JP2010087937A - Video detection device, video detection method and video detection program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To detect as user-intended non-constant video image from among video images photographed by an imaging apparatus. <P>SOLUTION: Frame images of photographed video images are time-sequentially input to a video detection device 1. A change region extraction section 12 extracts a change region by frame sequentially input from an imaging apparatus. A cut segmentation section 14 segments a change region image included in a specific time unit as a video cut. A feature vector calculation section 15 calculates a feature vector from the video cut. A non-constant cut detection section 19 uses a dimension-compressed feature vector and a detection parameter to detect candidates of the non-constant video cut. A non-constant cut detection feedback section 23 updates the detection parameter using a feature vector corresponding to a video cut regarded constant among the candidates of the non-constant video cut. A non-constant cut detection section 19 uses the updated detection parameter to detect the candidates of the non-constant video cut. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a technique for detecting an unsteady cut by image processing from video captured by an imaging device (for example, a camera or a video camera).

  Currently, many surveillance systems using cameras are used, but detection of anomalies from moving images by humans is labor intensive, and if this can be replaced by a computer, it will lead to a significant labor reduction. For example, a bank or the like can detect a suspicious behavior in an ATM operation. In elderly care, if there is an automatic notification system for seizures, the burden on the caregiver can be reduced. Therefore, there is a need for a camera monitoring system for notifying abnormality.

  As the conventional non-stationary video detection technique, the abnormal operation described in Non-Patent Document 1 is learned in advance and the non-stationary video detection is performed, and the non-stationary video detection is performed without learning described in Non-Patent Document 2. The method is known.

In the detection method described in Non-Patent Document 1, a rampage of a person in an elevator is detected by learning abnormal operations in advance. In addition, in the detection method described in Non-Patent Document 2, a video having a feature that is significantly different from the motion feature of the person extracted from the previous video is detected as an abnormal operation.
Kentaro Hayashi, Norihito Seki, Hideki Shiozaki, Taibunya Tai, Mitsumasa Mori, Hiroaki Kaji, “Image Application Security Sensors for Elevators”, Recent Technologies and Progressive Technology Lectures in Elevators / Amusement Facilities, 2005 ( January 2006) pp. 29-32 Takuya Nanzato, Nobuyuki Otsu, “Detection of Abnormal Motion from Multiple Human Motion Images”, IEICE Technical Report PRMU 2004-77, 2004-09 (September 2004), pp. 9-16

  However, in the above-described non-stationary video detection technique, the criterion for determining non-stationary regardless of the presence or absence of learning largely depends on the method, and it is difficult to detect a non-stationary video as intended by the user.

  Therefore, in view of such problems, the present invention enables detection of a non-stationary video intended by the user by feeding back a result of determination of the user's steady / non-stationary with respect to a candidate for non-stationary video to a detection process. This is a solution issue.

  The present invention is a technical idea created in order to solve the above-described problem, and acquires a partial video of a predetermined frame on the basis of a place where the background has changed, and uses a feature vector of the partial video to obtain a steady state. If the distance deviating from the space showing is equal to or greater than a predetermined threshold, it is determined as a candidate for a non-stationary video, and is presented to the user to determine whether it is stationary or non-stationary. Then, the predetermined threshold is updated by the video section designated as steady and fed back to the detection process.

  Specifically, the invention described in claim 1 is a video detection device that detects an unsteady video cut in a shooting space in a video shot of a space by the imaging device, and is continuously from the imaging device. Extraction means for extracting a change area in which a change appears in an input frame; feature quantity calculation means for obtaining a feature vector of change from a video cut obtained by cutting out an image of the change area into an arbitrary column; and A value that deviates from the distribution of the identification space indicating the steady state of the imaging space is obtained, and when the value exceeds a threshold value, detection means that is a non-stationary video cut candidate, and among the candidates, the user is stationary Feedback means for sequentially updating the threshold value using the feature vector of the video cut considered to be reflected in the detection process of the detection means.

  The invention described in claim 2 is a video detection device that detects non-stationary video cuts in a shooting space in a video shot of a space by the imaging device, and is continuously input from the imaging device. A moving object is extracted from the frame, a vector calculating means for obtaining a feature quantity in the motion direction of the extracted moving object as a motion vector, and a motion vector cut obtained by extracting the motion vector into an arbitrary column to obtain a motion feature vector of the moving object. A feature amount calculation unit and a value that the feature vector deviates from the distribution of the identification space indicating the steady state of the shooting space is obtained, and when the value exceeds a threshold value, it is determined as a non-stationary video cut candidate. The threshold value is sequentially updated using the detection vector and the feature vector of the video cut that the user considers to be stationary among the candidates, and is reflected in the detection process of the detection unit. It is characterized in that it comprises a readback means.

  The invention according to claim 3 further includes compression means for compressing the dimension of the feature vector, and in the detection means, the value of the dimension-compressed feature vector deviating from the distribution of the identification space. It is characterized by seeking.

  The invention according to claim 4 further includes display means for presenting a non-stationary video cut candidate detected by the detection means to a user, and the display means detects the non-stationary image detected before the threshold is updated. It is characterized by being able to display video cut candidates together.

  The invention described in claim 5 is a video detection method for detecting a non-stationary video cut in a shooting space in a video shot of a space by the imaging device, wherein the extraction means is continuously connected to the imaging device. A first step of extracting a change area in which a change appears in the frame input to the second step, and a second step in which the feature amount calculation means obtains a change feature vector from a video cut obtained by cutting out an image of the change area into an arbitrary column; The detection means obtains a value out of the feature vector from the distribution of the identification space indicating the steady state of the shooting space, and sets the value as a non-stationary video cut candidate when the value exceeds a threshold value. 3 steps, and the feedback means sequentially updates the threshold value using the feature vector of the video cut that the user considers to be stationary among the candidates, and reflects it in the detection process of the detection means. It is characterized by having 4 and steps.

  The invention described in claim 6 is a video detection method for detecting a non-stationary video cut in a shooting space in a video shot of a space by the imaging device, wherein the vector calculation means is continuously connected to the imaging device. A first step of extracting a moving object from a frame that is input automatically, and obtaining a feature quantity in the motion direction of the extracted moving object as a motion vector, and a motion vector in which the feature quantity calculation means cuts the motion vector into an arbitrary column A second step of obtaining a feature vector of motion of the moving object from the cut; and a detecting means obtains a value that is out of the feature vector from the distribution of the identification space indicating the steady state of the shooting space, and the value exceeds a threshold value And a third step of selecting a non-stationary video cut as a candidate and the feature vector of the video cut that the feedback unit considers to be stationary among the candidates. Sequentially updates the threshold using, is characterized by a fourth step of reflecting the detection process of the detection means.

  The invention according to claim 7 further includes a step of compressing the dimension of the feature vector, and in the third step, a value for which the dimension-compressed feature vector deviates from the distribution of the identification space is obtained. It is characterized by that.

  The invention according to claim 8 further includes a step of presenting a non-stationary video cut candidate detected in the third step to the user, and in this step, the non-stationary detected before the threshold is updated. It is characterized by being able to display video cut candidates together.

  The invention described in claim 9 is a video detection program, characterized by causing a computer to function as each means constituting the video detection device according to any one of claims 1 to 4.

  According to the first to ninth aspects of the present invention, the determination result of the user with respect to the candidate for the unsteady video is reflected in the subsequent detection processing, so that the detection intended by the user is possible.

  In particular, according to the invention of claim 4.8, the current non-stationary video candidate and the past non-stationary video candidate can be compared and confirmed.

  In the present invention, an unsteady image cut in the space is detected in an image obtained by continuously capturing a certain space with an imaging device such as a camera fixed at a predetermined position. Examples of the shooting space include an ATM corner of a bank and an elevator car. In this case, the internal space is continuously photographed by a camera installed on the ceiling of an ATM corner or the like. At this time, when an unsteady state such as the intrusion of a suspicious person is photographed by the camera, a video cut of the scene is detected.

  Here, a description will be given assuming a situation (a still image series, a video stream, etc.) in which images (frame images) captured by the imaging device are input to the video detection device in a time series. In addition, although an example in which the video detection device is configured by a computer will be described here, the video detection device is not limited to this, and includes, for example, an IC (Integrated Circuit) chip on which video detection processing logic is mounted. It may be a calculator.

(1) First Embodiment FIG. 1 shows a video detection apparatus 1 according to a first embodiment of the present invention. In the video detection apparatus 1, image data taken by a camera 2 is time-sequentially via a network. Has been entered. Here, it is assumed that a digital camera or a digital video camera is used for the camera 2 as the imaging device.

  As shown in FIG. 1, the video detection apparatus 1 includes an input unit 3 such as a keyboard and a mouse, a read-only memory (ROM) 4 that stores fixed data, and a processing unit that performs control and calculation processing of each unit. (CPU: Central Processor Unit, etc.) 5, a rewritable memory (RAM) 6 for temporarily storing processing data and the like, a communication interface unit (I / F) 7 used for network connection with the camera 2 and the like, An external storage unit 8 such as a hard disk drive device and a recording medium drive unit 9 for reading a recording medium 10 such as a CD-ROM or a DVD-ROM are provided. Further, a display device 11 such as a display for displaying processing data or the like is connected to the video detection device 1.

  As shown in FIG. 2, each of the components 3 to 9 includes a change region extraction unit 12, a change region storage unit 13, a cut cutout unit 14, a feature vector calculation unit 15, a dimension compression parameter storage unit 16, and a dimension compression unit 17. , Functioning as an outlier detection parameter storage unit 18, an unsteady cut detection unit 19, an unsteady cut storage unit 20, an unsteady cut display unit 21, a steady cut acquisition unit 22, and an unsteady cut detection feedback unit 23. The functions of the functional blocks 12 to 23 are realized by the processing unit (CPU) 5 reading a video detection program. Among these, each said memory | storage part 13.16.18.20 is constructed | assembled as a database on the said hard-disk drive apparatus.

  Specifically, as shown in FIG. 3, image data (frames) photographed by the camera 2 is sequentially input to the change area extraction unit 12 in time series. The change area where the change appears is extracted from the image data, and the data of the extracted change area image is sequentially output to the change area storage unit 13.

  The change area storage unit 13 stores the input data of the change area image in time series to form a change area image sequence, and outputs the data of the image sequence to the cut cutout unit 14.

  The cut cutout unit 14 cuts out a certain number of frames from the input change area image sequence, and outputs the cut out frames to the feature vector calculation unit 15 as change area video cuts.

  The feature vector calculation unit 15 calculates a feature vector from the input change area video cut, and outputs the calculated feature vector to the dimension compression unit 17.

  The dimension compression unit 17 compresses the dimension of the feature vector input from the feature vector calculation unit 15 using the dimension compression parameter input from the dimension compression parameter storage unit 16, and the obtained dimension compression vector Is output to the unsteady cut detector 19.

  The unsteady cut detection unit 19 applies a predetermined identification method to the outlier detection parameter input from the outlier detection parameter storage unit 18 to identify the dimension compression vector input from the dimension compression unit 17. Determine if it is out of the spatial distribution. Then, a video cut corresponding to the dimensional compression vector that has been removed is detected as a candidate for a non-stationary video cut (hereinafter referred to as a non-stationary candidate), and the detected candidate is output to the non-stationary cut storage unit 20.

  The unsteady cut storage unit 20 sequentially stores the unsteady candidates input from the unsteady cut detection unit 19 and outputs the stored unsteady candidate group to the unsteady cut display unit 21.

  The unsteady cut display unit 21 outputs the unsteady candidate group input from the unsteady cut storage unit 20 to the steady cut acquisition unit 22 and causes the display device 11 to display the unsteady candidate group. For each video of the displayed non-stationary candidate group, the user determines whether each video is a stationary video or a non-stationary video, and inputs the determination result through the input unit 3.

  The steady cut acquisition unit 22 acquires a steady / unsteady label attached to each video as a determination result made by the user for each video in the non-stationary candidate group. Then, from the non-stationary candidate group output from the non-stationary cut display unit 21, a video cut group to which a stationary label is assigned is extracted, and is used as a stationary video cut group to the non-stationary cut detection feedback unit 23. Output.

  When the unsteady cut detection feedback unit 23 receives the steady image cut group from the steady cut acquisition unit 22, the unsteady cut detection feedback unit 23 learns the identification method using a dimensional compression vector corresponding to the cut group. As a result, the outlier detection parameter is updated, and the updated parameter is output to the outlier detection parameter storage unit 18.

  The non-stationary cut detection unit 19 performs subsequent detection of non-stationary candidates using the outlier detection parameter updated in this way. Thereby, the determination result of the user is fed back to the subsequent detection process as needed, and the detection result closer to the user's intention can be output. The details of the input / output data of each of the functional blocks 12 to 23 are shown in Table 1.

  Hereinafter, processing steps executed by the functional blocks 12 to 23 will be described with reference to the flowchart of FIG. S01 to S10 in FIG. 4 are processed based on the input / output data shown in Table 1.

  S01: First, the change area extraction unit 12 compares the current input image data captured by the camera 2 with the average image data of the past several frames captured in advance by the camera 2 in the input image. A changed region image is obtained by extracting a region where a change from the steady state appears. It is assumed that the average image data of the past several frames is stored in the hard disk drive device or the like.

  Specifically, the change region extraction unit 12 applies a change region extraction method (or change detection method) that takes into account temporal changes in the background, and extracts a change region image from the current input image taken by the camera 2. get. As the change area extraction method, a method of extracting an area that exceeds a certain threshold value by a difference in pixel values between an average image of the past several frames and the current frame is considered as a change area. In this way, by using the change area extraction method that takes into account temporal changes, for example, even if the pixel value changes according to changes in morning, noon, night, and sunshine, the change area can be extracted robustly. Can do. At this time, a pixel value 1 is assigned to the change area of the change area image, and a pixel value 0 is assigned to the area where there is no change.

  Then, the change area extraction unit 12 sequentially stores the data of the change area image acquired in this manner in the change area storage unit 13 to generate a change area image sequence.

  S02: The cut cutout unit 14 reads out the image sequence of the change area stored in the change region storage unit 13, and cuts out a certain number (or a fixed time) of frames from the image sequence. Then, the fixed number of frames that have been cut out are output to the feature vector calculation unit 15 as a change area video cut.

  Here, the number of frames (or time) to be cut out is defined in advance in the video detection program. Note that such cutout of the change area video cut may be performed every time a certain number of frames are accumulated in the change area storage unit 13. Further, the cut data of the changed area video cut may be temporarily stored in the memory (RAM) 6 or the like, or may be stored in the external storage unit 8.

  S03: The feature vector calculation unit 15 calculates a feature vector corresponding to the video cut from the change area video cut cut out in S02, and outputs the calculated vector to the dimension compression unit 17.

  Here, the feature vector is an x × y × n-dimensional feature vector obtained by multiplying the cut-out constant frame number n, where the image size is x × y (vertical x pixel, horizontal y pixel) in the change area image. . Note that the calculated feature vector data may be temporarily stored in the memory (RAM) 6 or the like, or may be stored in the external storage unit 8.

  S04: The dimension compression unit 17 compresses the dimension of the feature vector calculated in S03 using a dimension compression parameter, and obtains a feature vector with a small number of dimensions. Here, principal component analysis (PCA), which is a compression method used in the field of similar scene retrieval from video, is used for dimension compression.

  Specifically, a principal component analysis of a feature vector F (k) (k = 1,..., K, K is the number of cuts) corresponding to a video cut acquired in advance is performed. Here, when the feature vector F (k) is projected onto the compression coordinate system, the feature vector F (k) is projected onto the principal axis up to the order exceeding the predetermined contribution rate, and the coefficient obtained at that time is expressed as a dimension. Obtained as a compression parameter and stored in advance in the dimension compression parameter storage unit 16. Then, the dimension compression unit 17 reads the dimension compression parameter from the dimension compression parameter storage unit 16, and uses this to compress the dimension of the feature vector calculated in S03. Then, the compressed vector is used as a dimensional compression vector and is output to the unsteady cut detection unit 19. Thus, by compressing the dimension of the feature vector, the information amount of the vector can be reduced.

  When the environmental condition changes with a long-time video, the dimensional compression may be performed by obtaining the dimensional compression parameter for each group of a certain amount of time.

  S05: The unsteady cut detection unit 19 uses the outlier detection parameter stored in the outlier detection parameter storage unit 18 in advance, and the dimension compression vector calculated in S04 deviates from the distribution of the identification space indicating the steady state. And the video cut corresponding to the off-dimensional compression vector is detected as a non-stationary candidate.

  Here, outlier detection is performed using a support vector machine (SVM) as an identification method. For example, the 1 class SVM uses the fact that the outlier point in the original feature space is mapped near the origin of the higher-order feature space when the feature vector in the original feature space is mapped to the higher-order feature space, This is a method of detecting outliers. That is, by using the outlier detection parameter, the distance from the origin of the dimension compression vector mapped on the identification axis is regarded as an outlier (unsteadiness), and this unsteadiness is a predetermined value (threshold value). Exceeding) shall be the off-point. Then, the video cuts corresponding to the dimension compression vectors determined to be out of place are set as non-stationary candidates, which are sequentially stored in the non-stationary cut storage unit 20 to generate a non-stationary candidate group.

  It is assumed that the threshold value for identifying an outlier is defined in the video detection program. In addition, the outlier detection parameter may be obtained by using an arbitrary video cut acquired in advance, or when environmental conditions change in a long-time video, for every certain amount of time. You may ask for it.

  S06: The unsteady cut display unit 21 reads the unsteady candidate group from the unsteady cut storage unit 20 and outputs the candidate group to the steady cut acquisition unit 22. Moreover, as shown in FIG. 5, the read non-stationary candidate group is displayed on the display device 11 such as a display in the order of higher non-stationary degree. Thus, the user can view the current non-stationary video cut candidates and the past non-stationary video cut candidates while comparing them. At this time, the unsteadiness degree for each image may be displayed in the region R in FIG.

  Further, as shown in FIG. 6, a stationary / unsteady selection interface (region S) may be displayed together with the unsteady video cut candidates and the unsteadiness degree. Thereby, the user can select by clicking whether or not each displayed video is steady. Here, radio buttons are displayed as the selection interface, but the interface is not limited to this, and may be appropriately changed according to the specification.

  S07: For each image of the non-stationary candidate group displayed in S06, it is confirmed whether or not a determination of whether the user is stationary or unsteady is input from the input unit 3 such as a mouse. When there is an input, the process proceeds to S08, and when there is no input, the process proceeds to S10.

  S08: The steady cut acquisition unit 22 acquires a steady / unsteady label given to each video as a user determination result for each video of the non-stationary candidate group displayed in S06. Among these, a video cut corresponding to a stationary label is extracted from the non-stationary candidate group received from the non-stationary cut display unit 21 and is output to the non-stationary cut detection feedback unit 23 as a stationary video cut group. . Note that the extracted data of the regular video cut group may be temporarily stored in the memory (RAM) 6 or the like, or may be stored in the external storage unit 8.

  S09: The non-stationary cut detection feedback unit 23 learns the SVM using the dimension compression vector corresponding to the stationary video cut group extracted in S08. Thereby, the outlier detection parameter is updated to a value reflecting the steady / unsteady determination made by the user in S06. The unsteady cut detection feedback unit 23 stores the outlier detection parameter updated in this way in the outlier detection parameter storage unit 18.

  This updated outlier detection parameter is used for detection of non-stationary candidates after the next time. That is, in the process of S05, the non-stationary cut detection unit 19 detects non-stationary candidates using the outlier detection parameters that are sequentially updated. Therefore, in the subsequent detection process, a non-stationary candidate reflecting the user's intention is detected, and a video cut with a high degree of unsteadiness is displayed on the display device 11 in S06.

  S10: Next, it is determined whether the end condition is satisfied. Here, when the processing of all the frames is completed, the detection process is terminated. If all the frames have not been processed, the process returns to S01, and S01 to S09 are repeated until the end condition is satisfied.

  As described above, according to the video detection device 1 according to the present embodiment, the user's judgment on the non-stationary video cut candidate is fed back to the subsequent non-stationary video cut detection process. It is possible to detect a steady image cut.

  In addition, by repeating the determination of stationary / unsteady with respect to the unsteady video cut candidates and the update of the outlier detection parameter, the detection of unsteady video cuts customized for each user can be realized with high accuracy. In this sense, false detection is prevented and the work burden on the user is reduced.

(2) Second Embodiment FIG. 7 shows a video detection device 26 according to a second embodiment of the present invention, in which the dimension compression parameter storage unit 16 and the dimension compression unit 17 are eliminated. In addition, instead of the change region extraction unit 12 and the change region storage unit 13, a motion vector calculation unit 24 and a motion vector storage unit 25 are provided.

  That is, the detection device 26 calculates a motion vector instead of extracting a change area from the input image, and detects a candidate for an unsteady video cut using this motion vector. The motion vector storage unit 25 is constructed as a database on the hard disk drive device in the same manner as the change area storage unit 13.

  Specifically, as shown in FIG. 8, the motion vector calculation unit 24 calculates a motion vector from image data sequentially input from the camera 2 and sequentially outputs the vector to the motion vector storage unit 25.

  The motion vector storage unit 25 stores the input motion vectors in time series to form a motion vector sequence, and outputs the vector sequence to the cut cutout unit 14.

  The cut cutout unit 14 cuts out a certain number of frames from the input motion vector sequence, and outputs the cut out frames to the feature vector calculation unit 15 as motion vector cuts.

  The feature vector calculation unit 15 calculates a feature vector from the input motion vector cut, and outputs the calculated feature vector to the unsteady cut detection unit 19.

  Then, as in the first embodiment, the non-stationary candidate is detected using the calculated feature vector, the outlier detection parameter is updated using the user's determination result for the detected non-stationary candidate, and the updated The parameter is fed back to the subsequent detection process.

  Table 2 shows details of input / output data of each of the functional blocks 14.15.18 to 25 in the present embodiment.

  Hereinafter, processing steps executed by each of the functional blocks 14.15.18 to 25 in the present embodiment will be described based on the flowchart of FIG. S21 to S29 in FIG. 9 are processed based on the input / output data in Table 2.

  S21: First, the motion vector calculation unit 24 detects the optical flow from the current input image data photographed by the camera 2. The optical flow represents the amount of movement and direction at each point on the image, and is obtained by searching for corresponding points between temporally continuous images. For the detection of optical flow, a method based on luminance gradient and a method based on region matching are conceivable, but a method in which a flow is stably generated may be selected depending on the scene.

  Then, the pixel position (u, v) and the motion direction (U, V) at the detection start point of the detected optical flow are calculated as motion vectors, and the motion vector data are sequentially stored in the motion vector storage unit 25, Generate a motion vector sequence.

  S22: The cut cutout unit 14 reads the motion vector sequence stored in the motion vector storage unit 25, and cuts out a certain number of frames (or a fixed time) from the vector sequence. Then, the fixed number of cut out frames are output to the feature vector calculation unit 15 as a motion vector cut.

  The motion vector cut may be cut out every time a certain number of frames are accumulated in the motion vector storage unit 25. The extracted motion vector cut may be temporarily stored in the memory (RAM) 6 or the like, or may be stored in the external storage unit 8.

  S23: The feature vector calculation unit 15 calculates a feature vector corresponding to the motion vector cut from the motion vector cut cut out by the cut cutout unit 14, and sends the calculated vector to the unsteady cut detection unit 19. Output. Here, the feature vector is a feature vector corresponding to a certain number n of motion vectors (u, v, U, V).

  The calculated feature vector may be temporarily stored in the memory (RAM) 6 or the like, or may be stored in the external storage unit 8. Similarly to the first embodiment, the dimension compression parameter storage unit 16 and the dimension compression unit 17 are provided to perform dimension compression of the calculated feature vector, and the compressed feature vector is converted into the unsteady cut detection unit. 19 may be output.

  S24 to S29: Similarly to S05 to S10, non-stationary candidates are detected using the feature vector calculated in S23 and outlier detection parameters, and these are sequentially stored in the non-stationary cut storage unit 20 (S24). . Then, the stored unsteady candidate group is displayed on the display device 11 (S25), and a steady image cut group is acquired based on the user's judgment on each image of the displayed unsteady candidate group (S27). Then, the outlier detection parameter is updated by learning the SVM using the feature vector corresponding to the acquired stationary video cut group, and the updated parameter is stored in the outlier detection parameter storage unit 18. (S28). In the process of S24, the unsteady cut detection unit 19 detects unsteady candidates using the outlier detection parameters that are sequentially updated.

  As described above, according to the video detection device 26 according to the present embodiment, although the non-stationary video cut candidate is detected using the motion vector calculated from the input image, the detected non-stationary video cut candidate is detected. Since the user's judgment is fed back to the subsequent detection process, the same effect as that of the first embodiment can be obtained.

  In addition, this invention is not limited to the said embodiment, For example, a computer can be provided as a video detection program which functions as each functional block 12-25 of the said video detection apparatus 1.26. This program may cause the computer to execute all the processing steps of the first embodiment, or may execute a part of the processing steps.

  This program is provided to the computer by downloading from a website or the like. The program is stored in a recording medium 10 such as a CD-ROM, a DVD-ROM, a CD-R, a CD-RW, a DVD-R, a DVD-RW, an MO, an HDD, or a Blu-ray Disk (registered trademark). May be provided to a computer.

1 is a configuration diagram of a video detection apparatus according to a first embodiment of the present invention. The main functional block diagram. The input / output data diagram. FIG. Display example of the same unsteady video cut group. The other example of the display of the same unsteady image cut group. The main functional block diagram of the image | video detection apparatus which concerns on 2nd Embodiment of this invention. The input / output data diagram. FIG.

Explanation of symbols

1.26 ... Video detection device 2 ... Camera (imaging device)
3 ... Input unit 4 ... Read only memory (ROM)
5. Processing unit (CPU)
6 ... Rewritable memory (RAM)
DESCRIPTION OF SYMBOLS 7 ... Communication interface part 8 ... External storage part 9 ... Recording medium drive part 10 ... Recording medium 11 ... Display apparatus 12 ... Change area extraction part (extraction means)
DESCRIPTION OF SYMBOLS 13 ... Change area memory | storage part 14 ... Cut cutout part 15 ... Feature vector calculation part (feature amount calculation means)
16 ... Dimension compression parameter storage unit 17 ... Dimension compression unit (compression means)
18 ... Outlier detection parameter storage unit 19 ... Unsteady cut detection unit (detection means)
20 ... Unsteady cut storage unit 21 ... Unsteady cut display unit (display means)
22 ... Steady cut acquisition unit 23 ... Unsteady cut detection feedback unit (feedback means)
24 ... Motion vector calculation unit (vector calculation means)
25: Motion vector storage unit R: Display area of non-stationary degree S: Display area of selected interface

Claims (9)

A video detection device for detecting a non-stationary video cut in a shooting space in a video shot of a space with an imaging device,
Extracting means for extracting a change area in which a change appears in frames continuously input from the imaging device;
Feature amount calculation means for obtaining a feature vector of change from a video cut obtained by cutting out the image of the change area into an arbitrary column;
A detection means for obtaining a value that deviates from the distribution of the identification space indicating the steady state of the shooting space, and when the value exceeds a threshold value,
Feedback means for sequentially updating the threshold value using the feature vector of the video cut that the user considers to be stationary among the candidates, and reflecting it in the detection process of the detection means;
A video detection apparatus comprising: A video detection device for detecting a non-stationary video cut in a shooting space in a video shot of a space with an imaging device,
A vector calculation means for extracting a moving object from frames continuously input from the imaging device and obtaining a feature quantity in the movement direction of the extracted moving object as a motion vector;
A feature quantity calculating means for obtaining a motion feature vector of a moving object from a motion vector cut obtained by cutting the motion vector into an arbitrary column;
A detection means for obtaining a value that deviates from the distribution of the identification space indicating the steady state of the shooting space, and when the value exceeds a threshold value,
Feedback means for sequentially updating the threshold value using the feature vector of the video cut that the user considers to be stationary among the candidates, and reflecting it in the detection process of the detection means;
A video detection apparatus comprising: And further comprising compression means for compressing the dimension of the feature vector,
3. The video detection apparatus according to claim 1, wherein the detection unit obtains a value in which the dimension-compressed feature vector deviates from the distribution of the identification space. 4. And further comprising display means for presenting a user with the unsteady video cut candidate detected by the detection means,
The video detection apparatus according to any one of claims 1 to 3, wherein the display unit is capable of displaying a non-stationary video cut candidate detected before the threshold is updated. A video detection method for detecting a non-stationary video cut in a shooting space in a video shot of a space with an imaging device,
A first step of extracting a change area in which a change appears in frames continuously input from the imaging device;
A second step in which a feature quantity calculating means obtains a feature vector of change from a video cut obtained by cutting out the image of the change area into an arbitrary column;
The detection means obtains a value that deviates the feature vector from the distribution of the identification space indicating the steady state of the shooting space, and when the value exceeds a threshold value, a third candidate is set as a non-stationary video cut candidate. Steps,
A fourth step in which the feedback means sequentially updates the threshold value using the feature vector of the video cut that the user considers to be stationary among the candidates, and reflects it in the detection process of the detection means;
A video detection method comprising: A video detection method for detecting a non-stationary video cut in a shooting space in a video shot of a space with an imaging device,
A first step in which a vector calculating means extracts a moving object from frames continuously input from the imaging device, and obtains a feature quantity in the moving direction of the extracted moving object as a motion vector;
A second step of calculating a feature vector of motion of the moving object from a motion vector cut obtained by cutting out the motion vector into an arbitrary column;
The detection means obtains a value that deviates the feature vector from the distribution of the identification space indicating the steady state of the shooting space, and when the value exceeds a threshold value, a third candidate is set as a non-stationary video cut candidate. Steps,
A fourth step in which the feedback means sequentially updates the threshold value using the feature vector of the video cut that the user considers to be stationary among the candidates, and reflects it in the detection process of the detection means;
A video detection method comprising: Further comprising compressing the dimension of the feature vector;
The video detection method according to claim 5, wherein, in the third step, a value in which the dimension-compressed feature vector deviates from the distribution of the identification space is obtained. Further comprising the step of presenting a non-stationary video cut candidate detected in the third step to the user,
The video detection method according to any one of claims 5 to 7, wherein in the step, candidates for unsteady video cuts detected before the threshold value is updated can be displayed together.   A video detection program for causing a computer to function as each means constituting the video detection device according to claim 1.

Images