Mobile QR Code QR CODE

  1. (Enterprise Department, Innodep Inc./ Seoul City, 08375, Korea
  2. (Information Security, Graduate School of Korea University / Seoul City, 02841, Korea
  3. (Information Security, Graduate School of Korea University / Seoul City, 02841, Korea

PTZ control, Random exclusive, Border defense, Intelligent surveillance system, Deep learning

1. Introduction

South Korea’s continuing low birth rate is reducing the number of people with military obligations, which is expected to reduce to 110,000 troops by 2022. In order to prepare for this, defense boundary tasks are being automated using information technology. The task is to develop a system that detects enemy soldiers using cameras at a defense boundary and analyzing captured images by an analysis engine with deep learning technology is applied [1]. All enemies including camouflaged ones would be detected by teaching the engine using pictures.

However, the system needs an image of sufficient resolution to identify an enemy through an image, and the area must be enlarged to acquire such an image. As a result, the camera cannot see all the areas for which it is responsible at the same time. Thus, all regions time should be monitored in parts through PTZ(Pan Tilt Zoom) control. In this paper, we propose a method for effective PTZ control for monitoring with a time-division method. We compare the proposed method with a previous method by simulating scenarios and measuring the detection probability and detection time.

2. Related Work

Most of the research on PTZ control uses a camera for wide view and PTZ cameras for zooming in on and tracking objects [2-5]. Some studies such as [2,3] handle a situation where wide-range camera overviews all the responsible regions, and the PTZ cameras view a sub-region for tracking. Another study [4] approaches the problem in a similar manner for face recognition and feature extraction through PTZ control. Another study [5] also adopts a wide-view camera and PTZ cameras. The former is used for tracking multiple targets, and the latter is used for tracking a selected target. If the target goes out of the PTZ camera’s boundary, it ``hands over'' to another PTZ camera.

Another proposal [6] adopts multiple PTZ cameras for 3D visualization, which is helpful for monitoring agents with situational awareness. In another study [6], algorithms were presented for 3D location determination using multiple cameras and algorithms for detecting and tracking. In another study [7], a rule-based method was adopted for PTZ scheduling. It utilizes reasoning modules to determine the PTZ values, and the controlled situation is learned iteratively.

As presented above, the majority of research adopts a camera for a whole view and other PTZ cameras for tracking, zooming, and another specific purpose of not losing sight of events in a responsible area. Among all those presented, our aim is the most similar to one proposal [6] in that we also would like to develop a system that can help operators to acknowledge a situation in an instant. The situational difference from the aforementioned research is that our system monitors a vast area of national border, which is about 2 kilometers away from the location of a camera. Although there is a wide-view camera in our system, it just gives an overview. We assume it is not possible to do detection with imagery from the wide-range camera because the resolution of enemies would be below several pixels.

3. Methodology & Simulation

In this paper, we propose three PTZ control methods for monitoring a given region and focus on the performance of each method. We simulated each method and measured the detection probability and average detection time. Fig. 1 shows the camera’s surveillance areas divided into 30 zones. In this environment, the camera looks at one of the divided zones. To divide a zone, the following constraints need to be considered.

1. The minimum detectable resolution of the video analytics engine is 50 pixels horizontally and vertically.

2. The maximum resolution that can be analyzed is 720p.

Considering this, the area divided in an actual image is shown in Fig. 2. The width of Fig. 2 is about 160 m, and the area is divided into 30 areas. With this condition, a person who is 1.7m high will be represented with 50 pixels In this situation, the camera controls the PTZ to monitor the enemy and monitors the area assigned to the camera. There are three ways to control PTZ. The first method is generally used for PTZ scheduling and is simply a sequential scanning method. PTZ is controlled by an adjacent cell from the upper left to the lower right.

The second method is to move the monitoring region in a random manner. The same place can be revisited in a short time, although there can be some areas that are not viewed at all. The third method is to move randomly but to traverse all 30 areas in order to see the whole areas completely. In other words, to see the same zone again, it is necessary to traverse the rest of the zone.

When an image is shown to an analysis engine in a simulation, it is assumed that targets are detected 100\% successfully. It is supposed that an enemy comes from the top and moves in the same way as in Fig. 1. It is assumed that the enemy stays for about 10 seconds in one area. If the simulation does not look at the enemy’s area for more than 50 seconds, it is determined that the detection has failed because it passed. The simulation was done 1000 times, and the success rate and average detection time were calculated afterwards.

Fig. 1. Surveillance zones divided into 30 areas.
Fig. 2. Actual CCTV image divided into 30 areas for PTZ control.

4. Results & Discussion

The simulation results show the rate of successful detection and the average time to successful detection depending on the interval time of PTZ control. Fig. 3 shows the detection success rate and the average time for successful detection as the time interval for PTZ control is increased from 2 seconds to 10 seconds. As shown in the figure, the detection success rate decreases as the time for viewing each area becomes longer.

For a detection success rate of more than 50% with the three methods, the PTZ control interval should be 3 seconds or less. The random exclusive method, which was the third method presented in section 2, was the most effective. The random search method, which was the second method in section 2, shows a similar detection success rate. These two random methods show enhanced successful detection times, which were about 33 seconds compared to 39 seconds for the sequential method. On the other hand, the probability drops sharply when the PTZ control interval is over 4 seconds. In that section, the sequential method (a general search method) shows the highest detection success rate.

Fig. 4 shows the detection probability depending on the number of cameras monitoring the same area. For each case, the PTZ control interval was 2 seconds. As can be expected, the detection success rate increases when multiple cameras detect the same given region together, and the detection success rate increases very rapidly when moving in a random fashion. For the 3-camera case in the random exclusive method, the successful detection rate was over 95%, and the detection time was also shortened to only 15 seconds from 34 seconds in the one-camera monitoring case. Less than half the time is needed for detection if 2 more cameras are included in the monitoring job, which means 10 sub-regions are assigned to each camera.

As can be seen from the results, it was found that the random traversal method is more effective than the sequential traversal of the monitoring area. Among the methods, the random exclusive traversal method showed the best performance. Random traversal also makes sense because the movement pattern of the CCTV should not be estimated by enemies. Therefore, this paper proposed two methods for controlling PTZ randomly.

Fig. 5 shows the successful detection rate depending on the PTZ control interval for the random exclusive case when n, which is the number of cameras, varies from 1 to 10. When the camera is controlled randomly, it provides a guide on how much detection probability should be needed for a given surveillance area. For example, if a successful detection rate over 90% is needed, and the number of sub-regions to be monitored is 20, two cameras with a 2-second PTZ control interval should be adopted. Fig. 6 shows the average detection time under the same circumstances.

Fig. 3. Detection probability of CCTV with PTZ control interval.
Fig. 4. Detection probability of CCTV with multiple cameras.
Fig. 5. Detection probability of random exclusive control method with multiple cameras and PTZ intervals.
Fig. 6. Detection time of random exclusive control method with multiple cameras and PTZ intervals.

5. Conclusion

Nowadays, AI technology that supersedes human roles is continuing to be developed. In the field of national border defense, AI systems are also being developed to compensate for the decline in troops resulting from low birth rates, and they will be introduced in the future. In this paper, to operate these AI systems in accordance with the intended purpose, we have proposed a PTZ control method to effectively monitor an area with this system. For detection and recognition with computer analysis, it is essential to secure enough resolution of the target, so it is inevitable for a camera to zoom in on an area, resulting in a blind region that a camera cannot see. Therefore, it should look around by controlling PTZ.

In this context, we simulated a situation of an enemy invasion using the constraint parameters of an intelligent system and calculated the detection probability and the average time for detection depending on the PTZ-control time interval in three possible PTZ control methods. By comparing the successful detection rate and time to detection of the three methods, we confirmed that the random methods detected the targets with higher detection probability and short detection time compared to the normal sequential method. The results of this paper could be referred to when deploying cameras for real defense boundary surveillance. We hope that the results of this paper can contribute to effective defense boundary surveillance.


This work was supported by the ICT R&D program of MSIP/IITP [2017-0-00250, Intelligent Defense Boundary Surveillance Technology Using Collaborative Reinforced Learning of Embedded Edge Camera and Image Analysis].


Dhulekar P. A., et al. , Nov. 2019, Surveillance System for Detection of Suspicious Activities at War Field, in Proc. of IEEE ICACCTDOI
Qureshi F. Z., et al. , Aug. 2009, Planning ahead for PTZ camera assignment and handoff, in Proc. of ICDSCDOI
Qureshi F. Z., et al. , Nov. 2006, Surveillance camera scheduling: a virtual vision approach, in SpringerDOI
Krahnstoever N., et al. , Oct. 2008, Collaborative Real-Time Control of Active Cameras in Large Scale Surveillance Systems, M2SFA2 2008 Workshop on Multi-camera and Multi-modal Sensor FusionURL
Ding C., et al. , Feb. 2012, Collaborative Sensing in a Distributed PTZ Camera Network, IEEE Trans. On Image Processing, Vol. 21, No. issue 7, pp. 3282-3295DOI
Collins R. T., et al. , Oct 2001, Algorithm for cooperative multisensory surveillance, in Proc. of IEEE, Vol. 89, No. 10, pp. 1456-1477DOI
Starzyk W., et al. , Oct 2011, Learning proactive control strategies for PTZ cameras, in Proc. of ICDSC 2011DOI


Taewoo Kim

Taewoo Kim is a research engineer at Innodep Inc., Seoul, Korea. He is the manager of the enterprise department in the Innodep company. He received his B.S. and M.S. degrees in electronic engineering from Korea Aerospace University, Korea in 2012 and 2014, respectively. He has been engaged in several national projects regarding artificial intelligence (AI) and a video management system. His research interests include AI, machine learning, data analysis, and computer system development.

Hyungheon Kim

Hyungheon Kim is a senior researcher at Innodep Inc., Seoul, Korea. He is the team manager of the enterprise department in the Innodep company. He received his B.S. and M.S. degrees in the science of software from Gacheon University in 2008 and 2010. He is in the doctoral program in the Department of Convergence Security at Korea University's Graduate School of Information Security. His research interests include convergence security, video surveillance, and data visualization.

Youngkyun Cha

Youngkyun Cha has been the general manager in the future technology convergence HQ since 2019 and a specially appointed professor in the graduate school of cybersecurity at Korea University since 2017. He was the president and CEO of KT group affiliates from 2015 to 2016. He was the executive director at Samsung group affiliate from 1991 to 2015. His research areas include physical security and convergence security.