The invention relates to a security system as generically defined by the preamble to claim 1 and to a method for operating the security system as generically defined by the preamble to claim 4. Such security systems are often equipped with stationary cameras. Detecting movement or change with stationary cameras is a basic function of systems for radio-based security technology. The products range from surveillance cameras that issue alarms to digital video recorders which allow a content-based search for moving objects. Detecting moving objects is also a basic function in analyzing image sequences and is thus an, important component for instance of systems for man-machine interaction (such as gesture control) or biometric systems (for instance, face detection with ensuing face recognition).
Both the systems described in the scientific literature and those on the market for detecting moving objects implicitly or explicitly use a camera sensor model which assumes that the time-related noise in a pixel (“pixel noise”) is independent of the gray value. Such systems are described for instance in the following places in the literature:
A. Elgammal, D. Harwood, L. Davis, “Non-Parametric Model for Background Subtraction”, FRAME-RATE workshop, 1999.
K. Toyama, J. Krumm, B. Brumitt and B. Meyers, “Wallflower: Principles and Practice of Background Maintenance”, ICCV 1999.
A. Elgammal, R. Duraiswami, D. Harwood, L. Davis, “Background and Foreground Modeling Using Nonparametric Kernel Density Estimation for Video Surveillance”, Proc. of the IEEE, Vol. 90, No. 7, July 2002, pp. 1151-1163.
M. Meyer, M. Hotter, T. Ohmacht, “A New System for Video-Based Detection of Moving Objects and its Integration into Digital Networks”, in Proceedings of IEEE Intern. Conference on Security Technology, Lexington, USA, 1996, pp. 105-110.
T. Aach, A. Kaup, R. Mester, “Change Detection in Image Sequences using Gibbs Random Fields: A Bayesian Approach”, Proceedings Intern. Workshop on Intelligent Signal Processing and Communication Systems, Sendai, Japan, October 1993, pp. 56-61.
The assumption of gray value independent of the pixel noise in the prior art is clearly incorrect, especially in the widely used sensors employing CCD technology. Instead, in reality, an increase in the noise variance of a pixel with the corresponding gray value must be expected. The usual simplifying assumption in the industry of pixel noise independent of the gray value has an adverse effect on the performance of the entire security system. For instance, in conventional security systems this assumption means that there must be a fixed decision threshold relating to the gray value, if a distinction is to be made between a gray value change because of sensor noise and a gray value change because a moving object has been detected. However, since the noise behavior in most image sensors is gray value-dependent, this means that the aforementioned decision threshold set is too sensitive for bright image regions and too insensitive for dark image regions.
The security system of the invention having the characteristics of claim 1, conversely, leads to a substantial improvement over conventional security systems. Because the decision threshold is designed to be gray value-dependent, the security system can be better adapted to both bright and dark image regions. This leads to substantially more-enhanced sensitivity of the security system. Because a gray value-dependent noise behavior is taken into account in defining the decision threshold, it is now possible even to detect dark objects in dark image regions, without generating mistaken detections caused by pixel noise in bright image regions. Advantageously, the detection precision is thus increased without causing an increase in the rate of mistaken detections. The lowest possible rate of mistaken detections, however, is of especially great significance in security technology.
The invention is described in further detail below in conjunction with the drawings.
The assumption of gray value independent of the pixel noise in the prior art is clearly incorrect, especially in the widely used sensors employing CCD technology. Instead, in reality, an increase in the noise variance of a pixel with the corresponding gray value must be expected. The usual simplifying assumption in the industry of pixel noise independent of the gray value has an adverse effect on the performance of the entire security system. For instance, in conventional security systems this assumption means that a fixed decision threshold relating to the gray value is understood if a distinction is to be made between a gray value change because of sensor noise and a gray value change because a moving object has been detected. However, since the noise behavior in most image sensors is gray value-dependent, this means that the aforementioned decision threshold set is too sensitive for bright image regions and too insensitive for dark image regions. This situation is illustrated in
If the noise variance were gray value-independent, then
An optimal decision threshold would be gray value-dependent and would correspond in its qualitative course to the course of the curve marked “noise variance over the gray value”; that is, for dark image regions, the threshold would be lower than for bright pixels. In the case of a sensor with a linear course of this curve (see also
The security system of the invention having the characteristics of claim 1 conversely leads to a substantial improvement over conventional security systems. The invention is based on the recognition that substantially better results can be attained if the decision threshold is adapted adaptively. Because the decision threshold is now designed to be gray value-dependent, the security system can be better adapted to both bright and dark image regions. This leads to substantially more-enhanced sensitivity of the security system. Because a gray value-dependent noise behavior is taken into account in defining the decision threshold, it is now possible even to detect dark objects in dark image regions, without generating mistaken detections caused by pixel noise in bright image regions. Advantageously, the detection precision is thus increased without causing an increase in the rate of mistaken detections. The lowest possible rate of mistaken detections, however, is of especially great significance in security technology.
One exemplary embodiment of the security system 100 according to the invention and its operating phases will be described below, in conjunction with
The security system 100 includes at least one camera 3 with an image sensor 4, and this camera is associated with both subsystems 101, 102 and is active in both operating phases of the security system 100. The security system 100 also includes a plurality of function modules 1, 6, 8, 9, 15, which are linked in terms of circuitry or at least functionally to the camera 3. The subsystem 101, besides the camera 3, includes a function module 1 with a light source. The brightness of this light source is controllable as a function of time. The subsystem 101 further includes a function module 6 for displaying a digital image sequence from the output signals of the image sensor of the camera 3. Finally, the subsystem 101 includes a function module 8 for displaying the noise variance as a function of the gray value from the digital image sequence. The subsystem 102, besides the camera 3 with the image sensor 4, includes a function module 13, which in turn comprises two function modules 13a, 13b. The function module 13a serves to calculate or estimate the gray value variance from the output signals of the image sensor 4 of the camera 3. The function module 13b makes a comparison with a threshold value possible. The security system 100 further includes a memory 9, to which both subsystems have access.
In this security system 100, two operating phases can be distinguished, which will now be discussed in succession. In the first operating phase, initialization of the security system 100 is done in the offline mode (flow chart in
The second operating phase of the security system 100 is schematically shown in the subregion 102 of the drawing. The system operates in ongoing operation as follows. A natural scene (recording field 10) that corresponds to the area being monitored is examined in terms of the scene contents for whether a change in pixels of the images taken by the camera 3 is occurring because of sensor noise, or because of a moving object. Once the natural scene 10 has been recorded (step 60 in
In a further step (step 64 in
As described above, it is useful during the initializing phase for data about the operating state of the sensor 4 as well as camera parameters, such as the amplification, to be forwarded to the function module 8 that determines the noise curve, so that possible changes in the gray-value-dependent noise characteristic can be taken into account. For instance, it is possible that the amplifier noise of the image sensor 4, in low light conditions, will cover the noise in a picture element and thus change the gray-value-dependent of the noise. To make it possible to utilize this option even during ongoing operation, the operating state of the image sensor 4 must be forwarded to the function module 13b for the threshold value decision.
The essential nucleus of the invention is thus the use of an adaptive, gray-value-dependent threshold value decision for detecting objects. By this provision, the performance and precision of recognition by such a security system is enhanced substantially. The threshold values are expediently measured in advance in the form of characteristic curves as a function of the gray value and of the camera parameters and are stored in a memory 9.
Number | Date | Country | Kind |
---|---|---|---|
10 2004 018 410.0 | Apr 2004 | DE | national |
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/EP05/50991 | 3/4/2005 | WO | 7/31/2006 |