Video motion detectors

Please, note that the article is automatically translated from Russian into English, so the translation may not be accurate.

Nikolai Ptitsyn

The article deals with current problems of motion video detection (video analysis of movement) in security surveillance systems. Compares the server, embedded and hybrid algorithms implementation. The features-Video High Definition (HD). Listed algorithmic and engineering approaches to reduce false positives while maintaining high sensitivity detector. Practical recommendations on the choice-Video.

Problems and problem statement

VMD is a hardware / software, part of the surveillance system and provides automatic detection of moving objects in a streaming video. Thus, when a person or vehicle in a field of view camera transmits a signal-Video (metadata) to the control of the device or video (watch video).

In practice, the technical means should not only detect moving objects, but to ensure their continued support (tracing) in the field of view of one or even multiple cameras. This function is needed to reduce the frequency of false positives, and especially repeated, as well as to visualize the trajectory of the on-screen operator.

Video analyzer motion greatly increases the productivity of staff safety, reduce the psychological burden and reduce the reaction time. Metadata generated by the detector can be used as an index for quick search for events in the video archive. In this case there is no need for a grueling multi-day view the video operator.

Good detector can significantly increase the efficiency of storage and bandwidth by reducing redundant videomaterilov. In the perimeter security systems saving disk space and network traffic in the tens or thousands of times depending on the busiest controlled areas.

-Video can be implemented programmatically in the server security system or integrated directly into the camera. In the latter case, the camera manufacturer uses a digital signal processor (DSP) or a custom chip (ASIC) for processing streaming video in real time. Many of today's ip-cameras have a embedded motion detection.According to Frederick Nielson, CEO of Axis, in the coming years-Video will be an important driving force behind market-circuit television cameras.

In the sterile (laboratory) conditions, the problem of motion video analysis is straightforward and from the standpoint of software video processing algorithms, and in terms of computational load on the processor. In the field, the problem takes on a fundamentally different level of complexity and goes to the class of non-trivial task of computer vision and pattern recognition on which to continue working scientists and engineers. 
The main problems are not solved by the developers before the end, are as follows:

• elimination of false positives, which are due to natural changes in the external environment (eg, solar glare, shadows, wiggling bushes, trees, water), as well as objects, not representing the interest of security service (eg, animals, birds, insects on the camera's lens, the clouds );
• good sensitivity for detecting objects in unstable stage of observation (eg, significant changes in natural or artificial lighting, as well as the variability of the background, such as open water, forest, driven by strong winds);
• ensure simplicity and ease of use-Video (three-dimensional gauge the prospects, the job classes of detected objects, filtering false positives).

The biggest challenge is to find a compromise while also addressing the three above-mentioned problems. On the one hand-Video should have high sensitivity to the objects of the unpredictable forms. For example, it must effectively detect people in camouflage suit, a group of people of any size, vehicles of any kind. On the other hand, the detector should not respond to the "harmless" to the external environment, such as swaying trees, and running the shadows of clouds.

Solving these problems requires a high level of artificial intelligence, complexity, can detect motion in the field of view, but also to provide targets of interest of a security guard. Thus, the detector has to classify moving objects on the current and irrelevant in all possible manifestations of the world. The false alarm rate should not exceed the threshold value at which the application-Video is becoming uneconomical. Permissible period between false alarms depends on the usage scenario and may vary from several hours to several months.

HD-Video

More keenly discussed issues evident in the video analysis systems, high-definition (HD), that is, with more than one megapixel sensors. The main differences in the HD-Video compared to systems standard definition (SD) are:

• The greater detail of the scene HD leads to frequent false alarms. In particular, our own camera movement caused by wind and vibration equipment, leading to significant changes in the image.Application of digital and mechanical image stabilizer becomes mandatory.
• Megapixel camera has a longer range or radius of the review. The range of scales of objects can be much greater than in the systems of SD. In the systems of HD should be applied in principle other algorithms promising calibration and optical correction. Be sure to use multiscale algorithms for simulation of the background and the segmentation of objects.
• The data flow in the system HD is many times the data stream SD. Most of the algorithms used in the Intelligent Video has nonlinear complexity with respect to the frame size and processor load increases by several orders of magnitude. Thus, it is necessary to optimize the fundamental and often create new algorithms for streaming HD.

Architecture and integration issues

Server-Video

The classical approach consists in the software implementation-Video on the server that handles streaming video coming from analog or ip-cameras. As the hardware platform is commonly used x86-compatible computers with specialized video capture cards and hardware accelerators compression. Such server-Video are implemented in domestic systems, "Intelligence" of the company «iTV», «Trace" company «DSSL», etc.

The main advantages of a server, a centralized approach is simple and flexible integration of video analytics. A significant drawback is that it handles the signal-Video, distorted analog section or compression algorithm.Another disadvantage - a significant burden on the communication channels and poor scalability of the server.In the case of an overload the server starts to skip frames, which leads to missed events and false positives.These failures are compounded significantly when switching to standard HD.

Built-Video

An alternative approach is to use the built videodektorov. In this case, processing streaming video directly at the camera before compression and transmission in the Situation Centre. Popular hardware platforms are SoC DaVinci company Texas Instruments, Blackfin company Analog Devices, ETRAX company Axis, Nexperia product line, NXP and development of Stretch. It is also possible to use the architecture x86, for example, a processor Intel Atom. Built-Video providers include such companies as IOImage, ObjectVideo, Cernium in the West, Synesis in the CIS and other

It just would like to refute the popular misconception that the server-Video on x86 have a higher accuracy than its single-chip counterparts. On the contrary, the theoretical accuracy of the built-Video above server due to its higher quality input image and computing power dedicated processor, all resources are loaded processing only one channel.

Video analysis of data in systems with embedded detectors significantly decentralized, which provides good scalability up to hundreds of thousands of cameras. In this case, the load on the channel of communication and storage can be minimized by filtering and selective transfer footage directly into the camera, which is especially important in wireless networks.

The disadvantages are the complexity of embedded videodektorov integration due to lack of common standards and the higher cost of hardware. However, the initial investment is quickly recouped if the accuracy of the embedded detector is made of high quality.

Hybrid-Video

There are also hybrid solutions that combine the server and embedded approaches. Thus, the company Agent Vi has implemented distributed processing: built the algorithm selects the local motion vector in the frame, while the server is processing algorithm for a higher level. However, such technology is poorly applicable to scenes with a variable background and create a significant burden on the communication channel during the transition to standard HD.

The accuracy of video analysis

Formal evaluation of the accuracy of detection is a challenging organizational and engineering task because of the infinite variety of scenarios, the behavior of offenders and the manifestations of the environment. Specialists have traditionally considered an error of two kinds: (1) missed the violation and (2) false alarm. It is obvious that missed the violation is more unpleasant mistake than a false alarm. The high frequency of false positives can lead to a decrease in the effectiveness of security guards and eventually neutralize the benefit-Video. Allowable ratio of errors of the first and second class defines the technical requirements of the project.

Research unit of the Ministry of Internal Affairs of Great Britain has created a collection of video footage i-LIDS (Imagery library for intelligent detection systems) and formal methods of testing detectors. British scientists have applied the metric of F _1, representing an integrated security system accuracy. The parameter α determines the effect of errors of the first and second kind integral metric F _1. Thus, for the services of a rapid response parameter α takes the value 0.65, to log events video - the value of 75.00.

The detectors have an average quality within F _1, about 0.98, good detectors - more than 0.99.

There are many additional parameters and conditions that affect the estimation accuracy. Note the most important:

• maximum response time since the advent of object in the field of view camera, with which he is deemed to have the detected;
• method of accounting for the repeated signals of the same object (if the object is not of interest to a security guard, then repeat the signals will reduce the productivity of employees);
• stabilization time after power-Video, or a significant change in the scene.

Tab below the technical properties that, according to the author's research, are needed to achieve high precision and versatility-Video.

Realization of high precision-Video: technical features and their benefits

Conclusion

Stable detection of movement of targets, especially outdoors, is an actual scientific and engineering problem at the junction of disciplines such as machine vision, pattern recognition and digital signal processing.

In terms of architecture and embedded implementing server-Video, available today, they have certain advantages and disadvantages. It is safe to say that the embedded detectors, as it develops, forcing the server.

In terms of recognition efficiency video embedded detectors can be identical or even better server. Adequate embedded implementation of the algorithm provides better accuracy and responsiveness than the server implementation of the same algorithm. As shown in the article, such an advantage built analytics due to the lack of compression distortion and decentralized processing large volumes of video data.

When choosing a detector should pay attention to a set of scripts and video stories for which testing and optimization algorithms. Solutions available on the market for security to be significantly different frequency of false positives, sensitivity, robustness, diversity, environment and ease of configuration. At the same time achieving an effective work-Video HD is the task of fundamentally more complicated than the standard definition of the detector.

It should be noted that today, many manufacturers of cameras and video management systems offer free-Video, as they can not guarantee the stable operation of the detector in the field.