JP4701100B2 - Abnormal behavior detection device - Google Patents

Description

  The present invention relates to an abnormal behavior detection device that detects abnormal behavior of a human or an animal.

  In order to deal with social unrest, such as the increase in crime rates, the number of cameras installed for the purpose of monitoring suspicious persons and suspicious vehicles is increasing. As described above, monitoring using a large number of cameras requires a monitoring support technique for efficiently monitoring a monitoring area with limited monitoring personnel resources.

  As such a monitoring support technique, there is a “feature quantity extraction method and apparatus from three-dimensional data” disclosed in Patent Document 1. Here, a method for extracting a feature amount of a moving image called a three-dimensional higher-order local autocorrelation feature is disclosed. And the method of applying this feature quantity to action recognition and gait authentication is shown.

  Non-Patent Document 1 discloses a method for calculating the degree of abnormality of a person's action in a video using a cubic higher-order local autocorrelation feature.

JP 2005-92346 A Takuya Minamisato, Nobuyuki Otsu, “Detection of abnormal motion from multiple human moving images”, IPSJ SIG 2004-CVIM-145, September 11, 2004

  The above prior art has a problem that the degree of abnormal behavior of a person or animal in a video is calculated as a scalar quantity, and it cannot be immediately determined whether or not abnormal behavior has actually occurred.

  The present invention has been made in consideration of the above problems, and calculates the degree of abnormal behavior of a person or animal in a video and accurately determines whether or not abnormal behavior has occurred based on the degree of abnormality. An object of the present invention is to provide a device for detecting abnormal behavior.

  In the abnormal behavior detection apparatus according to the present invention, the degree of abnormality of the monitored video acquired by the video acquisition unit is calculated, and whether or not abnormal behavior has occurred is determined based on the threshold value based on the calculated degree of abnormality.

  According to the present invention, it is possible to accurately determine whether or not an abnormality has occurred based on the degree of abnormality in the video. As a result, when an abnormality occurs, it is possible to immediately respond to the abnormality by issuing a warning or notifying the guard.

  Hereinafter, embodiments of the invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing a functional configuration of an abnormal behavior detecting apparatus according to an embodiment of the present invention. This apparatus includes a video acquisition unit 100, an abnormality degree calculation unit 102, an abnormality determination unit 104, a determination threshold calculation unit 110, and a notification unit 112, and an abnormality is detected based on the monitoring target video acquired by the video acquisition unit 100. Detect behavior. Hereinafter, it demonstrates in order.

  The video acquisition unit 100 is a video shooting device such as a video camera or a video playback device such as a video recorder, and acquires a video to be input to this apparatus. The video shooting device is used when a real-time video currently being shot is input. The video playback device is used when video stored in the past is used as input.

  The abnormality degree calculation unit 102 calculates the abnormality degree of the video acquired by the video acquisition unit 100. Here, the degree of abnormality is a scalar quantity indicating the degree of abnormality of behavior of a moving object such as a person or an animal in the video.

  The abnormality determination unit 104 determines whether or not abnormal behavior has occurred from the abnormality degree calculated by the abnormality degree calculation unit 102, and outputs the result to the determination result 108. When the determination threshold 106 is used as a determination criterion and the degree of abnormality is smaller than the determination threshold 106, it is determined that no abnormal behavior has occurred. Conversely, if the degree of abnormality is equal to or greater than the determination threshold 106, it is determined that abnormal behavior has occurred.

  The determination threshold calculation unit 110 calculates the determination threshold 106 necessary for the determination process of the abnormality determination unit 104. Here, calculation is performed based on the determination result 108 so that the determination accuracy of the abnormality determination unit 104 is the best.

  If the abnormal action has occurred based on the determination result 108, the reporting unit 112 notifies the external device to that effect. The external device that has received the notification can output an alarm to the voice / monitoring screen. Also, facilities such as elevators can be stopped for safety. Furthermore, it is possible to notify a remote place such as a monitoring center or a portable terminal to promote the response.

The degree-of-abnormality calculation unit 102, the abnormality determination unit 104, the determination threshold value calculation unit 110, and the reporting unit 112 are
It can be realized by a CPU, an arithmetic processing unit such as a CPU, or a personal computer. Further, the determination threshold value and the determination result are stored in a storage device such as a semiconductor memory, and are appropriately read out and used for various calculations.

  Next, the flow of processing when an abnormality determination is performed by the abnormal behavior detection apparatus of the present embodiment will be described using the flowchart of FIG.

  In step 200, the processing from step 202 to step 210 is repeated at a predetermined frequency set in advance until the user gives an end instruction.

  In step 202, the video obtained by the video acquisition unit 100 is captured as digital data using the degree-of-abnormality calculation unit 102.

  In step 204, the degree of abnormality of the video acquired in step 202 is calculated using the degree of abnormality calculation unit 102.

  In step 206, the abnormality determination unit 104 is used to determine whether or not abnormal behavior has occurred using the degree of abnormality calculated in step 204.

  In step 208, the determination result of step 206 is evaluated, and if it is determined that abnormal behavior has occurred, step 210 is executed.

  In step 210, the reporting unit 112 is used to notify the external device that an abnormal behavior has occurred.

  Next, the internal configuration of the abnormality degree calculation unit 102 of FIG. 1 will be described in detail using the block diagram of FIG. As described above, the degree-of-abnormality calculation unit 102 calculates the degree of abnormality of the video acquired by the video acquisition unit 100 as a scalar amount, and outputs it to the abnormality determination unit 104. The abnormality degree calculation unit 102 includes a motion extraction unit 300, a feature amount calculation unit 302, a feature amount conversion unit 304, and an abnormality degree evaluation unit 308. Hereinafter, it demonstrates in order.

  The motion extraction unit 300 extracts a moving part from the video acquired by the video acquisition unit 100. This is to remove a stationary part that is not related to the determination of abnormal behavior such as the background. A known image processing method can be used for motion extraction (see Japanese Patent Application Laid-Open No. 2005-92346). For example, a method of simply taking a difference between two frames or a method of taking a difference between frames after performing edge extraction processing can be used. Further, after taking the difference between the frames, a binarization process is applied so that the pixel value takes a value of 0 or 1 in order to remove the influence of noise such as illumination fluctuation.

  The feature amount calculation unit 302 calculates the feature amount of the video generated by the motion extraction unit 300. For the calculation, a known three-dimensional higher-order local autocorrelation feature is used (for example, see JP-A-2005-92346). This is a method of calculating a geometric feature of voxel data composed of three consecutive frames of video as a 251-dimensional feature vector. A method for calculating the feature amount will be described later. Further, an optical flow may be used for calculating the feature amount of the video. The optical flow is a method of calculating a motion between frames as a vector by paying attention to a minute area of an image, and is described in detail in, for example, the book “Digital Image Processing” (CG-ARTS Association) page 243. . All the components of the vector calculated by the optical flow may be combined and used as the feature amount. In addition, a statistic such as an average / variance of vectors may be used as the feature quantity.

The feature amount conversion unit 304 linearly converts the feature amount vector calculated by the feature amount calculation unit 302 using the conversion matrix 306. By this conversion, abnormal behavior components included in the feature vector are extracted. Here, if the feature quantity vector calculated by the feature quantity calculation unit 302 is x, the transformation matrix 306 is A, and the transformed feature quantity vector is x ′, this transformation can be expressed by the following equation (1).

x ′ = Ax (1)
The transformation matrix 306 is a matrix obtained by multivariate analysis such as principal component analysis, and a calculation method will be described later. When a 251-dimensional higher-order local autocorrelation feature is used as the video feature amount, the transformation matrix 306 is a matrix having a size of (252−n) × 251 (n = 1, 2,..., 251). And the feature-value linearly transformed by this matrix becomes a 252-n- dimensional vector.

  The degree of abnormality evaluation unit 308 calculates the degree of abnormality by evaluating the degree of deviation between the new feature amount vector calculated by the feature amount conversion unit 304 and the normal data 310. Then, the calculation result is output to the abnormality determination unit 104. Here, the normal data 310 is a set of feature amounts for normal behavior. A specific method for calculating the degree of abnormality will be described later.

  Next, details of the abnormality degree calculation processing in step 204 of FIG. 2 will be described using the flowchart of FIG.

  In step 400, the motion extraction unit 300 is used to extract a portion with motion from the video acquired by the video acquisition unit 100.

In step 402, the feature amount calculation unit 302 is used to calculate the feature amount of the video generated in step 400. In step 404, the feature amount conversion unit 304 is used to linearly convert the feature amount vector calculated in step 402. A new feature vector.

  In step 406, the degree of abnormality is calculated by evaluating the degree of divergence between the new feature vector calculated in step 404 and the normal data 310 using the degree-of-abnormality evaluation unit 308.

  Next, details of the moving image feature amount calculation processing shown in step 402 of FIG. 4 will be described with reference to FIGS.

  FIG. 5 is a diagram for explaining the input data of the above-described stereoscopic higher-order local autocorrelation feature. The target of feature quantity calculation is a moving image, that is, a time-series continuous frame (image). At least 3 frames are required to calculate the cubic higher-order local autocorrelation features. For example, when a frame 500 with a frame number n is given, three frames, a frame 502 and a frame 504 (corresponding to frame numbers n-1 and n-2, respectively) positioned before the frame 500 are targeted for feature amount calculation. .

  Now, assuming that the resolution of a frame is h pixels in the vertical direction and w pixels in the horizontal direction, a voxel (cube) of h × w × 3 can be configured by combining the three frames. The three-dimensional higher-order local autocorrelation feature is extracted by applying a 3 × 3 × 3 mask pattern 506 to all the voxel elements while sequentially moving them.

  Note that although three consecutive frames are set as processing targets here, arbitrary f frames may be set as processing targets. In this case, h × w × f voxels are to be processed, and an average feature amount of moving images for f frames is calculated.

  FIG. 6 is a diagram illustrating an example of a mask pattern used for calculation of a stereoscopic higher-order local autocorrelation feature. The mask pattern is used to calculate a local correlation feature of a voxel, and is composed of 3 × 3 × 3 voxels.

  Pattern 1 is a pattern for counting up the number of pixels in the central voxel 600 when the voxel data for the input video is sequentially scanned. Similarly, pattern 2 is a pattern for counting up the number of voxels 602 that are 1 in addition to the central voxel 604.

  There are 251 mask patterns in the three-dimensional higher-order local autocorrelation features for a binary image, and the number of cases where each pattern is satisfied is counted to extract the features of the input video as a 251-dimensional feature vector. To do.

  Next, details of the moving image feature quantity calculation processing shown in step 402 of FIG. 4 will be described using the flowchart of FIG.

  In step 700, the feature vector is initialized.

  In step 702, the processing from step 704 to step 708 is repeated for all voxels of the video to be processed. That is, as shown in FIG. 5, all the voxels to be processed are sequentially scanned using the mask pattern 506.

  In step 704, the processing from step 706 to step 708 is repeated for all 251 types of mask patterns shown in FIG.

  In step 706, it is determined whether all the pixels corresponding to the mask pattern to be processed are 1. If the determination result is true, step 708 is executed.

  In step 708, the feature vector component corresponding to the mask pattern to be processed is incremented by one.

  Through this series of processing, a 251-dimensional feature vector based on the three-dimensional higher-order local autocorrelation can be calculated.

  Next, the calculation procedure of the transformation matrix 306 in FIG. 3 will be described using the flowchart in FIG.

  In step 800, the processing from step 400 to step 402 is repeated for one or more normal scene images that have been recorded in advance in a storage device such as a semiconductor memory for learning.

In step 400, as shown in FIG. 4, the motion extraction unit 300 is used to extract a portion with motion from the video generated by the video acquisition unit 100.

In step 402, as shown in FIG. 4, the feature amount calculation unit 302 is used to calculate the feature amount of the video generated in step 400.

  In step 802, principal component analysis is performed on the set of feature quantities for the calculated normal scene. Principal component analysis is one of multivariate analysis methods. By creating a composite variable called a principal component from several variables so as to be uncorrelated with each other, it is possible to summarize information held by a plurality of variables. This principal component analysis is a method often used for the analysis of multivariate data. For example, it is described in detail in the book “Multivariate analysis that can be understood immediately” (Tokyo Book). Omitted. By performing principal component analysis on a set of 251-dimensional feature vectors, the principal components and eigenvalues of 251 are obtained.

  In step 804, a subspace with a low contribution rate to normal behavior is calculated based on the result of the principal component analysis in step 802. A matrix for converting the feature vector into this partial space is referred to as a conversion matrix 306.

  Next, the details of the subspace calculation process shown in step 804 of FIG. 8 will be described with reference to FIG. This figure represents the cumulative contribution rate for each principal component obtained by the principal component analysis shown in step 802. Cumulative contribution rate is an index that shows how much information that the analysis target data originally had with the principal components up to that point can be explained by adding the contribution rate of each principal component in descending order. is there. For example, the cumulative contribution ratio 900 up to the third principal component is 90%, and the first principal component to the third principal component represent 90% of the information amount of the original data. On the other hand, the remaining information amount from the fourth principal component to the 251st principal component represents only 10% of the information amount of the original data.

  From this, it can be said that the partial space composed of the first principal component to the third principal component has a large contribution rate to normal behavior. Moreover, it can be said that the partial space comprised from the 4th main component to the 251st main component has a small contribution rate to normal action.

  In this way, a partial space with a small contribution rate to normal behavior can be obtained using the cumulative contribution rate as a criterion.

  Next, with reference to FIG. 10, a method of the abnormality degree evaluation process shown in step 406 of FIG. 4 will be described. The degree of abnormality is evaluated in a partial space with a small contribution rate to normal behavior shown in FIG. This is because in this partial space, the variance of the feature amount of normal behavior is small, whereas the variance is large in the case of behavior other than normal, that is, abnormal behavior.

Now, as this partial space, a space composed of the nth principal component and the subsequent components is considered. Originally, this subspace has 252 to n dimensions, but FIG. 10 shows two axes of the n- th principal component and the ( n + 1) -th principal component, which are principal components having a large contribution rate, for convenience. The feature quantity set 1000 is a plot of normal data 310. In the subspace where the contribution rate to normal behavior is small, the set is distributed around the center of gravity xn1004 of the set. Therefore, it can be determined that the feature vector x1002 for the image currently being evaluated is normal when the feature vector x1002 is close to the feature set 1000, and is abnormal when the feature vector x1002 is separated. Here, the distance 1006 between them is defined as the degree of abnormality.

Although the distance between the feature quantity vector x1002 and the feature quantity set 1000 may be calculated by the Euclidean distance with a low calculation cost, the Mahalanobis distance considering the variance of the feature quantity set 1000 is used here. If the inverse matrix of the variance-covariance matrix of the feature quantity set 1000 is S ⁻¹ , the Mahalanobis distance D can be calculated by the following equation (2).

D ² = (x−x _n ) ^t S ⁻¹ (x−x _n ) (2)
Next, the flow of processing when the determination threshold value 106 is calculated by the abnormal behavior detection apparatus of the present embodiment will be described using the flowchart of FIG.

  In step 1100, the abnormality degree calculation unit 102 is used to repeat the processing in step 1102 for all evaluation scenes. The evaluation scene is a video library of normal scenes and abnormal scenes, and is labeled with determination results to be determined by the apparatus.

  In step 1102, the determination threshold value calculation unit 110 is used to calculate the maximum abnormality level of each frame of the evaluation scene to be processed as a representative value of the abnormality level of the scene.

  In step 1104, the determination threshold calculation unit 110 is used to create an error curve, which will be described later, based on the maximum degree of abnormality for each scene calculated in step 1102. The error curve shows how the false alarm rate that erroneously determines a normal scene as an abnormal scene and the false alarm rate that erroneously determines an abnormal scene as a normal scene change according to the determination threshold. It is a curve to represent.

  In step 1106, an appropriate threshold value is determined based on the error curve created in step 1104. Details of this determination method will be described later.

Next, the procedure for calculating the maximum degree of abnormality for each scene shown in step 1102 of FIG. 11 will be described in detail with reference to FIG. The horizontal axis of the graph is the frame number, the vertical axis is the degree of abnormality, and the transition of the degree of abnormality for each frame when the evaluation scene is sequentially evaluated is plotted. A curve 1200 is an evaluation result of the degree of abnormality for each frame of the scene 1. The maximum value of the degree of abnormality is the case of the point 1204, and this value is set as the maximum degree of abnormality of the scene 1. Similarly, the curve 1202 is an evaluation result of the degree of abnormality with respect to the scene 2, and the degree of abnormality corresponding to the point 1206 is set as the maximum degree of abnormality of the scene 2.

  Next, the error curve calculation procedure shown in step 1104 of FIG. 11 will be described in detail with reference to FIGS. 13 and 14.

  FIG. 13 is a bar graph showing the maximum degree of abnormality for each scene calculated in step 1102 of FIG. In this example, scenes 1 to 7 are normal scenes, and scenes 8 to 14 are abnormal scenes. From this graph, it can be seen that the maximum abnormality degree tends to be small for a normal scene, and conversely, the maximum abnormality degree tends to be large for an abnormal scene.

  Here, a discrimination threshold value 1300 is introduced as a criterion for determining whether an abnormality is normal or normal. If the maximum degree of abnormality is smaller than the determination threshold 1300, the scene can be regarded as normal. On the other hand, when the maximum abnormality degree is the determination threshold value 1300 or more, the scene can be regarded as abnormal.

  The anomaly detection performance of this device can be evaluated by the false alarm rate and the false alarm rate. The false alarm rate is a rate at which a normal scene is erroneously determined as an abnormal scene, and the smaller the value, the better. In the case of FIG. 13, the maximum degree of abnormality 1302 of the scene 5 exceeds the determination threshold value 1300 in the normal seven scenes, and is erroneously determined as abnormal. In this case, the false alarm rate is 1/7 = 14%. On the other hand, the unreported rate is a rate at which an abnormal scene is erroneously determined as a normal scene, and the smaller the value, the better. In the case of FIG. 13, since the maximum abnormality degree 1304 of the scene 12 is smaller than the determination threshold value 1300 among the abnormal 7 scenes, it is erroneously determined as normal. In this case, the unreported rate is 1/7 = 14%.

The false alarm rate and the false alarm rate vary depending on the discrimination threshold value. This is shown in FIG. Curve 1400 represents the false alarm rate and curve 1402 represents the false alarm rate. As can be seen from the figure, there is a trade-off between the false alarm rate and the false alarm rate. That is, if the discrimination threshold is greatly adjusted to lower the false alarm rate, the false alarm rate increases. Conversely, the false alarm rate increases when the misregistration threshold is adjusted to a low value to reduce the false alarm rate. In step 1106 of FIG. 11, the determination threshold 106 is set based on this graph so that the detection performance desirable for the user of this apparatus is obtained.

  As candidates for the determination threshold 106, for example, there is a point 1404 where the false alarm rate and the false alarm rate are equal, that is, EER (Equal Error Rate). In this case, the determination threshold value 1406 corresponding to EER is set as the determination threshold value 106.

  In addition, a determination threshold 1408 when the unreported rate is 0% and a determination threshold 1410 when the misreporting rate is 0% may be set as the determination threshold.

  Furthermore, since the selection conditions for these discrimination thresholds differ depending on the purpose of use and usage conditions of the apparatus, the user of the apparatus may be able to select them.

  Further, the error curve of FIG. 14 may be presented to the user using an output means such as a display. In this case, the discrimination threshold may be selected on the screen using an input means such as a mouse / keyboard.

  These means have the effect of improving the degree of freedom of setting.

  According to the embodiment described above, it is possible to automatically determine whether or not an abnormality has occurred based on the degree of abnormality in the video. As a result, when an abnormality occurs, it is possible to immediately respond to the abnormality by issuing a warning or notifying the guard. Furthermore, it becomes possible to automatically determine a determination threshold value, which is a determination criterion for occurrence of abnormality, so as to be the best based on the evaluation result of the evaluation scene. Therefore, even when the target event for abnormality determination is changed, it is possible to flexibly cope with the situation only by creating an evaluation scene.

  In the embodiment described above, the partial space having a small contribution rate to the normal space shown in FIG. 9 is determined according to the preset cumulative contribution rate. However, the partial space is determined based on the fixed cumulative contribution rate. Instead of calculating, the partial space may be determined so that the detection accuracy is the best.

A method for determining such a partial space will be described with reference to FIG. In the drawing, the relationship between the EER value and n when the degree of abnormality is calculated in the partial space composed of the nth principal component to the 251st principal component is shown as an EER curve 1500. It is considered that the detection performance is better as the EER described in FIG. 14 is smaller. Therefore, the number n of principal components corresponding to the minimum value 1502 of the EER curve 1500 may be selected.

  As a result, the partial space can be automatically determined so that the detection accuracy is the best. Therefore, the best detection performance can be realized without user effort.

  Next, an example of a monitoring screen using the abnormal behavior detection device of the present invention is shown in FIG. This monitoring screen is displayed on a display device of a personal computer or a monitoring terminal incorporating the abnormal behavior detection device. In this example, the inside of the elevator car is photographed with a video camera.

  An area 1600 is an area for displaying a video currently being processed.

  An area 1602 is an area for sequentially displaying the transition of the calculated abnormality degree as a trend graph 1604 with the passage of time. A straight line 1606 is a discrimination threshold.

  An area 1608 is a display area for an abnormality determination result. Normality / abnormality can be distinguished and displayed, and according to the content of the determination result 108 for the current frame, which state is displayed.

  Further, the degree of abnormality may be displayed with a granularity of three or more levels instead of two levels of normal and abnormal. An area 1608 shows an example in which the degree of abnormality is displayed in three stages. For example, blue indicates normal, yellow indicates minor abnormality, and red indicates serious abnormality, so that the supervisor can grasp the situation intuitively.

  Next, details of the multi-stage abnormality determination process shown in the region 1608 of FIG. 16 will be described using the flowchart of FIG. Here, the ratio of the time occupied by the abnormality determination is calculated when going back to the past by a predetermined time from the present, and multi-stage determination is performed according to the ratio.

  In step 1700, an occupancy rate p for abnormality determination when going back in the past by a predetermined time from the present is calculated.

  In Step 1702, it is compared whether the occupation ratio p is smaller than a threshold value pb for blue determination designated in advance. If the comparison result is true, step 1704 is executed. If the comparison result is false, steps 1706 to 1710 are executed.

  In step 1704, the determination result is blue and the process is terminated.

  In step 1706, it is compared whether or not the occupation ratio p is smaller than a predetermined yellow determination threshold value py. If the comparison result is true, step 1708 is executed. If the comparison result is false, step 1710 is executed.

  In step 1708, the determination result is yellow and the process is terminated.

  In step 1710, the determination result is red, and the process ends.

  Although the above processing shows the case where the determination result is three stages, the determination may be made with a granularity higher than that. In that case, a threshold value may be set in advance according to the granularity. Thus, by showing the determination results in multiple stages, there is an effect that it is possible to provide the supervisor with easy understanding.

  According to the embodiment described above, it is possible to easily grasp the correspondence between the content of the video to be evaluated and the degree of abnormality of the video. For this reason, for example, even if there is a false alarm, the supervisor can easily confirm whether or not the abnormal behavior is actually occurring in the video.

  FIG. 18 shows an elevator apparatus equipped with the abnormal behavior detection apparatus according to the present invention. In the drawing, drive mechanisms such as a hoisting machine, a rope, and a counterweight are not shown. An image in the car is acquired by the video camera 30 provided in the car 40, and the video signal is transmitted to the abnormal behavior detection device 10 provided inside or outside the hoistway via the tail cord 50. . When detecting an abnormality in the video in the car 40, the abnormal behavior detection device 10 outputs a determination result signal indicating the abnormality to the elevator control device 20 provided in the hoistway or in the machine room. When the elevator control device 20 receives a determination result signal indicating an abnormality, the elevator control device 20 controls the hoisting machine so that the car 40 stops at the nearest floor and drives the door so as to open the car 40 and the landing door. Control the device. Alternatively, the control device 40 activates an alarm device such as an alarm provided in the car 40. According to such an elevator apparatus, it becomes possible for a passenger to quickly avoid an abnormal action person or to suppress an abnormal action in the car, thereby improving the security of the elevator apparatus. The abnormal behavior detection device 10 may be provided in the car 40. In this case, the determination result signal is transmitted to the control device 20 via the tail code 50.

It is a block diagram showing the functional structure of the abnormal action detection apparatus which is the position Example of this invention. It is a flowchart showing the flow of the process at the time of abnormality determination. It is a block diagram showing the function structure of an abnormality degree calculation part. It is a flowchart showing the flow of abnormality degree calculation processing. It is a figure explaining the frame used for calculation of a solid higher-order local autocorrelation. It is a figure explaining the mask pattern of a solid higher-order local autocorrelation. It is a flowchart showing the flow of a calculation process of a solid higher-order local autocorrelation feature. It is a flowchart showing the flow of the calculation process of a conversion matrix. It is a figure explaining the calculation process of a partial space. It is a figure explaining the evaluation method of an abnormality degree. It is a flowchart showing the flow of the process at the time of determination threshold value calculation. It is a figure explaining the maximum abnormal degree according to scene. It is a figure explaining a false alarm rate and a false alarm rate. It is a figure explaining an error curve. It is a figure explaining the determination process of a partial space. It is a figure explaining the example of a monitoring screen. It is a flowchart showing the flow of a three-stage abnormality determination process. An elevator apparatus provided with the abnormal action detection apparatus by this invention is shown.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 10 ... Abnormal action detection apparatus, 20 ... Elevator control apparatus, 30 ... Video camera, 40 ... Ride car, 50 ... Tail code, 100 ... Image | video acquisition part, 102 ... Abnormality calculation part, 104 ... Abnormality determination part, 106 ... Determination Threshold value 108... Determination result 110. Determination threshold value calculation unit 112.

Claims (6)

A video acquisition unit for acquiring video to be monitored;
An abnormality degree calculation unit for calculating the degree of abnormality of the video acquired by the video acquisition unit;
An abnormal behavior detection device comprising: an abnormality determination unit that determines, based on a threshold value, whether abnormal behavior has occurred from the degree of abnormality calculated by the abnormality level calculation unit,
The abnormality degree calculation unit
A motion extractor for extracting a portion with motion from the video;
A feature amount calculation unit that calculates a 251-dimensional first feature amount that is a stereoscopic higher-order local autocorrelation feature of the video generated by the motion extraction unit;
A feature amount conversion unit that converts the first feature amount into a second feature amount on a 252-n-dimensional subspace composed of the n-th principal component and the 251-th principal component by linear conversion by principal component analysis;
An abnormality degree evaluation unit that calculates an abnormality degree by comparing the second feature quantity of 252-n dimensions with a third feature quantity of 252-n dimensions for normal behavior;
When the feature amount conversion unit calculates a degree of abnormality in a subspace composed of the n-th principal component to the 251st principal component by performing a principal component analysis on the first feature amount, a false alarm rate and a false alarm rate are obtained. An abnormal behavior detection apparatus, wherein n is selected such that an equal EER (Equal Error Rate) is minimized, and a transformation matrix having a size of (252−n) × 251 is generated.   The abnormal behavior detecting device according to claim 1, wherein the abnormality degree evaluation unit calculates an abnormality degree based on a Euclidean distance between the second feature amount and a center of gravity of the third feature amount set.   The abnormal behavior detecting device according to claim 1, wherein the abnormality degree evaluation unit calculates an abnormality degree from a Mahalanobis distance of a set of the second feature quantity and the third feature quantity. In any one of claims 1 to 3, abnormal behavior detection, characterized in that it comprises the determination threshold calculation unit that sets the determination threshold value required for the determination process of the abnormality determination portion such determination result is best apparatus. 5. The determination threshold value calculation unit according to claim 4 , wherein the determination threshold value calculation unit sets a threshold value corresponding to any one of EER (Equal Error Rate), false error rate 0%, and false error rate 0%. An abnormal behavior detection device. The determination threshold value calculation unit according to claim 5 , further comprising an input unit that displays an error curve indicating a relationship between the determination threshold value, the false alarm rate, and the unreported rate on the display device, and selects the determination threshold value on the display device. To detect abnormal behavior.

Images