WO2005109893A2

WO2005109893A2 - System and method for detecting anomalies in a video image sequence

Info

Publication number: WO2005109893A2
Application number: PCT/US2005/014860
Authority: WO
Inventors: Robert Pless
Original assignee: Washington University
Priority date: 2004-04-30
Filing date: 2005-04-29
Publication date: 2005-11-17
Also published as: WO2005109893A3

Abstract

A system and method for detecting anomalies in a video. The system comprises a digital image input to receive a plurality of frames. Each frame has a plurality of pixels each corresponding to a pixel location in each of the frames. The system uses a background model based on spatio-temporal filter responses at each pixel location in the frame and an anomalous motion detector to generate a score for each pixel based on the background model at the pixel location. The score is compared with a threshold score indicative of a maximum score for the pixel location to fit the background model. An anomalous motion signaling system triggers an anomalous motion signal when the comparison of the score for at least one pixel with the threshold score indicates that the spatio-temporal filter responses at that pixel location differ sufficiently from the background model.

Description

System and Method for Detecting Anomalies in a Video Image Sequence Field of the Invention The present invention relates generally to systems and methods for processing video images and more particularly to identifying motion in video images.

Background [01] Motion detection techniques involving the analysis of video image sequences are well known in the art. Such techniques typically involve converting an analog image signal into a digital stream of data that characterizes video or picture elements commonly known as "pixels." The pixels are typically organized in frames of individual images that, in sequence, generate the video. Pixels are typically characterized by their position in the frame and by the intensity of light generated at the position. In color images, the digital signal also includes the intensities of the three color signals (red, green and blue) at each pixel.

[02] Motion detection techniques have found advantageous use in, for example, video surveillance systems that analyze video data collected from surveillance cameras and trigger an alarm when motion is detected. A variety of video surveillance systems and motion detection techniques have been developed. For example, one technique tracks the number of pixels that have changed from one frame to the next and responsively changes the rate of video capture as the number increases. In another technique, each new image in a video sequence is compared with either a reference frame or a previous video frame. A motion detection signal is generated if the overall difference between the new frame and the old frame exceeds some threshold. In response to the signal some action is taken. For instance, the current frame may be recorded, or an alarm may alert security personal to the motion. Improvements on this type of system have focused on refining ways to create a reference frame and on methods for varying the detection thresholds which are used to trigger an alarm.

[03] One problem with known motion detection techniques is that they tend to generate false alarms. False alarms may be due to the false interpretation of changes in pixel intensities as motion. Various systems have been developed in attempts to minimize the occurrence of such false alarms. In one system, new images are compared to a reference frame. The differences are then analyzed to determine whether changes are due to lighting variation or motion in the scene. Another system uses the retinex theory of human vision to try to highlight changes that come from sources other than lighting changes.

[04] While known systems have been quite useful in detecting motion, none have adequately addressed the problem of background motion. Background motion includes motion such as trees moving in the wind or waves in water. The detection of background motion is not typically useful in a surveillance system. One known system that has addressed the problem of background motion provides a user interface that allows a user to define areas of a region with significant background motion (e.g. trees moving with the wind, waves in water and other background motion). In these regions, the system does not respond to any independent motion that it detects.

[05] However in such systems, motion of interest (i.e. anomalous motion) may occur in the same region in which background motion occurs. Such motion would not be identified. In addition, it may not be possible to isolate regions of background motion. For example, a surveillance camera may be aimed at a body of water. Video from such a camera would contain background motion covering entire images due to the waves of water throughout the images. Finally, it would be preferable to automate the technique so as not to require that a user identify any regions of background motion.

[06] There exists a need for a motion detection technique that accounts for background motion and identifies anomalous motion.

Summary of the Invention [07] In view of the above, embodiments of the present invention provide a system for detecting anomalies in a video. The system comprises a digital image input to receive a plurality of frames. Each of the frames has a plurality of pixels and each pixel corresponds to a pixel location in each of the frames. The system has a background model based on spatio-temporal filter responses at each pixel location in the frame. The background model may be updated and stored in memory as each frame is processed. [08] The system also comprises an anomalous motion detector that generates a score for each pixel based on the background model at the pixel location. The anomalous motion detector compares the score to a threshold score indicative of a maximum score for the pixel location to fit the background model.

[09] An anomalous motion signaling system triggers an anomalous motion signal when the score for at least one pixel with the threshold score indicates that the spatio-temporal filter responses at that pixel location differ sufficiently from the background model.

[10] Embodiments of the system may also employ a blurring function to generate a blurred image comprising a blurred intensity level at each pixel location of each frame. A frame processor is also employed to store a set of pixel intensities for each pixel location. The pixel intensities comprise, for each pixel location, a first blurred intensity level, lo, at the pixel location; a second blurred intensity level, l_x, at an adjacent pixel location to the right of the pixel location; a third blurred intensity level, l_y, at another adjacent pixel location above the pixel location; and a fourth blurred intensity level, l_t, from the pixel location in the previous frame.

[11] Embodiments of the system may also have a pixel filter operable to generate filtered pixel data at each pixel location using the set of pixel intensities at that location. The filtered pixel data comprises, the first intensity lo; an x-direction intensity change Δl_x = l₀ - l_x; a y-direction intensity change Δl_y = lo - l_y; and a temporal intensity change Δl_t = lo - It-

[12] Embodiments of the system may also employ a scoring function to receive the filtered pixel data and to generate the score based on the background model.

[ 3] Embodiments of the system may also include a threshold database comprising a plurality of threshold scores. Each threshold score corresponds to one of the pixel locations on the frame. A threshold comparator function may receive the score generated by the scoring function for a given pixel location and retrieve the threshold score from the threshold database corresponding to the given pixel location. The system may also include a background model parameter database, which comprises at least one background model parameter; and a background model update function operable to receive the filtered pixel data and to update the at least one background model parameter in accordance with the filtered pixel data.

[14] The anomalous motion signaling system may trigger the anomalous motion signal when the scores of a plurality of pixels located within a predetermined region indicates that the spatio-temporal filter responses at the pixel locations within the predetermined region differ sufficiently from the background model.

[15] In another aspect of the present invention, embodiments of the system may use a plurality of background models each based on spatio-temporal filter responses at each pixel location in the frame. The anomalous motion detector may then generate a plurality of scores for each pixel, where each score is based on one of the background models at the pixel location. The anomalous motion detector may compare each score to one of a plurality of threshold scores, each corresponding to one of the background models and each indicative of a maximum score for the pixel location to fit the background model.

[16] In another aspect of the present invention, embodiments of the present invention provide a video system that include a video camera to capture a video image; an image digitizer operable to convert the video image to a digital signal comprising a plurality of frames each having a plurality of pixels each pixel corresponding to a pixel location in each of the frames. The video system also employs a background model based on spatio-temporal filter responses at each pixel location in the frame in such a manner that the background model would be updated as each frame is processed. An anomalous motion detector is also included to generate a score for each pixel based on the background model at the pixel location. The anomalous motion detector compares the score to a threshold score indicative of a maximum score for the pixel location to fit the background model. An anomalous motion signaling system triggers an alarm when the comparison of the score for at least one pixel with the threshold score indicates that the spatio-temporal filter responses at that pixel location differ sufficiently from the background model. [17] In another aspect of the present invention, embodiments of the present invention include a method for detecting anomalous motion in a video system. The method includes receiving a digital image as a series of frames comprising a plurality of pixels each at a pixel location; generating a set of filtered pixel values for each pixel in each frame; generating a score at each pixel location using the filtered pixel values and a background model at each pixel location; comparing the score at each pixel location with a threshold score for each pixel location; and generating an anomalous motion signal if the score is greater than the threshold.

[18] The above-mentioned and other features, utilities, and advantages of the invention will become apparent from the following detailed description of the preferred embodiments of the invention together with the accompanying drawings.

[19] Other systems, methods, features, and advantages of the invention will become apparent to one having skill in the art upon examination of the following figures and detailed description. It is intended that all such additional systems, methods, features, and advantages be included within this description, be within the scope of the invention, and be protected by the accompanying drawings.

Brief Description of the Drawings [20] Figure 1 is a block diagram representation of a video system operable to detect anomalous motion in accordance with exemplary embodiments;

[21] Figure 2 is a block diagram of an exemplary system for detecting anomalous motion that may be used in the system shown in Figure 1 ;

[22] Figure 3 is a block diagram of a first exemplary embodiment of a background model function that may be used in the system shown in Figure 1 ;

[23] Figure 4 is a block diagram of a second exemplary embodiment of a background model function that may be used in the system shown in Figure 1.

Detailed Description [24] Figure 1 is a block diagram representation of a video system 5 operable to detect anomalous motion in accordance with one exemplary embodiment. The video system 5 comprises a video camera 10 coupled to a video signal analyzer 20. The video system 5 may also include a display 40 for presenting the images from the camera in a form perceptible by humans. The video system 5 operates in a computer system having one or more central processing units, mass-storage memory, local storage memory and an input/output system. The video signal analyzer 20 receives image data at an image input 15 and processes the image data in accordance with applications and functions designated for the video system 5. The video system 5 may have the video camera 10 in a single unit, or different computers may perform different functions described below in a more distributed system. Specific details of the computer system are implementation details that one of ordinary skill in the art would readily understand. In preferred embodiments, the video system 5 is used in security systems in which the video camera 10 collects images from an area of interest to detect whether anyone or anything has obtained access to the area. The video signal analyzer 20 receives the image data signal and analyzes the data to determine whether any part of the image is an entity that does not belong in the image, i.e. an article of interest.

[25] An article of interest appears as a moving object in an area of interest. The image that the camera 10 captures when the area of interest does not have an article of interest present is a background image. In prior art systems, the background image is presumed to be primarily a still image so that any motion detected in the background image would be detected as an article of interest. In reality, however, the background image would not be completely still. For example, a camera aimed at a lake with waves, or a field with trees that move with the wind would pick up images with motion in the background, i.e. background motion. In exemplary embodiments, the video signal analyzer 20 advantageously distinguishes between such motion and the motion indicative of an article of interest, i.e. anomalous motion.

[26] The video signal analyzer 20 in Figure 1 receives the image data at the image input 15 and processes the image data using an anomalous motion detector 22. The anomalous motion detector 22 analyzes the image over time to create a background model based on spatio-temporal filter responses. Spatio-temporal filter responses are numbers that are calculated for each pixel based on the relative intensity of that pixel, and its neighbors in the same image, and its neighbors in time, ie, the intensity value at the same pixel in the preceding image. The background model is data, functions, parameters, preferably organized in data structures in memory, and which are used for identifying regions of video which have unusual spatio-temporal filter responses. Exemplary embodiments provide for updating the background model or allowing the background model to update itself to accommodate to slow changes in the background that appear in the background image.

[27] The image data in a digital form comprises a plurality of frames received as a series of frames. Each frame comprises a plurality of picture elements, or pixels, that correspond to a location, or pixel location, in each of the frames. The pixel location is the location of a point on the image. The anomalous motion detector 22 advantageously determines typical filter responses to form the background model.

[28] As image data is received, the filter responses are compared with the background model to detect pixel locations at which the filter responses deviate from the expected filter responses. The anomalous motion detector 22 analyzes each pixel and may determine that anomalous motion is detected based on the analysis of a single pixel. Alternatively, the anomalous motion detector 22 may determine that anomalous motion is detected based only on the analysis of pixels that form an area on the image.

[29] When pixel locations are deemed to indicate anomalous motion, an anomalous motion signaling system 30 triggers an anomalous motion signal 35. In an exemplary embodiment, the anomalous motion signal 35 is used to provide alarms or other notice to draw human attention to the anomalous motion. This may be done in a variety of ways, including without limitation by audible and/or visual alarm 60, by triggering a video recorder 50 to capture the image of the anomalous motion, or by any suitable means in accordance with specific applications.

[30] In exemplary embodiments, the anomalous motion detector 22 analyzes the image data to detect anomalous motion by first converting the image data to a digital representation, if the image data is not already in digital form. If the camera 10 is a digital camera, or if another component that is not shown intervenes to convert the data to a digital form, the video signal analyzer 20 need not convert the image data to a digital form. The anomalous motion detector 22 then creates and maintains a background model based on pixel data at each pixel location in the frame. The background model is maintained by being updated as each frame is processed. The anomalous motion detector 22 analyzes each pixel in the image data and generates a score for each pixel based on the background model at each pixel location. The anomalous motion detector 22 compares the score to a threshold score indicative of a maximum score for the pixel location to be within a region of background motion. The anomalous motion signaling system triggers the anomalous motion signal 35 when the comparison of the score for at least one pixel with the threshold score indicates that the pixel filter responses do not fit the background model at this location.

[31] The video system 5 in Figure 1 may be used in a variety of applications. For example, the video system 5 may be a security system in which the camera 10 is aimed at an entrance or the surroundings of a property or other area that is to be kept secure from intruders. The background image formed by the camera may include background motion such as traffic, foliage moving with the wind, waves on the water, or other types of common background motion. The video system 5 in Figure 1 advantageously distinguishes between the background motion and the anomalous motion indicative of intruders.

1 Anomalous Motion Detector [32] Figure 2 shows a block diagram of an exemplary anomalous motion detector that may be used in the system shown in Figure 1. The anomalous motion detector in Figure 2 includes an image preprocessing system 100, a background model processor 140 and an anomaly processor 150. In the description that follows, the image is presumed to be monochromatic. Each pixel is therefore analyzed according to a single intensity level at each location. The present invention is not limited however to monochromatic images. One of ordinary skill in the art will appreciate that color images may also be processed in accordance with embodiments of the present invention by analyzing multiple intensities, such as, for example, the three components of color (i.e. red, green and blue, or Y, U, V). [33] The image preprocessing system 100 comprises a blurring function 105 operable to generate a blurred image. The blurring function 105 may be a 5x5 blurring filter that is either uniform or Gaussian. One of ordinary skill in the art will appreciate that a number of possible blurring filters may also be used. In the blurred image produced by the blurring function 105, each pixel value is a weighted sum of neighboring pixel values. The blurred image is sent to an image buffer 160, which maintains the current blurred image and the blurred image from the previous frame.

[34] A frame processor 110 processes a frame of image data pixel by pixel. The frame processor 110 retrieves the pixel data from the current blurred image and the previous blurred image in the image buffer 160 and stores a set of pixel intensities 170 for each pixel location. The pixel intensities 170 include, for each pixel location, a first blurred intensity level lo, which represents the blurred intensity level at the pixel location being processed. In addition, the pixel intensities 170 include a second blurred intensity level l_x at an adjacent pixel location to the right of the pixel location, a third blurred intensity level l_y at another adjacent pixel location above (i.e. "above" as viewed on the image display) the pixel location and a fourth blurred intensity level It from the pixel location in the previous frame. The pixel intensities 170 are input into a pixel filter 120, which generates a set of filtered pixel data 180. As shown in Figure 2, the filtered pixel data 180 comprises the first intensity lo, an x-direction intensity change Δl_x = l₀ - l_x, a y- direction intensity change Δl_y = l₀ - l_y, and a temporal intensity change Δl_t = lo - It- One of ordinary skill in the art will appreciate that different filters may be used to calculate filter responses I, Δl_x, Δl_y, Δl_t. In addition, I, ΔI_Xι Δl_y, Δl_t. are four exemplary filter responses, and that other filter responses may also be used. For example, filter responses can include color intensity parameters, i.e. R, G, B, l_x, l_y, It.

[35] Once the filtered pixel data 180 is derived, it is input to the background model processor 140. The background model processor 140 maintains a set of background parameters that permit the background model processor to characterize normal filter responses of the background image. The background model processor 140 also maintains a threshold score for each pixel location. The threshold score indicates a threshold intensity level beyond which the intensity at the pixel location is not at an expected level indicative of the background. The background model processor 140 uses the filtered pixel data 180 and the background parameters to generate a score at the pixel location. The score is compared to the threshold score to determine whether the pixel location contains a pixel that reflects an unexpected intensity. An unusual set of filter responses generates an anomaly indicator 190, which in some embodiments, may be sufficient to signal anomalous motion. The anomaly processor 150 may trigger an anomalous motion signal on a single pixel. Or, alternatively, the anomaly processor 150 may collect multiple anomalous indicators and determine if they comprise an area of anomalous intensities such that an entity is determined to be present in the background area.

2 Background Model Processor [36] Figure 3 is a block diagram of one exemplary embodiment of a background model processor that may be used in the system shown in Figure 2. The background model processor in Figure 3 comprises a scoring function 200, a threshold comparator function 220, a background model update function 240, a background model parameter set 210 and a threshold data set 260. The background model processor may comprise one or more scoring functions 200. Figure 3 depicts operation of a background model processor that uses a single scoring function 200.

[37] The scoring function 200 may implement any suitable statistical function that may be used to generate a score for an filter responses at each pixel based on the historical filter responses of that pixel. The scoring function 200 retrieves one or more parameters based on the selected function for each pixel from the background model parameters database 210. The scoring function 200 then uses the parameters to compute a single value, a score, for each pixel. Each score is based upon the parameters for the pixel and the four filter responses at that pixel location, i.e. the filtered pixel data.

[38] The score is passed to the threshold comparator function 220, which retrieves a threshold score, which may be different for each pixel, from the threshold database 260. The threshold comparator function 220 compares the score with the threshold score for the pixel location. If the score is smaller than the threshold score, the parameters defining the background model are updated (as described below), and the background model processing for this pixel is complete. If this score is greater than the threshold, then the threshold comparator function 220 signals an anomaly indicator. Processing would then continue for the next pixel location.

[39] The background model update function 240 advantageously updates the background model parameter values in accordance with the background model selected (i.e. selected statistical function). The process of updating the background model parameters ensures that each pixel filter response set is being analyzed against historical data for that pixel intensity at that pixel location. In this way, the background model update function 240 incorporates background motion into the comparison thereby increasing the likelihood that only anomalous motion will be detected as being motion..

[40] As described above, the scoring function 200 may be any suitable statistical function. In one exemplary embodiment, the scoring function is the "Mahalinobis Distance". Those of ordinary skill in the art will appreciate that the Mahalinobis distance is one method for computing how likely a point is to come from a distribution, and that there are other methods for representing a probability distribution, and for determining how likely a new sample is to come from that motion. For this function, there are fifteen parameters to be stored in the background model parameters DB 210. These parameters are: u, ux, uy, ut, and M, Mx, My, Mt, Mxx, Mxy, Mxt, Myy, Myt, Mtt, and n. The values are used for the mean vector (V = (u, ux, ut, ut)) and the covariance matrix M (where M, Mx, My, Mt, Mxx, Mxy, Mxt, Myy, Myt, Mtt, are the values, in order, of the right upper triangular part of a symmetric matrix) and n is the number of data vectors that have already been included in the model. If the four values comprising the filtered pixel data at a pixel are defined to be a vector of the data at the pixel called D, then the Mahalinobis distance is defined by the matrix multiplication rule: (V - D)' M^"1 (V - D). This distance is a single number, and is defined as the score.

[41] The background model update function 240 updates the background model by updating the parameters listed above. The updated parameters, after updating with the data D are defined to be: V = V n / (n+1 ) + D / (n+1 ), which is the weighted average of V and D. The new matrix M is defined as: M = M n/(n+1 ) + D D' n / (n+1 f . And the value of n is increased: n = n+1. [42] The Mahalinobis distance is only one example of a function that can be used as the scoring function 200. Another example is a function known to those of ordinary skill in the art as the constant optic flow function. For this function there are 7 background model parameters, Mxx, Mxy, Mxt, Myy, Myt, Mtt, and n. The values Mxx, Mxy, Mxt, Myy, Myt, Mtt are the values listed, in order, of a right upper triangular part of a matrix M (in this case a 3 by 3 symmetric matrix), and n is the number of measurements so far included in the model. The scoring function is computed by defining to values u,v as follows:

• v = (Mxx Myt - Mxt Mxy)/ (Mxy Mxy - Mxx Mxx)

• u = ( -Myt - V Myy ) / Mxy,

which, those of ordinary skill in the art may recognize as the pseudo-inverse solution to a linear equation with the matrix inverse of a 2 by 2 matrix written out explicitly. Once these values are computed, the score for a particular pixel is (u lx + v ly + It)². In order to update this model, the new values of M are computed as follows:

Mxx = Mxx n / (n+1 ) + lx lx / (n + 1 )

Mxy = Mxy n / (n+1 ) + lx ly / (n + 1 ) Mxt = Mxt n / (n+1) + lx lt / (n + 1)

Myy = Myy n / (n+1 ) + ly ly / (n + 1 )

Myt = Myt n / (n+1) + ly lt / (n + 1)

Mtt = Mtt n / (n+1) + It It / (n + 1) n = n + 1.

[43] The above-described examples of scoring functions are only two examples of statistical functions that may be used in exemplary embodiments. For example, the scoring function 200 may implement a known intensity function in which a known intensity value defines the background. Another may be based on a constant intensity or a constant intensity and variance. Still other functions include the use of Gaussian distributions and Multiple Gaussian Distributions in (l₀, l_x, l_y, l_t) space and Linear Prediction based on History.

[44] Figure 4 is a block diagram of a second exemplary embodiment of a background model function that may be used in the system shown in Figure 1. The background model function in Figure 4 uses multiple scoring functions to allow scoring using a variety of background models. The function in Figure 4 inputs the filtered pixel data 180 into a first scoring function 320. The first scoring function 320 calculates a score in a manner similar to the functions described above with reference to Figure 3. A first threshold function 330 retrieves a threshold score for the particular pixel position. The threshold score may be particular to the background model used by the scoring function 320. The score is compared with the threshold score at 332a. If the score is less than the threshold score, the score is passed to a Min function 400. The background model parameters for the scoring function 320 are updated and processing for the pixel location is complete.

[45] If the score is greater than the threshold, the score is still passed to the MIN function 400. However, the comparator 332a signals an input filtered pixel data function 300 to pass the filtered pixel data to a second scoring function 340. The second scoring function calculates a second score and a second threshold function 350 passes the score to a comparator 332b. The second score is compared to the threshold score, which may be the same as the first threshold score, or may be different in accordance with the particular background model used by the second scoring function 340.

[46] If the second score is less than the threshold score, the second score is passed to the MIN function 400. The background model parameters for the scoring function 340 are updated and processing continues for the next pixel.

[47] If the second score is greater than the threshold score, the score is passed to the MIN function 400 and the comparator 332b signals the input filtered pixel data function 310 to input the filtered pixel data 180 into the next scoring function. As shown in Figure 4, the next scoring function is the last scoring function 370. The system may use any number of scoring functions deemed appropriate for a specific application.

[48] Referring back to Figure 4, the last scoring function 370 determines a last score and a last threshold function 390 retrieves the threshold score. The last score is compared with the threshold at 332c. Regardless of the outcome, the score is passed to the MIN function 400. The MIN function 400 outputs a final score that may or may not indicate anomalous motion. The MIN function 400 advantageously allows a quality control function to be added in exemplary embodiments to monitor the background models. For example, the background model that more commonly generates the lowest score is typically providing the most accurate model of the background.

[49] The background model processor shown in Figure 4 advantageously uses multiple scoring function to enhance the accuracy of the detection of anomalous motion. In addition, the proper selection of scoring functions may make the process more efficient. For example, the first scoring function 320 may be a simple known intensity for a particular pixel location. Or, the first scoring function 320 may be a constant intensity and variance function. These two functions are very simple and may permit early indication of a lack of anomalous motion sufficient to forego the need to proceed with additional scoring functions. The second scoring function 340 may be more complex, the Mahalinobis Distance, for example. A third may be still more complicated such that the more complex, computationally intensive, but perhaps more complex functions would only need to be used where the scores are very close to the threshold. This would allow the system to process from one pixel location to the next more quickly.

[50] Those of ordinary skill in the art will appreciate that the embodiments described above may be implemented in color image systems with appropriate modifications. The modifications relate to the fact that three intensities (for each component of color) are used instead of one in the embodiments described above. That is, with color video pixels, the measurement taken at each pixel would include 6 components (instead of 4). The six components include the 3 components of the color (for example, red, green, and blue, or, if the camera uses another representation of color, Y, U, V), and additionally the three derivative measurements (l_x l_y l_t) listed above. In this case, the matrix M will be a 6 x 6 matrix and the vector u will be a 6 element vector.

[51] In addition, embodiments of the present invention are not limited to any particular system configuration. Figure 1 shows a camera that is separate from the video signal analyzer, which is separate from the display and video recorder. All depicted components may be in a single standalone system, or a part of a larger video system. The video signal analyzer may also receive a compressed signal, such as an MP4 signal. While a compressed signal may be used, the video signal analyzer preferably decompresses the video signal for better performance.

Claims

claim:

1. A system for detecting anomalies in a video comprising: a digital image input to receive a plurality of frames each having a plurality of pixels each pixel corresponding to a pixel location in each of the frames; a background model based on spatio-temporal filter responses at each pixel location in the frame, the background model being stored in memory and updated as each frame is processed; an anomalous motion detector operable to generate a score for each pixel based on the background model at the pixel location, to compare the score to a threshold score indicative of a maximum score for the pixel location to fit the background model; and an anomalous motion signaling system operable to trigger an anomalous motion signal when the comparison of the score for at least one pixel with the threshold score indicates that the spatio-temporal filter responses at that pixel location differ sufficiently from the background model.

2. The system of claim 1 further comprising: a blurring function operable to generate a blurred image comprising a blurred intensity level at each pixel location of each frame; and a frame processor operable to store a set of pixel intensities for each pixel location, the pixel intensities comprising for each pixel location a first blurred intensity level lo at the pixel location, a second blurred intensity level l_x at an adjacent pixel location to the right of the pixel location, a third blurred intensity level l_y at another adjacent pixel location above the pixel location and a fourth blurred intensity level It from the pixel location in the previous frame.

3. The system of claim 2 further comprising: a pixel filter operable to generate filtered pixel data at each pixel location using the set of pixel intensities, the filtered pixel data comprising: the first intensity l₀; an x-direction intensity change Δl_x = lo - l_x; a y-direction intensity change Δl_y = lo - l_y; and a temporal intensity change Δl_t = lo - It-

4. The system of claim 3 further comprising a scoring function operable to receive the filtered pixel data and to generate the score based on the background model.

5. The system of claim 4 further comprising a threshold database comprising a plurality of threshold scores, each threshold score corresponding to one of the pixel locations on the frame.

6. The system of claim 5 further comprising a threshold comparator function to receive the score generated by the scoring function for a given pixel location and to retrieve the threshold score from the threshold database corresponding to the given pixel location.

7. The system of claim 4 further comprising: a background model parameter database comprising at least one background model parameter; and a background model update function operable to receive the filtered pixel data and to update the at least one background model parameter in accordance with the filtered pixel data.

8. The system of claim 1 wherein the anomalous motion signaling system triggers the anomalous motion signal when the scores of a plurality of pixels located within a predetermined region indicates that the spatio-temporal filter responses at the pixel locations within the predetermined region differ sufficiently from the background model.

9. A system for detecting anomalies in a video comprising: a digital image input to receive a plurality of frames each having a plurality of pixels each pixel corresponding to a pixel location in each of the frames; a plurality of background models stored in memory, each background model based on spatio-temporal filter responses at each pixel location in the frame; an anomalous motion detector operable to generate a plurality of scores for each pixel, each score based on one of the background models at the pixel location, the anomalous motion detector operable to compare each score to one of a plurality of threshold scores, each corresponding to one of the background models and each indicative of a maximum score for the pixel location to fit the background model; and an anomalous motion signaling system operable to trigger an alarm when the comparison of one of the plurality of scores with the threshold score for at least one pixel indicates that the spatio-temporal filter responses at the pixel location differ sufficiently from the background model.

10. The system of claim 9 further comprising: a blurring function operable to generate a blurred image comprising a blurred intensity level at each pixel location of each frame; and a frame processor operable to store a set of pixel intensities for each pixel location, the pixel intensities comprising for each pixel location a first blurred intensity level lo at the pixel location, a second blurred intensity level l_x at an adjacent pixel location to the right of the pixel location, a third blurred intensity level l_y at another adjacent pixel location above the pixel location and a fourth blurred intensity level It from the pixel location in the previous frame. 1. The system of claim 10 further comprising: a pixel filter operable to generate filtered pixel data at each pixel location using the set of pixel intensities, the filtered pixel data comprising: the first intensity l₀; an x-direction intensity change Δl_x = lo - l_x; a y-direction intensity change Δl_y = lo - l_y; and a temporal intensity change Δlt = lo - It.

12. The system of claim 11 further comprising a plurality of threshold databases corresponding to the background models, each threshold database comprising a plurality of threshold scores, each threshold score corresponding to one of the pixel locations on the frame.

13. The system of claim 12 further comprising a plurality of threshold comparator functions each coupled to a corresponding scoring function, the threshold comparator functions operable to receive the score generated by the corresponding scoring function for a given pixel location and to retrieve the threshold score from the threshold database corresponding to the given pixel location and to the background model.

14. The system of claim 9 further comprising: a plurality of background model parameter databases comprising at least one background model parameter, each background model parameter database corresponding to one of the background models; and a plurality of background model update functions, each corresponding to one of the background models, the background model update functions operable to receive the filtered pixel data and to update the at least one background model parameter in accordance with the filtered pixel data.

15. A video system comprising: a video camera operable to capture a video image; an image digitizer operable to convert the video image to a digital signal comprising a plurality of frames each having a plurality of pixels each pixel corresponding to a pixel location in each of the frames; a background model based on spatio-temporal filter responses at each pixel location in the frame, the background model being stored in memory and updated as each frame is processed; an anomalous motion detector operable to generate a score for each pixel based on the background model at the pixel location, to compare the score to a threshold score indicative of a maximum score for the pixel location to fit the background model; and an anomalous motion signaling system operable to trigger an alarm when the comparison of the score for at least one pixel with the threshold score indicates that the spatio-temporal filter responses at that pixel location differ sufficiently from the background model.

16. The video system of claim 15 further comprising: a blurring function operable to generate a blurred image comprising a blurred intensity level at each pixel location of each frame; and a frame processor operable to store a set of pixel intensities for each pixel location, the pixel intensities comprising for each pixel location a first blurred intensity level lo at the pixel location, a second blurred intensity level l_x at an adjacent pixel location to the right of the pixel location, a third blurred intensity level l_y at another adjacent pixel location above the pixel location and a fourth blurred intensity level l_t from the pixel location in the previous frame.

17. The video system of claim 16 further comprising: a pixel filter operable to generate filtered pixel data at each pixel location using the set of pixel intensities, the filtered pixel data comprising: the first blurred intensity lo', an x-direction intensity change Δl_x = lo - l_x; a y-direction intensity change Δl_y = lo - l_y; and a temporal intensity change Δlt = lo - It-

18. The video system of claim 17 further comprising a scoring function operable to receive the filtered pixel data and to generate the score based on the background model.

19. The video system of claim 18 further comprising a threshold database comprising a plurality of threshold scores, each threshold score corresponding to one of the pixel locations on the frame.

20. The video system of claim 19 further comprising a threshold comparator function to receive the score generated by the scoring function for a given pixel location and to retrieve the threshold score from the threshold database corresponding to the given pixel location.

21. The video system of claim 18 further comprising: a background model parameter database comprising at least one background model parameter; and a background model update function operable to receive the filtered pixel data and to update the at least one background model parameter in accordance with the filtered pixel data.

22. The video system of claim 15 wherein the anomalous motion signaling system triggers the anomalous motion signal when the scores of a plurality of pixels located within a predetermined region indicates that the spatio-temporal filter responses at the pixel locations within the predetermined region differ sufficiently from the background model.

23. A method for detecting anomalous motion in a video system comprising: receiving a digital image as a series of frames comprising a plurality of pixels each at a pixel location; generating a set of filtered pixel values for each pixel in each frame; generating a score at each pixel location using the filtered pixel values and a background model stored in memory for each pixel location; comparing the score at each pixel location with a threshold score for each pixel location; and generating an anomalous motion signal if the score is greater than the threshold.

24. The method of claim 23 further comprising: updating the background model at each pixel location using the filtered pixel data if the score is not greater than the threshold.

25. The method of claim 24 wherein the step of generating the set of filtered pixel values comprises: generating a blurred image comprising a blurred intensity level at each pixel location of each frame; storing a set of blurred pixel intensities for each pixel location, the pixel intensities comprising for each pixel location a first blurred intensity level lo at the pixel location, a second blurred intensity level l_x at an adjacent pixel location to the right of the pixel location, a third blurred intensity level l_y at another adjacent pixel location above the pixel location and a fourth blurred intensity level l_t from the pixel location in the previous frame; and generating each item of the filtered pixel data to comprise: the first blurred intensity lo; an x-direction intensity change Δl_x = lo - l_x; a y-direction intensity change Δl_y = lo - l_y; and a temporal intensity change Δlt = lo - It-

26. A system for detecting anomalous motion in a video system comprising: means for receiving a digital image as a series of frames comprising a plurality of pixels each at a pixel location; means for generating a set of filtered pixel values for each pixel in each frame; means for generating a score at each pixel location using the filtered pixel values and a background model at each pixel, location; means for comparing the score at each pixel location with a threshold score for each pixel location; and means for generating an anomalous motion signal if the score is greater than the threshold.

27. The system of claim 26 further comprising: means for updating the background model at each pixel location using the filtered pixel data if the score is not greater than the threshold.

28. The method of claim 27 wherein the means for generating the set of filtered pixel values comprises: means for generating a blurred image comprising a blurred intensity level at each pixel location of each frame; means for storing a set of blurred pixel intensities for each pixel location, the pixel intensities comprising for each pixel location a first blurred intensity level lo at the pixel location, a second blurred intensity level l_x at an adjacent pixel location to the right of the pixel location, a third blurred intensity level l_y at another adjacent pixel location above the pixel location and a fourth blurred intensity level I_t from the pixel location in the previous frame; and means for generating each item of the filtered pixel data to comprise: the first blurred intensity lo; an x-direction intensity change Δl_x = lo - l_x; a y-direction intensity change Δl_y = lo - l_y; and a temporal intensity change Δl_t = lo - It.