a subsystem involves the implementation of a situational analysis, that is, one that will focus on identifying alarming situations. When identifying an alarming situation, the main criteria for detection are the intersection of the control line, an abrupt change in the position of an object in space (fall, jump, etc.), the emergence of a source of fire, and others.

1.3. Mathematical apparatus

The task of pattern recognition can be viewed as the task of establishing differences between the original data, and not by identifying with individual images, but with their sets.

The objective function F in this case is a combination of factors — the error factor Erri between expert judgment and machine analysis and the compression ratio Kj, as shown in the formula.

F = Erri → min AND Kj → min

A good image quality in compression is considered to be the achievement of the coefficient K = 0.15 - 0.35. K is calculated by the formula.

K = size * 8 / width * height * fps * time

size - video clip size, byte

width - frame width, pixel

height - frame height, pixel

fps - frame rate per second

time - the duration of the compressible fragment, sec

The efficiency criterion Err is calculated using the formulas below and characterizes the difference between expert estimates and machine analysis. It should be assumed that expert analysis is a standard and takes the value Accexp = 1, while the accuracy of machine expertise varies within Accsys ∈ (0; 1].

ACCexp = TPexp + TNexp / TPexp + FPexp + TNexp + FNexp = 1
ACCsys = TPsys + TNsys / TPsys + FPsys + TNsys + FNsys ∈ (0; 1]
Err = ACCexp / ACCsys ∈ [1; ∞]

ACCexp, ACCsys - the value of the accuracy of expert evaluation and machine analysis, respectively,

TPexp, FPexp, TNexp, FNexp, TPsys, FPsys, TNsys, FNsys – true positive, false positive, true negative, false negative statement regarding peer review and machine analysis, respectively.

References

  1. S.N. Yaryshev, “TV security systems”. [link].
  2. S.N. Yaryshev, “Digital methods of recording and playing video”. [link].
  3. S. Orlov, “Video analytics tasks and solutions”., Network Solutions Log/LAN ¹6 2014. [link].
  4. Y. Bengio, “Learning deep architectures for AI”. [link].
  5. Kaiser L. [è äð.], “One model to learn them all”. [link].
  6. B. Alexe, T. Deselaers, and V. Ferrari. Measuring the objectness of image windows. TPAMI, 2012. [link].
  7. P. Arbelaez, B. Hariharan, C. Gu, S. Gupta, L. Bourdev, and J. Malik. Semantic segmentation using regions and parts. In CVPR, 2012.
  8. P. Arbelaez, J. Pont-Tuset, J. Barron, F. Marques, and J. Malik. Multiscale combinatorial grouping. In CVPR, 2014.
  9. J. Carreira, R. Caseiro, J. Batista, and C. Sminchisescu. Semantic segmentation with second-order pooling. In ECCV, 2012.
  10. J. Carreira and C. Sminchisescu. CPMC: Automatic object segmentation using constrained parametric min-cuts. TPAMI, 2012.