JP2015114685A - Video search apparatus, video search method, and storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve search speed and reduce search noise, while reducing the work for selecting a query image.SOLUTION: A video search apparatus performs: detecting moving paths of one or more moving bodies from each of a first video formed of a plurality of frames captured in a first location and a second video formed of a plurality of frames captured in a second location, and storing them on a storage device; extracting image feature quantities of each of the frames, of a moving body selected from among the one or more moving bodies detected from the first video, and storing them on the storage device; selecting a query image feature quantity to be used as a search query, out of the image feature quantities extracted, on the basis of the moving path of the selected moving body detected from the first video and the moving paths of the one or more moving bodies detected from the second video; and searching for image feature quantities of the one or more moving bodies extracted from the second video, by use of the query image feature quantity, and outputting a result of the search.

Description

  The present invention relates to a video search technique.

  With the widespread use of security cameras, there is an increasing need for searching for a desired person or vehicle from images taken at multiple points. However, many conventional security camera systems are systems including a security camera, a recorder, and a playback device, and it has been difficult to search for a desired scene from a huge amount of accumulated data.

  On the other hand, attention has been focused on a system using a similar image search technique. By using the similar image search technique, it is possible to search a frame in which a specific person or object is shown from a large amount of video information. The similar image search is a technique for acquiring, from a database, an image similar in appearance characteristics to an image of a search query designated by a user. In calculating the similarity of objects, numerical data called feature values are extracted from an effective area (significant area) for distinguishing between objects and compared. When applied to a security camera system, feature quantities are extracted from a salient region such as a person's face or clothes. For example, in Patent Document 1, an image acquired from a camera is divided into blocks, and the feature amount of the block is extracted based on a color histogram, which is used as a similar image search query.

  On the other hand, a technique is known in which a dynamic object is detected from continuous frames extracted from a video, and a dynamic object is associated between frames. For example, the frame can be divided into small areas, and the motion vector of each small area can be calculated between the frames. A dynamic object can be tracked by observing a motion vector and grouping small regions that move in the same manner. As a result, the dynamic object can be tracked as long as it exists in the frame, so that the object designated by the user can be searched from another frame included in the same video.

JP 2011-18238 A

  In Patent Document 1, since a search is performed using a query specified by a user, a desired search result may not be obtained if the query is not appropriate. Further, Patent Document 1 shows a method of searching for a block similar to a query from frames before and after a frame including a query image, and performing a similar image search in a database using all of them. In this method, Even if the saliency area on the same dynamic object is significantly different from the query specified by the user, it is not selected as a search query, so the search results are limited.

  In order to solve the above-described problem, the present invention provides a video search device having a processor and a storage device connected to the processor, wherein the first search device includes a plurality of frames taken at a first location. , And a second image made up of a plurality of frames taken at a second location, the moving path of one or more moving objects is detected and stored in the storage device, and the first An image feature amount for each frame of the selected moving body among the one or more moving bodies detected from the video is extracted and stored in the storage device, and the selection detected from the first video A query image feature to be used as a search query out of the extracted image feature amounts based on the movement route of the mobile object and the movement route of the one or more mobile objects detected from the second video. Select the amount and before Using the query image feature amount, and searches the image feature quantity of the one or more mobile extracted from the second image, and outputs the result of the search.

  According to the video search apparatus of the present invention, parameters for tracking a dynamic object and detecting a saliency area from a large number of input videos and determining a query suitable for each shooting location from the accumulated tracking information. When the user specifies an object to search for, the query image suitable for the search is automatically determined from the flow line of the object, and the similar image search is performed to reduce the user's work of selecting the query image. be able to. In addition, since only a query suitable for each shooting location is used, effects such as an improvement in search speed and a reduction in search noise can be obtained. Problems, configurations, and effects other than those described above will become apparent from the following description of embodiments.

1 is a configuration diagram of a video search system according to Embodiment 1 of the present invention. It is a hardware block diagram of the image | video search system which concerns on Example 1 of this invention. It is a figure which shows the structure and data example of a video database which concern on Example 1 of this invention. It is a figure for demonstrating operation | movement of the video search system which concerns on Example 1 of this invention. It is a flowchart explaining the process which the image | video search device which concerns on Example 1 of this invention registers the input image | video. It is a figure explaining the estimation process of the query parameter which the video search device concerning Example 1 of the present invention performs. It is explanatory drawing of the dispersion | distribution ratio of the image feature-value used in order for the video search device which concerns on Example 1 of this invention to determine a remarkable area | region. It is a figure which shows the structure and example of a query parameter storage part which concern on Example 1 of this invention. It is a flowchart explaining the process which estimates a query parameter from the data which the video search device based on Example 1 of this invention accumulated. It is a figure explaining the operation | movement which the query determination part which concerns on Example 1 of this invention determines a search query using tracking information. It is a flowchart explaining the process which the image search apparatus which concerns on Example 1 of this invention performs a similar image search with the search query determined using tracking information. It is a figure explaining the processing sequence of the image | video search system which concerns on Example 1 of this invention. It is a figure showing the structural example of the operation screen used when searching the object in an image | video using the image | video search device which concerns on Example 1 of this invention. It is a figure for demonstrating correction | amendment of the tracking information using the remarkable area | region by the video search system which concerns on Example 2 of this invention. It is a flowchart explaining the process which the image | video search device which concerns on Example 2 of this invention correct | amends tracking information using a remarkable area | region. It is a figure for demonstrating correction | amendment of the tracking information which considered the depth information by the video search system which concerns on Example 2 of this invention. It is a flowchart explaining the process which the video search apparatus which concerns on Example 2 of this invention adds depth information to tracking information using a remarkable area | region. It is explanatory drawing regarding presentation of the query which exists in the different image | video in Example 3 of this invention. It is a flowchart explaining the process which the image | video search device which concerns on Example 3 of this invention searches for a new type | mold remarkable area | region from a different image | video. It is a figure for demonstrating the video summary using the tracking information in Example 4 of this invention. It is a flowchart showing the process of the video summary using the tracking information which the video search system which concerns on Example 4 of this invention performs.

  <System configuration>

  FIG. 1A is a configuration diagram of a video search system 100 according to Embodiment 1 of the present invention.

  The video search system 100 selects an object in the video specified by the user from a different time zone of the video (for example, from a frame at a time different from the frame including the object specified by the user) or from a different video (for example, , A system for searching (from a video taken at a location different from the video including the object specified by the user), and a query suitable for searching using tracking information of a dynamic object (moving object) in the video This system aims to improve the speed and accuracy of search by generating an image for each video shooting location where a search is performed using the image.

  The video search system 100 includes a video storage device 101, an input device 102, a display device 103, and a video search device 104.

  The video storage device 101 is a storage medium for storing video data, and is configured using a hard disk drive built in a computer, or a storage system connected via a network such as NAS (Network Attached Storage) or SAN (Storage Area Network). can do. The video storage device 101 may be a cache memory that temporarily holds video data continuously input from a camera, for example.

  Note that the video data stored in the video storage device 101 may be data in any format as long as it can be used for tracking a captured dynamic object. For example, the stored video data may be moving image data shot by a video camera, or a series of still image data shot by a still camera at a predetermined interval.

  The input device 102 is an input interface for transmitting user operations to the video search device 104 such as a mouse, a keyboard, and a touch device. The display device 103 is an output interface such as a liquid crystal display, and is used for displaying a search result of the video search device 104, interactive operation with a user, and the like.

  The video search device 104 detects and accumulates the tracking of the dynamic object and the saliency area from each frame of the given video data. When the user designates an object to be searched from the accumulated frames, the video search device 104 is suitable for each video shooting location where the search is performed using the query image from a series of frames before and after the tracking information. A query image is selected and a similar image search is executed. It is assumed that the video handled by the video search device 104 is a fixed-point observation video shot at one or more places. Further, the search target object is an arbitrary dynamic object such as a person or a vehicle. The video search device 104 includes a video input unit 105, a frame registration unit 106, a dynamic object tracking unit 107, a tracking information registration unit 108, a saliency area detection unit 109, a saliency area registration unit 110, a video database 111, and a query parameter estimation unit 112. A query parameter storage unit 113, a query input unit 114, a query determination unit 115, and a similar image search unit 116.

  The video input unit 105 reads video data from the video storage device 101 and converts it into a data format used inside the video search device 104. Specifically, the video input unit 105 performs a video decoding process that decomposes video (moving image data format) into frames (still image data format). The obtained frame is sent to the frame registration unit 106, the dynamic object tracking unit 107, and the saliency area detection unit 109.

  The frame registration unit 106 writes information about the extracted frame and the source video to the video database 111. Details of data recorded in the video database 111 will be described later with reference to FIG.

  The dynamic object tracking unit 107 detects a dynamic object in the video and tracks the dynamic object by associating it with the dynamic object of the previous frame. The detection and tracking of dynamic objects is a method described in, for example, S. Baker and I. Matthews "Lucas-kanade 20 years on: A unifying framework", International Journal of Computer Vision, vol. 53, no. 3, 2004. It is realizable using arbitrary methods. The obtained tracking information of the dynamic object includes coordinate information of the dynamic object of each frame and an ID (tracking ID) uniquely assigned to each tracking.

  The tracking information registration unit 108 extracts an image feature amount from the dynamic object region of each frame obtained by the dynamic object tracking unit 107 and registers it in the video database 111. The image feature amount is, for example, data represented by a fixed-length vector, and is obtained by digitizing appearance information such as an image color and shape. Further, the tracking information registration unit 108 extracts the feature amount of the flow line (that is, the movement path of the dynamic object) from the coordinates of the dynamic object to which the same tracking ID is assigned, and registers it in the video database 111.

  The saliency area detection unit 109 detects a saliency area from the frame and obtains its coordinates. The salient area differs depending on the application. For example, if the image includes a person, it is a face area, head region, clothing color, clothes pattern, or belongings. If the image includes a vehicle, the wheel or front grille is used. Etc. The saliency detection unit 109 includes a plurality of detection modules for extracting a saliency area corresponding to the type of the object. When the type of the object appearing in the video cannot be limited, the plurality of detection modules are operated simultaneously in parallel. Also good.

  The saliency area registration unit 110 extracts an image feature amount from each detected saliency area, and registers it in the video database 111 together with the frame information of the detection source and the coordinate information of the saliency area. The image feature amount extraction method may be changed according to the type of the salient region, but the image feature amount must be extracted by the same method for the salient region of the same type. For example, the shape feature amount can be used for the face region, and the color feature amount can be used for the clothing region. For the face regions A and B detected in different frames, the color feature amount is used for A and B Cannot use shape features.

  The video database 111 is a database for storing video, frames, tracking information, dynamic objects, and saliency information. A similar image search can be performed for an item to which an image feature amount is given. The similar image search is a function of rearranging and outputting data in the order in which the query and the image feature amount are close to each other. For comparison of image feature amounts, for example, the Euclidean distance between vectors can be used. Access to the video database 111 includes registration processing from the frame registration unit 106, tracking information registration unit 108, and saliency registration unit 110, readout processing from the query parameter estimation unit 112, and query determination unit 115, and similar images. Occurs during search processing from the search unit 116. Details of the structure of the video database 111 will be described later with reference to FIG.

  The query parameter estimation unit 112 uses the tracking information and the saliency information stored in the video database 111 to estimate parameters for determining a query type suitable for each video shooting location. The estimated parameters are stored in the query parameter storage unit 113.

  The query parameter storage unit 113 holds parameters for determining a query type suitable for each video shooting location. Details of the structure of the query parameter storage unit 113 will be described later with reference to FIG.

  The query input unit 114 informs the video search device 104 of the user's operation given by the input device 102 when the user specifies an object to be searched for from the video stored in the video database 111.

  The query determination unit 115 determines one or more optimal queries for each video shooting location using the object specified by the user, tracking information thereof, and parameters read from the query parameter storage unit 113. The query is an image feature amount of the area of the dynamic object detected by the dynamic object tracking unit 107 or an image feature amount of the saliency area detected by the saliency area detection unit 109.

  The similar image search unit 116 performs a similar image search on the video database 111 using the feature amounts of one or more query images selected by the query determination unit 115. When the types of salient areas of the query are different, search results with different scales are obtained. Therefore, for example, the similar image search unit 116 normalizes the similarity, integrates the search results, and outputs the result to the display device 103.

  FIG. 1B is a hardware configuration diagram of the video search system 100 according to the first embodiment of the present invention.

  The video search device 104 can be realized by a general computer, for example. For example, the video search device 104 may include a processor 121 and a storage device 122 that are connected to each other. The storage device 122 is configured by any type of storage medium. For example, the storage device 122 may be configured by a combination of a semiconductor memory and a hard disk drive.

  In this example, the video input unit 105, the frame registration unit 106, the dynamic object tracking unit 107, the tracking information registration unit 108, the saliency area detection unit 109, the saliency area registration unit 110, the query parameter estimation unit 112, which are illustrated in FIG. Functional units such as the query input unit 114, the query determination unit 115, and the similar image search unit 116 are realized by the processor 121 executing the processing program 123 stored in the storage device 122. In other words, in this example, the processing executed by each functional unit described above is actually executed by the processor 121 based on the processing program 123 described above. In addition, the video database 111 and the query parameter storage unit 113 are included in the storage device 122.

  The video search device 104 further includes a network interface device (NIF) 124 connected to the processor, and the video storage device 101 may be NAS or SAN connected to the video search device 104 via the network interface device. Good. Alternatively, the video storage device 101 may be included in the storage device 122.

  FIG. 2 is a diagram illustrating a configuration and a data example of the video database 111 according to the first embodiment of the present invention. Here, a configuration example of the table format is shown, but the data format may be arbitrary.

  The video database 111 includes a video table 200, a frame table 210, a tracking information table 220, a dynamic object table 230, and a saliency area table 240. The table configuration of FIG. 2 and the field configuration of each table are configurations necessary for implementing the present invention, and tables and fields may be added according to the application.

  The video table 200 has a video ID field 201, a file name field 202, and a shooting location ID field 203. The video ID field 201 holds an identification number of each video data. The file name field 202 holds the file name of the video data read from the video storage device 101. When the video is input directly from the camera, the file name may be omitted. The shooting place ID field 203 holds the ID of the place where the fixed point is observed. The correspondence between the video data ID and the shooting location may be managed by an application or may be managed by adding a shooting location management table to the video database. When a fixed camera is used, the shooting location ID may be read as a camera ID. As in the example of FIG. 2, a plurality of video files may be registered for one shooting location. In this case, the plurality of video files include, for example, video data shot by different time zones by a single camera whose installation location and shooting direction are fixed.

  The frame table 210 includes a frame ID field 211, a video ID field 212, and an image data field 213. The frame ID field holds the identification number of each frame extracted from the video data. The video ID field 212 is a field that holds an identification number of a video from which a frame is extracted, and this identification number corresponds to a value held in the video ID field 201 managed by the video table 200. The image data field 213 is binary data of a still image of a frame, and holds data used when displaying a search result or the like on the display device 103.

  The tracking information table 220 includes a tracking ID field 221, a dynamic object ID list field 222, and a flow line feature amount field 223. The tracking ID field 221 holds an identification number used for tracking each dynamic object by the dynamic object tracking unit 107. The dynamic object ID list field 222 has a list of dynamic object IDs having the same tracking ID. The ID of the dynamic object is an identification number managed in a dynamic object table 230 described later. The flow line feature value field 223 holds a flow line feature value extracted from a time-series change in the coordinates of the dynamic object in the image. Since the image size varies depending on the video, the flow line feature amount is calculated from the normalized coordinates of the dynamic object.

  The dynamic object table 230 includes a dynamic object ID field 231, a tracking ID field 232, a frame ID field 233, a coordinate field 234, and a feature amount field 235. The dynamic object ID field 231 holds an identification number (that is, dynamic object ID) of each dynamic object detected by the dynamic object tracking unit 107. The tracking ID field 232 holds an identification number (that is, a tracking ID) used in the dynamic object tracking unit 107 to associate the same dynamic object between frames. This identification number corresponds to the identification number held in the tracking ID field 221 managed in the tracking information table 220. The frame ID field 233 holds the identification number of the frame in which the dynamic object is detected. This identification number corresponds to the identification number held in the frame ID field 211 managed by the frame table 210. The coordinate field 234 holds the coordinates in the image of the dynamic object. The coordinates are expressed, for example, in the form of “the horizontal coordinates of the upper left corner, the vertical coordinates of the upper left corner, the horizontal coordinates of the lower right corner, and the vertical coordinates of the lower right corner of the rectangle” of the circumscribed rectangle of the dynamic object. The feature quantity field 235 holds an image feature quantity extracted from the dynamic object image. The image feature amount is expressed by, for example, a fixed-length vector.

  The dynamic object ID does not identify the dynamic object itself, but identifies an image of the dynamic object. Therefore, when an image of the same dynamic object is included in a plurality of frames, a different (unique) dynamic object ID is given to each of the images. For example, as shown in FIG. 2, when dynamic object IDs 1, 2, and 3 are held in the dynamic object ID list field 222 corresponding to the tracking ID: 1 of the tracking information table 220, at least the dynamic The images of the three dynamic objects identified by the object IDs: 1, 2, and 3 (these are included in different frames, respectively) are images of the same dynamic object by the dynamic object tracking unit 107. It means that it was determined.

  The saliency area table 240 includes a saliency area ID field 241, a frame ID field 242, a coordinate field 243, and a feature field 244. The saliency area ID field 241 holds the identification number of each saliency area detected by the saliency area detection unit 109. The frame ID field 242 holds the identification number of the frame in which the saliency area is detected. This identification number corresponds to the identification number held in the frame ID field 211 managed by the frame table 210. The coordinate field 243 holds the coordinates in the image of the saliency area. The feature quantity field 244 holds the image feature quantity extracted from the saliency area. The saliency area table 240 is prepared for the saliency area types determined by the system designer. It is also possible to perform a search using only the image feature amount of a dynamic object without preparing a saliency area table.

  <Operation of each part>

  The overall configuration of the video search system 100 has been described above. In the following, the operational principle of the video search system 100 is outlined, and the detailed operation of each functional unit is described.

  FIG. 3 is a diagram for explaining the operation of the video search system 100 according to the first embodiment of the present invention.

  When searching for an object in the video, the video search system 100 displays, for example, one frame of the video and performs a similar image search using the object shown in the frame as a query. An explanatory diagram 301 in FIG. 3 shows a state in which the user selects the search target 302 in the input frame. The search target 302 is a dynamic object shown in a plurality of frames including the illustrated input frame in the video. One of the images (that is, still images) of the search target 302 included in the plurality of frames is selected as a search query by the process described later.

  An arrow attached to the search target 302 represents the direction of the object (for example, the direction of the front of the body when the search target 302 is a person). In general, when the direction of an object is different, the image feature amount changes. Further, depending on the frame to be selected, there may be a case where an area characterizing the search target 302 (a salient area) is not shown in the first place. For example, in the case of a person, if the face is facing backward, the face area that is the feature is not captured, and therefore, the search using the face feature cannot be performed. Searching for an appropriate query for similar image search from a plurality of frames in the video is very laborious, which increases the time required to obtain a search result and decreases the search accuracy.

  In the present invention, by using tracking information of a dynamic object, one or more appropriate queries are determined from the same flow line, and a similar image search using the query is performed. The explanatory diagram 303 represents the tracking information of the search target 302 in the form of a route (flow line) moved in the image. Specifically, the curve displayed in the explanatory diagram 303 is obtained by concatenating the coordinates on the screen of the search target 302 in the continuous frames extracted from the video in order of the time at which each frame was captured, The direction of the arrow at the end of the curve indicates the direction in which the search target 302 has moved, and the contour in the explanatory diagram 303 corresponds to the contour of each frame. In addition, the flow line shown in the explanatory diagram 303 is a moving path of the search target 302 on the screen shot so as to look down at the shooting location from an obliquely upward direction. For this reason, the lower side of the screen corresponds to the near side (that is, the side close to the camera in the shooting range at the shooting location), and the upper side of the screen corresponds to the back side (that is, the side far from the camera in the shooting range of the shooting location). Also in the following explanatory diagrams, unless otherwise specified, the flow line of the dynamic object is displayed in the same manner as described above.

  When a certain flow line is a moving path of an image of a certain dynamic object, in the following description, the dynamic object is referred to as “dynamic object on the flow line”, and the dynamic object image is referred to as “on the flow line. In some cases, the “image” and the saliency area of the image on the flow line are described as “a saliency area on the flow line”.

  In the explanatory diagram 303, for example, the images of the search target 302 at the points A, B, C, and D on the flow line are taken from different directions as shown in the explanatory diagram 304. Since the appearances of the images differ from each other, different image feature amounts can be obtained. The plurality of images thus obtained are all query candidates that can be used as similar image search queries, and similar image searches may be performed using all of the obtained query candidates. However, since a large number of query candidates are obtained from the continuous frames as described above, the search time is increased by using the large number of query candidates. Also, there remains a problem with a search result integration method (for example, a method for determining the order in which search results are displayed).

  On the other hand, in the present invention, an appropriate query is automatically determined using the flow line information accumulated for each photographing location, and the number of searches is reduced. For example, the explanatory diagram 305 shows a flow line of each dynamic object shown in a video imaged at a certain place (denoted as place 1 in FIG. 3). In this example, many dynamic objects move from the front to the back on the screen, as indicated by the arrow of the flow line, so there are many images of the front of the dynamic object in the video taken at this shooting location. It is thought that it is not included. For this reason, it is difficult to perform a similar image search using, as a query, a feature that appears in front of an object such as a feature of a person's face, for example, for a video shot at such a place.

  The input frame shown in the explanatory diagram 301 includes an image of the search target 302 facing a certain direction. This image does not necessarily include a feature suitable as a query for searching for a video shot at the place 1. However, as shown in the explanatory diagrams 303 and 304, the frames before and after that include images of the search target 302 that face in various directions, and some of them search for the video shot at the location 1. May contain appropriate features as a query. Specifically, if the continuous frame including the input frame includes the image of the search target 302 that moves from the front to the back on the screen, like many dynamic objects in the place 1, There is a high possibility that the image includes a feature suitable as a query for searching for a video shot at the location 1.

  Therefore, the video search system 100 searches for the moment when the search target 302 moves toward the back of the screen from the tracking information of the input video, as with many dynamic objects at the location 1, and the frame of the frame taken at that time is searched. A similar image search is performed using features of a salient region other than the front of the search target 302 extracted from the image. For example, when the search target 302 is a person, a search is performed using the clothes color feature 306 on the back instead of the face feature on the front as a query. On the other hand, in the shooting location 2 shown in the explanatory diagram 307, not only the image of the dynamic object moving from the front side to the back on the screen but also the image of the dynamic object moving from the back side to the front on the screen is displayed. Since there is a high possibility that a face is captured in the latter, a search using the facial feature 308 as a query is performed.

  As an effect of the present embodiment, since the number of queries is automatically increased using tracking information, the work cost is reduced, and the search time can be reduced by selecting a query according to the shooting location. In addition, since only a query representing a prominent region is selected for each shooting location, an effect of reducing search noise can be expected as compared with a case where a search is performed using all query candidates.

  In order to implement the present invention, it is necessary to first track a dynamic object and detect a saliency area in the video accumulation stage and register it in a database. Further, after a large number of videos are accumulated, it is necessary to derive parameters for generating an appropriate query for each shooting location. At the time of search, one or more queries are generated and searched using these registration information and accumulated information. In the following, the operation of each unit in video registration, parameter derivation, and search will be described.

  FIG. 4 is a flowchart for explaining processing in which the video search apparatus 104 according to the first embodiment of the present invention registers the input video. Hereinafter, each step of FIG. 4 will be described.

(FIG. 4: Step S401)
The video input unit 105 decodes video data input from the video storage device 101 and extracts a frame as a still image.

(FIG. 4: Steps S402 to S410)
Each unit in the video search device 104 executes steps S402 to S410 for each frame extracted in step S401.

(FIG. 4: Step S403)
The frame registration unit 106 registers the frame and extraction source video information in the video database 111.

(FIG. 4: Step S404)
The dynamic object tracking unit 107 detects a dynamic object from the frame.

(FIG. 4: Step S405)
The dynamic object tracking unit 107 determines whether or not the dynamic object detected in step S404 is also present in the previous frame, and sets it to the previous frame (the frame at the time immediately before the current frame). If it also exists, the tracking information registration unit 108 executes step S407. On the other hand, if the dynamic object detected in step S404 does not exist in the previous frame, the dynamic information is a new dynamic object that has appeared in the current frame, and the tracking information registration unit 108 determines in step S406. Execute.

(FIG. 4: Step S406)
The tracking information registration unit 108 newly registers the dynamic object newly detected in step S405 as a tracking target in the tracking information table 220 of the video database 111.

(FIG. 4: Step S407)
The tracking information registration unit 108 extracts the image feature amount from each dynamic object, extracts the extracted image feature amount, the same tracking ID as the dynamic object of the previous frame specified in step S405, the frame ID of the current frame, and The coordinates of each dynamic object in the current frame are registered in the feature quantity field 235, tracking ID field 232, frame ID field 233, and coordinate field 234 of the dynamic object table 230, respectively. In addition, the tracking information registration unit 108 adds the obtained dynamic object ID to the dynamic object ID list field 222 of the tracking information table 220.

(FIG. 4: Step S408)
The saliency detection unit 109 detects a saliency area from the frame. When a plurality of types of saliency detection modules are prepared, detection processing is performed for the number of detection modules.

(FIG. 4: Step S409)
The saliency area registration unit 110 extracts an image feature amount from the saliency area detected in step S <b> 408 and registers it in the saliency area table 240 of the video database 111.

  Since steps S404 to S407 and steps S408 to S409 are independent processes, they may be executed in parallel using a plurality of calculation resources.

  This completes the description of the video registration process. Next, processing for estimating parameters used for determining an appropriate query using registered data will be described.

  FIG. 5 is a diagram for explaining query parameter estimation processing executed by the video search apparatus 104 according to the first embodiment of the present invention.

  When a certain number of videos or more are accumulated in the video database 111, a large number of flow lines can be obtained for each shooting location. In the explanatory diagram 501, the flow line obtained with respect to the place 2 similar to the explanatory diagram 307 in FIG. 3 is shown as an example. For each flow line, a flow line feature amount is stored in the tracking information table 220. The query parameter estimation unit 112 first performs clustering processing on these flow line feature quantities to find representative flow lines 502A and 502B indicated by thick arrows in the explanatory diagram 502. A general technique such as the k-means method can be used for the clustering process.

  Next, the query parameter estimation unit 112 acquires the saliency area detected on the flow line belonging to each cluster from the video database 111. As a result, for each type of saliency area, the number of detected saliency areas and a set of image feature quantities of the detected saliency areas are obtained. The saliency areas of the type in which the number detected at this stage is less than the predetermined number are excluded, and the most suitable saliency area is selected from the remaining saliency areas.

  As a method for determining a saliency area suitable for search, for example, a method using a dispersion ratio of image feature values is conceivable.

  FIG. 6 is an explanatory diagram of a dispersion ratio of image feature values used by the video search apparatus 104 according to the first embodiment of the present invention to determine a saliency area.

  The dispersion ratio of the image feature amount is a ratio between the dispersion value of the image feature amount in the salient region detected within the same flow line (intra-flow line dispersion) and the dispersion value between the flow lines (inter-flow line dispersion). (Dispersion ratio = distribution between flow lines / average dispersion within flow lines). An explanatory diagram 601 schematically shows an example of the dispersion of the image feature amount in the salient region of each flow line when the dispersion ratio is large. In this example, the time variation of the image feature amount in the same flow line, that is, the same object is small, and the difference in the image feature amount between the flow lines, that is, between different objects is large. Easy to find.

  On the other hand, an explanatory diagram 602 schematically shows an example of the dispersion of the image feature amount in the saliency area of each flow line when the dispersion ratio is small. In this example, the feature amount space of one object and the feature amount space of a different object cannot be separated, so there is a high possibility that an object that is different from the object to be originally searched is erroneously found, and an effective search result. Hard to get.

  The query parameter estimation unit 112 obtains a variance ratio of the image feature amount for each saliency area, selects a saliency area having a high variance ratio, and registers it in the query parameter accumulation unit 113.

  In the example of FIG. 5, as shown in the explanatory diagrams 501 and 502, the acquired plurality of flow lines includes a cluster including a plurality of flow lines from the front side to the left back side and a plurality of flow lines from the left back side to the front side. And a cluster including the flow line. In this example, the representative flow lines 502A and 502B of each cluster are not one of a plurality of actually acquired flow lines, but representative ones generated from a plurality of flow lines included in each cluster. It is a flow line. In addition, the flow line included in each cluster is also referred to as a similar flow line of the representative flow line of each cluster.

  Explanation diagrams 503 and 505 in FIG. 5 show examples of information related to the saliency areas detected on the similar flow lines of the representative flow line 502A and the representative flow line 502B, respectively. Specifically, in the explanatory diagrams 503 and 505, as examples of information regarding the saliency area, the saliency area type, the number of saliency areas detected for each type, an example of the saliency area image, and the distribution of the feature amount The ratio is displayed. In the example of FIG. 5, since each dynamic object is a person, the types of saliency areas include “moving object” (that is, the entire dynamic object), “face”, and “clothing color”, but other types may also be included. Good.

  In the example of FIG. 5, since the similar flow line of the representative flow line 502A includes many flow lines from the front side to the left back side on the screen, a large number of saliency areas whose types are “moving object” and “clothing color” are detected. However, a saliency area whose type is “face” is not detected. In this example, since the variance ratio of the image feature quantity of “clothing color” is larger than that of “moving object”, “clothing color” is also described as a saliency area type 504 suitable for search (hereinafter referred to as “effective saliency area type”). Selected). On the other hand, since the similar flow line of the representative flow line 502A includes many flow lines from the left back to the front on the screen, a sufficient number of salient regions whose type is “face” is detected, and the image feature amount Since the variance ratio is the largest, “face” is selected as the effective area type 506 in which the face is effective.

  More specifically, a type that satisfies a predetermined condition may be selected, for example, the number of detections and the variance value both exceed a predetermined value. When a plurality of types satisfy one of the above conditions for one cluster, all of them may be selected, or the types are further narrowed down according to other conditions, such as selecting one with the maximum variance value. Also good.

  FIG. 7 is a diagram illustrating a configuration and a data example of the query parameter storage unit 113 according to the first embodiment of the present invention. Here, a configuration example of the table format is shown, but the data format may be arbitrary.

  The query parameter storage unit 113 can be expressed by a table structure having a parameter ID field 700, a shooting location ID field 701, an area coordinate field 702, a representative flow line feature quantity field 703, and a salient area type field 704.

  The parameter ID field 700 holds an identification number (that is, parameter ID) of each parameter. This is an ID given to each cluster of flow lines described above.

  The shooting location ID field 701 holds an identification number (that is, shooting location ID) of each shooting location. The shooting location ID corresponds to the value held in the shooting location ID field 203 of the video table 200 in the video database 111. The area coordinate field 702 holds coordinates representing the distribution range of flow lines belonging to the flow line cluster. The representative flow line feature quantity field 703 holds an average feature quantity of flow line clusters (that is, an average of flow line feature quantities of flow lines belonging to the flow line cluster). The salient region type field 704 holds one or more effective salient region types selected by the method described above with reference to FIGS. 5 and 6.

  FIG. 8 is a flowchart for explaining processing in which the video search device 104 according to the first embodiment of the present invention estimates query parameters from accumulated data. Hereinafter, each step of FIG. 8 will be described.

(FIG. 8: Steps S801 to S809)
The query parameter estimation unit 112 executes steps S801 to S809 with each shooting location as a processing target.

(FIG. 8: Step S802)
The query parameter estimation unit 112 acquires tracking information extracted from the video of the shooting location to be processed from the video database 111. Thereby, for example, information on the flow line as shown in the explanatory diagram 501 of FIG. 5 is acquired.

(FIG. 8: Step S803)
The query parameter estimation unit 112 clusters the tracking information acquired in step S802 based on the flow line feature amount. Thus, for example, as shown in FIG. 5, a plurality of flow lines are classified into two clusters, and representative flow lines 502A and 502B representing the respective clusters are obtained.

(FIG. 8: Steps S804 to S808)
The query parameter estimation unit 112 executes steps S804 to S808 with each cluster obtained in step S803 as a processing target.

(FIG. 8: Step S805)
The query parameter estimation unit 112 acquires a salient region on the flow line from the tracking information belonging to the processing target cluster. For example, the query parameter estimation unit 112 may determine the coordinates of the dynamic object corresponding to a certain tracking ID and a certain frame ID (that is, the value held in the coordinate field 234) and the coordinates of the salient region corresponding to the same frame ID ( That is, when the overlapping ratio with the value of the coordinate field 243 is equal to or greater than a predetermined value, it is determined that the saliency area is a saliency area on the flow line identified by the tracking ID. The superimposition rate is, for example, the ratio of the size of the overlapping portion between the range and the coordinate range of the dynamic object to the size of the coordinate range of the saliency area. By collecting the saliency areas on each flow line obtained in this way for the cluster to be processed, for example, a saliency area as shown in the explanatory diagram 503 or 505 in FIG. 5 is acquired.

(FIG. 8: Step S806)
The query parameter estimation unit 112 derives the number of detections and the variance of the feature amount for each saliency area type, and estimates the effective saliency type by the method described with reference to FIGS. 5 and 6. Thus, for example, the saliency area type 504 or 506 shown in FIG. 5 is acquired.

(FIG. 8: Step S807)
The query parameter estimation unit 112 registers the parameter obtained in step S806 in the query parameter storage unit 113.

  The above is the description regarding the pre-processing for improving the efficiency of the similar image search using the tracking information of the dynamic object. Below, the search process of this invention is demonstrated.

  FIG. 9 is a diagram for explaining an operation in which the query determination unit 115 according to the first embodiment of the present invention determines a search query using tracking information, and is a diagram for explaining the conceptual diagram of FIG. 3 in more detail. .

  When the user designates a dynamic object to be searched (for example, search target 302 in FIG. 3), as shown in an explanatory diagram 901, flow line information of the object is obtained. For example, the information regarding the flow line similar to the explanatory diagram 303 of FIG. 3 is obtained. Next, the query determination unit 115 divides the obtained flow line into one or more partial flow lines. In the example of FIG. 9, partial flow lines 901a to 901e are obtained by the division.

  The flow line segmentation is performed so that all (or most) images on each partial flow line are images obtained by photographing one dynamic object from substantially the same direction (in other words, one dynamic object). It is desirable that the coordinates of a plurality of images obtained by photographing an object from substantially the same direction be connected in the order of photographing times to form one partial flow line). Specifically, for example, paying attention to the shooting time of each image on the flow line, the flow line may be divided at a predetermined time interval, or using the change in the direction of the flow line (for example, one The flow line may be divided so that the traveling direction of the flow line at each point in the partial flow line is included in a predetermined range. The query determination unit 115 extracts the flow line feature amount from each of the partial flow lines 901a to 901e included in the partial flow line set 902 obtained in this way, and makes it searchable.

  Next, the query determination unit 115 performs a nearest neighbor flow line search 903 for a set of partial flow lines using each representative flow line accumulated in the query parameter accumulation unit 113 as a query. The nearest-neighbor flow line search is a process of finding an element having the smallest distance between the feature vector and the query from the set.

  For example, when a search using a query image is performed for a video shot at location 2 in the explanatory diagram 307, in the nearest flow line search 903, each representative flow line 502A and 502B becomes a flow line query. A partial flow line having the smallest distance from the flow line feature vector is searched. In the example of FIG. 9, partial flow lines 901a and 901d are obtained by the nearest flow line search 903 using the representative flow lines 502A and 502B as queries.

  A small distance between the flow line feature amount of the representative flow line 502A and the flow line feature amount of the partial flow line 901a means that the similar flow line of the representative flow line 502A and the partial flow line 901a are similar. . In the example of FIG. 9, both the representative flow line 502A and the partial flow line 901a correspond to the movement of the dynamic object moving from the front of the screen toward the back.

  For this reason, the image of the dynamic object on the similar flow line of the representative flow line 502A and the image of the dynamic object on the partial flow line 901a may be images obtained by capturing the respective dynamic objects from substantially the same direction. High nature. This means that the saliency area of the same type as that of the effective saliency area related to the former is likely to be included in the latter. The relationship between the representative flow line 502B and the partial flow line 901d is the same.

  In the example of FIG. 9, as described above, the partial flow line 901a for the representative flow line 502A in which the clothing color saliency area is effective, and the partial flow line 901d for the representative flow line 502B in which the saliency area of the face is effective. Is selected. In this case, as shown in the explanatory diagram 904, the query determination unit 115 is extracted from the image feature amount of clothes color extracted from the image on the partial flow line 901a and the image on the partial flow line 901d as a search query. The face image feature amount is determined as a search query.

  When there are a plurality of saliency areas in the partial flow line (for example, when a plurality of saliency areas are included in a plurality of frames constituting the partial flow line), the query determination unit 115 selects any of them. Then, it may be determined as a search query, but a more prominent region more suitable as a search query may be selected based on other conditions. For example, the query determination unit 115 may select a place where the size of the saliency area is large or a place where the speed of the dynamic object is slow (to reduce subject blurring) and determine it as a search query. Further, if the saliency area detection module has a function of outputting the reliability of the detection result, the value may be used to determine, for example, the image feature amount of the saliency area with high reliability as the search query.

  FIG. 10 is a flowchart illustrating a process in which the video search apparatus 104 according to the first embodiment of the present invention performs a similar image search using a search query determined using tracking information. Hereinafter, each step of FIG. 10 will be described.

(FIG. 10: Step S1001)
The query determination unit 115 reads tracking information of the search target 302 specified by the user through the query input unit 114 from the video database 111. As a result, for example, flow line information as shown in the explanatory diagram 901 of FIG. 9 is read.

(FIG. 10: Step S1002)
The query determination unit 115 generates a partial flow line set based on the tracking information obtained in step S1001, and extracts the flow line feature quantity of each partial flow line. Thereby, for example, a set of partial flow lines 902 shown in FIG. 9 is obtained.

(FIG. 10: Step S1003)
The query determination unit 115 reads the parameters of each representative flow line at each shooting location from the query parameter storage unit 113. Thereby, for example, the parameters of the representative flow lines 502A and 502B shown in FIG. 9 are read.

(FIG. 10: Steps S1004 to S1008)
The query determination unit 115 executes steps S1004 to S1008 for each representative flow line parameter read in step S1003.

(FIG. 10: Step S1005)
The query determination unit 115 searches for the nearest flow line from the partial flow line set using the feature quantity of the representative flow line as a query. This procedure corresponds to the nearest flow line search 903 in FIG.

(FIG. 10: Step S1006)
The query determination unit 115 selects the saliency area on the nearest flow line obtained in step S1005, and uses the parameters of the representative flow line read in step S1003 as the search query including the image feature amount of the saliency area. It is determined as a search query for the specified shooting location and area. Accordingly, for example, as shown in an explanatory diagram 904 of FIG. 9, a search query including an image feature quantity of clothes color is represented for the representative flow line 502A, and a search query including an image feature quantity of the face is represented for the representative flow line 502B. It is determined.

(FIG. 10: Step S1007)
The similar image search unit 116 acquires a similar image search result from the video database 111 using the search query determined in step S1006. A general similar image search technique can be used for this processing.

(FIG. 10: Step S1009)
When the execution of steps S1004 to S1008 is completed for the parameters of each representative flow line, the similar image search unit 116 integrates and displays the search results for each shooting location and each representative flow line obtained in steps S1004 to S1008. Display on the device 103. Since each search result is a search result using different types of salient regions as queries, the similar image search unit 116 normalizes the similarity when integrating. The search results may be displayed separately for each shooting location.

  FIG. 11 is a diagram illustrating a processing sequence of the video search system 100 according to the first embodiment of the present invention. Specifically, the video registration process, the query parameter estimation process, and the search of the video search system 100 described above are illustrated. It is a figure explaining the processing sequence of the user 1101, the computer 1102, the video database 111, and the query parameter storage part 113 in a process. A computer 1102 is a computer that implements the video search device 104. In FIG. 11, the video database 111 and the query parameter storage unit 113 are displayed separately from the computer 1102 for the sake of explanation, but these may be included in the computer 1102. Steps S1132, S1133, and S1134 in FIG. 11 are processes related to video registration processing, query parameter estimation processing, and search processing, respectively. Hereinafter, each step of FIG. 11 will be described.

[Video Registration Process] (FIG. 11: Steps S1003 to S1112)
When the user 1101 inputs video from the video storage device 101 to the computer 1102 (S1103), in the computer 1102, the frame registration unit 106 registers the frame extracted by the video input unit 105 in the video database 111 (S1104). 111 notifies the completion of registration (S1105).

  Next, in the computer 1102, the dynamic object tracking unit 107 detects and tracks the dynamic object in the extracted frame (S1106), and the tracking information registration unit 108 registers the tracking information in the video database 111 (S1107). ), The video database 111 notifies registration completion (S1108). Further, the saliency area detection unit 109 detects the saliency area in the extracted frame (S1109), the saliency area registration unit 110 registers the saliency area in the video database 111 (S1110), and the video database 111 notifies the registration completion. (S1111). When all the frames have been processed, the video registration completion is notified to the user 1101 (S1112).

[Query Parameter Estimation Processing] (FIG. 11: Steps S1113 to S1119)
When the user 1101 issues a request for query parameter estimation processing to the video search device 104 (S1113), in the computer 1102, the query parameter estimation unit 112 requests tracking information for each shooting location from the video database 111 (S1114). (S1115).

  The query parameter estimation unit 112 derives a parameter necessary for determining a query by the method described above with reference to FIGS. 5 to 8 (S1116), registers the parameter in the query parameter storage unit 113 (S1117), and sets the query parameter. The storage unit 113 notifies registration completion (S1118). When the parameter estimation process is completed for all shooting locations, the user 1101 is notified of the completion of the process (S1119).

[Search Process] (FIG. 11: Steps S1120 to S1131)
When the user 1101 designates a search target dynamic object (for example, the search target 302) from the frames stored in the video database 111 (S1120), in the computer 1102, the query determination unit 115 tracks the search target dynamic object. Information is requested from the video database 111 (S1121) and acquired (S1122), and a parameter is requested from the query parameter storage unit (S1123) and acquired (S1124).

  The query determination unit 115 determines a query for each shooting location by the method described above with reference to FIGS. 9 to 10 using the tracking information of the dynamic object to be searched and the query parameters for each shooting location (S1125). ) To the user 1101 (S1126). When the user 1101 confirms the presented query and issues a search request (S1127), in the computer 1102, the similar image search unit 116 performs a similar image search using the determined query (S1128), and the video database. A similar image search result is acquired from 111 (S1129). The computer 1102 integrates search results obtained from a plurality of queries as necessary (S1130) and presents them to the user (S1131).

  FIG. 12 is a diagram illustrating a configuration example of an operation screen used when searching for an object in a video using the video search device 104 according to the first embodiment of the present invention. This screen is presented to the user on the display device 103. The user operates the cursor 1207 displayed on the screen using the input device 102 to give a processing instruction to the video search device 104.

  The operation screen of FIG. 12 includes a video selection button 1201, a video display area 1202, a query display area 1203, a search button 1204, and a search result display area 1205.

  First, the user clicks a video selection button 1201 to select an arbitrary video recorded in the video database 111. The selected video is displayed in the video display area 1202. The user designates an arbitrary dynamic object 302 displayed in the video display area 1202 as a search target.

  The video search apparatus 104 searches for an appropriate query on the flow line of the specified dynamic object by the method described above with reference to FIGS. 9 to 10, and displays it in the query display area 1203.

  The user confirms the displayed query, makes adjustments if necessary, and clicks a search button 1204 to issue a search request.

  The video search device 104 displays the similar image search result of each query in the search result display area 1205.

  The above is the description regarding the first embodiment of the present invention. According to the present embodiment, the user can easily specify the search target, and the similar image search corresponding to the shooting location can be executed. Further, by using only a query limited to each shooting location, the search time can be shortened and search noise can be reduced.

  In the first embodiment, the method for improving the efficiency of the similar image search using the tracking information of the dynamic object in the video has been described. Since tracking of a dynamic object is usually performed by using only information between adjacent frames, tracking is performed when the object is stationary for a long time or when there is an obstacle between the object and the camera. May be interrupted. When tracking for a long time is not possible, the number of query candidates obtained from each flow line is reduced, so that the effect of the first embodiment may not be sufficiently exhibited.

  Therefore, in the second embodiment, a method for correcting tracking information using a saliency area accumulated in the video database 111 will be described. Except for the differences described below, each part of the video search system 100 according to the second embodiment has the same function as each part denoted by the same reference numeral as the first embodiment shown in FIGS. Those explanations are omitted.

  FIG. 13 is a diagram for explaining the tracking information correction using the saliency area by the video search system 100 according to the second embodiment of the present invention.

  For example, an explanatory diagram 1301 shows three flow lines detected within a certain period in a certain video. The flow line 1 and the flow line 3 are flow lines of different persons. On the other hand, although the flow line 1 and the flow line 2 should originally be one flow line of the same person, the person has been divided in the middle because the person has passed the back side of the shield 1302. Therefore, as shown on the left side of FIG. 13, the tracking information table 220 and the dynamic object table 230 of the video database 111 before correction are given two flow lines of the same person with tracking IDs 1 and 2, respectively. , Recorded as a separate flow line.

  Therefore, the video search system 100 according to the present embodiment performs a similar image search using the saliency areas on each flow line as a query and the saliency areas on different flow lines within a predetermined time at the same shooting location. As a result, if significant areas with similarities of a predetermined value or more are found on different flow lines, it is determined that those flow lines are flow lines of the same object, and the tracking information is corrected.

  From the tracking information table 220 after correction shown on the right side of FIG. 13, the entry with tracking ID: 2 that existed before correction is deleted, and the contents of the dynamic object ID list of the entry with tracking ID: 2 are the tracking ID. : 1 is added to the entry, so that two flow lines are integrated into one. Correspondingly, the tracking ID: 2 of the dynamic object table 230 before correction is also changed to the tracking ID: 1, as shown in the corrected dynamic object table 230 shown on the right side of FIG. A portion surrounded by a thick frame in the table of FIG. 13 is a portion corrected as a result of this processing.

  FIG. 14 is a flowchart for describing processing in which the video search apparatus 104 according to the second embodiment of the present invention corrects tracking information using a saliency area. Hereinafter, each step of FIG. 14 will be described.

(FIG. 14: Steps S1401 to S1406)
The tracking information registration unit 108 executes steps S1401 to S1406 for each flow line detected within a predetermined time.

(FIG. 14: Step S1402)
The tracking information registration unit 108 reads the salient region on the flow line from the video database 111.

(FIG. 14: Step S1403)
The tracking information registration unit 108 performs a similar image search using the saliency area as a query.

(FIG. 14: Step S1404)
The tracking information registration unit 108 determines whether or not there is a saliency area whose similarity is equal to or higher than a predetermined value on different flow lines within a predetermined time, and if there is, executes step 1405.

(FIG. 14: Step S1405)
The tracking information registration unit 108 combines the flow lines and updates the tracking information table 220 and the dynamic object table 230 of the video database 111 accordingly.

  Since the method according to the first embodiment selects a query for each flow line, it is necessary to appropriately describe the feature amount of the flow line. However, the feature of the flow line obtained by tracking the dynamic object represents a two-dimensional motion in the video to the last, and depth information (in other words, the distance from the camera to the dynamic object) is not considered. Therefore, depending on the installation method of the camera, there is a possibility that the state of the saliency area on the flow line may change even if the flow line is the same.

  FIG. 15 is a diagram for explaining tracking information correction in consideration of depth information by the video search system 100 according to the second embodiment of the present invention.

  For example, in the explanatory diagrams 1501 and 1502 of FIG. 15, the flow lines of the dynamic objects (people in this example) included in the images taken at the place 1 and the place 2 and a plurality of images on the respective flow lines, respectively. An example of In this example, the three images of dynamic object IDs: 1 to 3 shown in the explanatory diagram 1501 are images obtained by photographing one dynamic object near the start point, the middle point, and the end point of the flow line, respectively. . The saliency areas 1501A to 1501C are saliency areas (faces in this example) of the images of the dynamic object IDs: 1 to 3, respectively. Similarly, the three images of dynamic object IDs 11 to 13 shown in the explanatory diagram 1502 are images obtained by photographing one dynamic object near the start point, the intermediate point, and the end point of the flow line, respectively. The saliency areas 1502A to 1502C are saliency areas of the images of the dynamic object IDs 11 to 13, respectively.

  Although the shapes of these two flow lines coincide, there is almost no change in the depth direction of the position of the dynamic object from the start point to the end point of the flow line shown in the explanatory diagram 1501 (in other words, the dynamic object The distance between the camera and the camera hardly changes), whereas the dynamic object in the explanatory diagram 1502 moves from the back toward the front. As shown in the size 1503 of the saliency area on the flow line, the saliency areas 1501A to 1501C have the same size (for example, 10 cm × 10 cm), whereas the saliency areas 1502A, B, and C have the same size. It can be seen from the change (for example, 5 cm × 5 cm, 7 cm × 7 cm, and 10 cm × 10 cm, respectively).

  This means that even if the shape of the moving path of the dynamic object is the same as the change in the coordinates on the screen shot at each location, the dynamic as the change in the three-dimensional coordinates in the space of each shooting location. This means that the actual movement path of the object is greatly different. In such a case, the appearance of the saliency area on each flow line may be greatly different, so in order to select a more appropriate search query, as a change in the three-dimensional coordinates in the space of the shooting location It is desirable to use a moving path of a dynamic object (that is, a flow line to which depth information is given).

  Therefore, in this embodiment, the tracking information registration unit 108 assigns depth information to the flow line using the size 1503 of the saliency area on the flow line and the prior knowledge 1504 for the saliency area. The prior knowledge 1504 includes the standard size of the saliency area for each type, for example, 25 cm × 25 cm when the saliency area type is a face, and the camera installation position (particularly height) at each shooting location, Information such as the installation direction (particularly the depression angle) and the focal length of the camera lens is included. Based on these pieces of information and the size of the saliency area actually captured, the distance from the camera of the dynamic object including the saliency area can be estimated.

  The dynamic object table 230 may further include a depth feature amount field 236 that holds depth information of the dynamic object. For example, when the size of the salient region on the flow line is 10 cm × 10 cm, the ratio “10/25” between the size and the standard size 25 cm × 25 cm is held in the depth feature quantity field 236. The tracking information registration unit 108 generates a flow line indicating the moving path of the dynamic object on the screen in the three-dimensional space of the shooting location based on the depth feature amount and the prior knowledge about the installation position of the camera. It converts into the flow line which shows a movement path | route, and calculates the feature-value of the flow line after conversion. The calculated feature amount is held in the flow line feature amount 224 in consideration of the depth of the tracking information table 220, for example.

  FIG. 16 is a flowchart for describing processing in which the video search apparatus 104 according to the second embodiment of the present invention adds depth information to tracking information using a saliency area. Hereinafter, each step of FIG. 16 will be described.

(FIG. 16: Steps S1601 to S1607)
The tracking information registration unit 108 executes steps S1601 to S1607 for each dynamic object on the flow line.

(FIG. 16: Step S1602)
The tracking information registration unit 108 checks whether or not there is a saliency area on the flow line by checking the overlapping rate between the coordinates of the detected saliency area and the coordinates of the dynamic object in the same manner as in step 805 in FIG. If there is a saliency area, step S1603 is executed. If not, step S1604 is executed.

(FIG. 16: Step S1603)
The tracking information registration unit 108 derives a depth feature from the prior knowledge 1504. The reliability of the depth feature can be increased by using a plurality of saliency areas.

(FIG. 16: Step S1604)
If no saliency area is detected, the tracking information registration unit 108 interpolates the depth feature from the depth information derived from the saliency areas of the adjacent frames before and after.

(FIG. 16: Step S1605)
The tracking information registration unit 108 adds the obtained depth information to the dynamic object table 230 of the video database 111.

(FIG. 16: Step S1607)
When the tracking information registration unit 108 obtains all the dynamic depth information on the flow line, the tracking information registration unit 108 extracts the flow line feature amount considering the depth and adds the extracted flow line feature amount to the tracking information table 220 of the video database 111.

  The method for determining the search query based on the flow line feature amount extracted by the above processing is the same as that in the first embodiment, and thus the description thereof is omitted. Thus, by using the corrected tracking information, it is possible to increase the search accuracy of the nearest flow line search 903 based on the flow line feature amount. This makes it possible to determine a search query that is more suitable for the video at each shooting location, and as a result, it is possible to improve the accuracy of the object search described in the first embodiment.

  In the first embodiment, the search accuracy is improved by performing a similar image search using a salient region in another frame on the same flow line. However, if there is no saliency area characterizing the search object on the flow line, the video obtained by the search may be limited. In the third embodiment, a method for notifying a user of a salient area of a search target included in a video different from that specified by the user will be described. Except for differences described below, each part of the video search system 100 according to the third embodiment has the same function as each part denoted by the same reference numeral in the first embodiment shown in FIGS. Those explanations are omitted.

  FIG. 17 is an explanatory diagram relating to presentation of queries existing in different videos according to the third embodiment of the present invention.

  For example, an explanatory diagram 1701 shows a flow line of a search target person extracted from a video photographed at a place 1. On the flow line, the face of the person to be searched and the salient area of the clothing color appear. On the other hand, the explanatory diagram 1702 shows the flow lines of the same person as the search target extracted from the video shot at the place 2 different from the place 1. Whether these flow lines belong to the same person is determined based on the image feature amounts of the salient areas 1701A and 1702A of the face on the respective flow lines. Further, from the image on the flow line in the explanatory diagram 1702, a saliency area 1702B of a characteristic possession (for example, a bag) of the person is found. If it is possible to notify the user of a noticeable area of such a different video (for example, a video shot at a different place or a video shot at the same place at different times), the user searches for more videos. be able to.

  For example, the explanatory diagram 1703 shows a flow line extracted from a video shot at a place 3 different from both the place 1 and the place 2. In this example, the flow line is the flow line of the same person as shown in the explanatory diagrams 1701 and 1702, but neither the face nor the clothing color is detected as a prominent region on the flow line, and the bag is prominent. A region 1703A is detected. In this case, even if the image feature amount of the salient region of the face or clothing color extracted from the video of the place 1 is used as a search query, the person to be searched cannot be searched from the video of the place 3, but the bag is prominent. If the image feature amount in the area 1702B is used as a search query, the person can be searched.

  Screens 1704 and 1705 are examples of screens displayed on the display device 103 in order to notify the user of salient areas existing in different videos. The screen 1704 displays a frame of a video in which the user has selected a search target person. The display device 103 may further display a saliency area detected from another video by the above method in a pop-up. In the example of the screen 1704, a bag that is the property of the person to be searched, extracted from a video captured by another camera (external camera 2), and a video captured by another camera (external camera 4). The hats belonging to the person to be searched, detected from the above, are displayed by pop-ups 1704A and 1704B, respectively.

  On the other hand, on the screen 1705, the relationship between the saliency areas detected from different images is displayed in a graph. In the example of the screen 1705, a node 1705A representing the location 1, a node 1705B representing the salient area of the face, and a node 1705C representing the salient area of the clothing color are displayed, and the nodes 1705A and 1705B are joined by an edge. And 1705C are also connected by an edge. This indicates that a salient area (for example, salient area 1701A) of the face and a salient area of clothing color are detected from the image of the person to be searched, which is taken at location 1.

  Further, the screen 1705 displays nodes 1705D, 1705E, 1705F, and 1705G representing the location 2, the salient area of the face, the salient area of the clothing color, and the salient area of the bag, respectively. The node 1705D includes nodes 1705E, 1705F, and 1705G. Are connected with each of the edges. Further, the node 1705E is coupled to the node 1705B, and the node 1705F is coupled to the node 1705C at the edges. These are a saliency area (for example, saliency area 1702A) of a face on a certain flow line extracted from the video of place 2 and a saliency area for clothing color, respectively. ) And clothes color salient areas, indicating that a salient area of the bag (for example, salient area 1702B) is further detected as a salient area on the flow line.

  Further, the screen 1705 displays nodes 1705H, 1705I, and 1705J representing the location 4, the salient area of the face, and the salient area of the hat, respectively, and the node 1705H is coupled to each of the nodes 1705I and 1705J by edges. Further, the node 1705I is coupled to the node 1705B at the edge. In these, the salient area of the face on a certain flow line extracted from the video of the place 4 is similar to the salient area (for example, the salient area 1701A) of the face to be searched for the place 1, and the salient area on the flow line is This indicates that a more prominent region of the hat has been detected.

  The user can designate a new salient area used for the search query with reference to the above display. For example, when the user designates the pop-up 1704A or the node 1705G using the input device 102, a similar image search is executed using the image feature amount of the salient area of the bag as a search query. Thereby, an image including the remarkable area 1703A of the bag at the place 3 can be acquired as a search result. For example, in the video taken at the place 3, the face of the person to be searched and the clothes color are not shown to the extent that they can be searched, but when the bag is shown, the face or clothes color detected at the place 1 In the similar image search using the image feature amount, the image of the person cannot be acquired from the video at the place 3. However, as described above, by using the image feature amount of the bag acquired at the location 2 as a search query, it is possible to acquire the image of the person from the video at the location 3.

  FIG. 18 is a flowchart for explaining processing in which the video search apparatus 104 according to the third embodiment of the present invention searches for a new type of saliency area from different videos. Hereinafter, each step of FIG. 18 will be described.

(FIG. 18: Step S1801)
The query determination unit 115 selects a query for each shooting location from dynamic objects specified by the user. This process is the same as the process up to step S1006 in FIG.

(FIG. 18: Steps S1802 to S1805)
The query determination unit 115 executes steps S1802 to S1805 for the query selected for each shooting location.

(FIG. 18: Step S1803)
The similar image search unit 116 performs a similar image search for the designated shooting location using the selected query.

(FIG. 18: Step S1804)
The similar image search unit 116 notifies the user by a display method such as a screen 1704 or 1705 in FIG. 17 when a new type of remarkable area is found on the flow line to which the search result belongs. When the user designates any saliency area based on this notification, the similar image search unit 116 executes step S1007 in FIG. 10 using a search query including the image feature amount of the designated saliency area.

  In the above embodiment, the use of searching for an object designated by the user has been described. On the other hand, there are cases where the user does not assume a specific search target and wants to efficiently grasp all objects that have appeared within a predetermined period. In the fourth embodiment, a method for summarizing and displaying a long-time video will be described. Except for differences described below, each part of the video search system 100 according to the fourth embodiment has the same function as each part denoted by the same reference numeral in the first embodiment shown in FIGS. Those explanations are omitted.

  FIG. 19 is a diagram for explaining video summarization using tracking information according to the fourth embodiment of the present invention.

  Since the video database 111 holds information on dynamic objects detected in each frame, for example, a graph 1901 is generated with time (frame number) on the horizontal axis and the number of detected dynamic objects on the vertical axis. be able to. When the user operates the cursor 1207 using the input device 102 to select, for example, a time zone 1905 in which many dynamic objects exist, all the dynamic objects detected in the time zone 1905 are superimposed on the frame. Is displayed. However, in this state, a large number of dynamic objects are mixed and visibility is poor. An explanatory diagram 1902 is an example of a screen displayed by the display device 103. In this example, images on the flow lines of four persons are displayed in one frame, but since many images are displayed for each person, the screen is crowded and visibility is reduced.

  Therefore, the video search system 100 according to the present embodiment displays only one dynamic object image for each flow line using the tracking information of the video database 111. When the dynamic objects overlap, the object image to be superimposed is moved on the flow line and adjusted so that the objects do not overlap each other. An explanatory diagram 1903 is an example of a screen displayed by the display device 103 of this embodiment. In this example, a flow line 1903A of a certain person is displayed, and only one image 1903B is displayed among a plurality of images of the person on the flow line 1903A. Similarly, for each person, a flow line and one image on the flow line are displayed, and the image on the flow line is displayed so as not to be superimposed on an image of another person already displayed. This eliminates screen congestion and improves visibility.

  In addition, by using the query determination method described in the first embodiment and highlighting the saliency area that becomes a query for each dynamic object, each object can be grasped more efficiently. An explanatory diagram 1903 is another example of a screen displayed by the display device 103 of this embodiment. In this example, for a certain person, in addition to the flow line 1903A and the image 1903B on the flow line, a saliency area 1904A on the flow line is displayed in a pop-up manner. The same applies to other persons.

  Note that not all the images on one flow line include the image of the saliency area. The video search system 100 according to the present embodiment may preferentially select an image including a saliency area when selecting one to be displayed from among a plurality of images of each person.

  FIG. 20 is a flowchart showing video summary processing using tracking information executed by the video search system 100 according to the fourth embodiment of the present invention. Hereinafter, each step of FIG. 20 will be described.

(FIG. 20: Step S2001)
The query determination unit 115 reads out all the flow line information within the shooting location and time designated by the user.

(FIG. 20: Steps S2002 to S2008)
The query determination unit 115 executes steps S2002 to S2008 for each flow line obtained in step S2001.

(FIG. 20: Step S2003)
The query determination unit 115 searches for a saliency area suitable for the query on the flow line. This process is the same as the process described in FIG.

(FIG. 20: Step S2004)
The query determination unit 115 reads the coordinates of the dynamic object in the frame where the saliency area exists from the video database 111.

(FIG. 20: Step S2005)
The query determination unit 115 determines whether or not the coordinate range of the dynamic object read in step S2004 overlaps the coordinate range of the displayed dynamic object, and if so, executes step S2006, If not, step S2007 is executed.

(FIG. 20: Step S2006)
The query determination unit 115 moves the coordinates of the dynamic object on the flow line, and returns to step S2005.

(FIG. 20: Step S2007)
The video search device 104 superimposes an image of a dynamic object on the flow line and displays it on the display device 103.

  With the above processing, it becomes possible for the user to efficiently grasp the dynamic object that has appeared within the specified time and the saliency area by using the tracking information of the dynamic object and the saliency area detection.

  In addition, this invention is not limited to the Example mentioned above, Various modifications are included. For example, the above-described embodiments have been described in detail for easy understanding of the present invention, and are not necessarily limited to those having all the configurations described. Further, a part of the configuration of one embodiment can be replaced with the configuration of another embodiment, and the configuration of another embodiment can be added to the configuration of one embodiment. Further, it is possible to add, delete, and replace other configurations for a part of the configuration of each embodiment.

  Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. Further, each of the above-described configurations, functions, and the like may be realized by software by interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files for realizing each function is a memory, a hard disk drive, a storage device such as an SSD (Solid State Drive), or a computer-readable non-transitory data such as an IC card, an SD card, or a DVD. It can be stored in a storage medium.

  Further, the drawings show control lines and information lines that are considered necessary for explaining the embodiments, and not necessarily all control lines and information lines included in an actual product to which the present invention is applied. Not necessarily. Actually, it may be considered that almost all the components are connected to each other.

100: video search system 101: video storage device 102: input device 103: display device 104: video search device 105: video input unit 106: frame registration unit 107: dynamic object tracking unit 108: tracking information registration unit 109: salient area Detection unit 110: Striking area registration unit 111: Video database 112: Query parameter estimation unit 113: Query parameter storage unit 114: Query input unit 115: Query determination unit 116: Similar image search unit

Claims (15)

A video search device having a processor and a storage device connected to the processor,
One or more moving paths of one or more moving objects from each of the first video composed of a plurality of frames photographed at the first location and the second video composed of the plurality of frames photographed at the second location. Is detected and stored in the storage device,
Extracting the image feature amount for each frame of the selected moving body from the one or more moving bodies detected from the first video, and storing the extracted image feature amount in the storage device;
The extracted image feature amount based on the movement path of the selected moving body detected from the first video and the movement path of the one or more moving bodies detected from the second video. Select the query image feature to use as the search query,
Using the query image feature amount, search for the image feature amount of the one or more moving objects extracted from the second video,
A video search apparatus that outputs the search result. The video search device according to claim 1,
The second video includes a plurality of images of the moving body,
The video search device
The movement paths of the plurality of moving bodies detected from the second video are classified into a plurality of clusters based on the feature amount of each movement path, and a representative path that represents the movement path of each cluster is generated. ,
Generating a plurality of partial paths by dividing the movement path of the selected moving body;
Based on the feature amount of the representative route of the plurality of clusters and the feature amount of the plurality of partial routes, a partial route that is most similar to any one of the representative routes is searched from among the plurality of partial routes.
An image search apparatus, wherein an image feature amount of the selected moving body on the partial path obtained by the search is selected as the query image feature amount. The video search device according to claim 1,
The storage device stores information on a moving path of the first moving body and the second moving body extracted from the first video,
The video search device further includes an image of the first moving body detected from the first video based on an image feature amount of the first moving body and an image feature amount of the second moving body. When the image of the second moving body is determined to be similar to the image of the second moving body, the information about the moving path of the second moving body is integrated with the information about the moving path of the first moving body. Search device. The video search device according to claim 1,
Both the image of the selected moving body and the image of the moving body detected from the second video include a predetermined first type area, and the first type of the first type included in the selected image. The image of the area is similar to the image of the first type area included in the image of the moving object detected from the second video, and the moving object extracted from the second video is When the image further includes an area of the second type, the video search apparatus outputs information related to the area of the second type. The video search device according to claim 1,
A display device;
When images of a plurality of moving bodies are detected from any video and a plurality of images of the respective moving bodies are detected, one of the plurality of images of the respective moving bodies is selected, and each of the moving bodies is selected. The video search apparatus is characterized in that the selected image is displayed so as not to overlap with the selected image of another moving body. A video search method executed by a video search device having a processor and a storage device connected to the processor,
One or more moving paths of one or more moving objects from each of the first video composed of a plurality of frames photographed at the first location and the second video composed of the plurality of frames photographed at the second location. Detecting and storing in the storage device;
A second procedure of extracting an image feature amount for each frame of a selected moving body from the one or more moving bodies detected from the first video and storing the extracted image feature amount in the storage device;
The extracted image feature amount based on the movement path of the selected moving body detected from the first video and the movement path of the one or more moving bodies detected from the second video. A third procedure for selecting a query image feature value to be used as a search query,
A fourth procedure for searching for image feature quantities of the one or more moving objects extracted from the second video using the query image feature quantities;
And a fifth procedure for outputting a result of the search. The video search method according to claim 6,
The second video includes a plurality of images of the moving body,
The third procedure includes
The movement paths of the plurality of moving bodies detected from the second video are classified into a plurality of clusters based on the feature amounts of the respective movement paths, and representative paths that represent the movement paths of the respective clusters are generated. Procedure and
Generating a plurality of partial paths by dividing the movement path of the selected moving body;
A procedure for searching for a partial route that is most similar to any one of the representative routes out of the plurality of partial routes based on the feature amount of the representative route of the plurality of clusters and the feature amount of the plurality of partial routes. When,
And a procedure for selecting an image feature quantity of the selected moving object on the partial path obtained by the search as the query image feature quantity. The video search method according to claim 6,
The storage device stores information on a moving path of the first moving body and the second moving body extracted from the first video,
The video search method further includes an image of the first moving body detected from the first video based on an image feature amount of the first moving body and an image feature amount of the second moving body. And when the image of the second moving body is determined to be similar, the method includes a step of integrating information related to the moving path of the second moving body into information related to the moving path of the first moving body. Video search method. The video search method according to claim 6,
Both the image of the selected moving body and the image of the moving body detected from the second video include a predetermined first type area, and the first type of the first type included in the selected image. The image of the area is similar to the image of the first type area included in the image of the moving object detected from the second video, and the moving object extracted from the second video is A video search method, further comprising a step of outputting information on the second type area when the image further includes a second type area. The video search method according to claim 6,
When images of a plurality of moving bodies are detected from any video and a plurality of images of the respective moving bodies are detected, one of the plurality of images of the respective moving bodies is selected, and each of the moving bodies is selected. The video search method further includes a step of displaying the selected image so as not to overlap with the selected image of another moving body. A non-transitory computer-readable storage medium storing a program for controlling a computer,
The computer includes a processor and a storage device connected to the processor,
The program is
One or more moving paths of one or more moving objects from each of the first video composed of a plurality of frames photographed at the first location and the second video composed of the plurality of frames photographed at the second location. Detecting and storing in the storage device;
A second procedure of extracting an image feature amount for each frame of a selected moving body from the one or more moving bodies detected from the first video and storing the extracted image feature amount in the storage device;
The extracted image feature amount based on the movement path of the selected moving body detected from the first video and the movement path of the one or more moving bodies detected from the second video. A third procedure for selecting a query image feature value to be used as a search query,
A fourth procedure for searching for image feature quantities of the one or more moving objects extracted from the second video using the query image feature quantities;
A non-transitory computer-readable storage medium that causes the processor to execute a fifth procedure for outputting the search result. A non-transitory computer readable storage medium according to claim 11, comprising:
The second video includes a plurality of images of the moving body,
The third procedure includes
The movement paths of the plurality of moving bodies detected from the second video are classified into a plurality of clusters based on the feature amounts of the respective movement paths, and representative paths that represent the movement paths of the respective clusters are generated. Procedure and
Generating a plurality of partial paths by dividing the movement path of the selected moving body;
A procedure for searching for a partial route that is most similar to any one of the representative routes out of the plurality of partial routes based on the feature amount of the representative route of the plurality of clusters and the feature amount of the plurality of partial routes. When,
A non-transitory computer-readable storage medium, comprising: selecting an image feature quantity of the selected moving body on the partial path obtained by the search as the query image feature quantity. A non-transitory computer readable storage medium according to claim 11, comprising:
The storage device stores information on a moving path of the first moving body and the second moving body extracted from the first video,
The program further includes the image of the first moving body detected from the first video and the image of the first moving body based on the image feature amount of the first moving body and the image feature amount of the second moving body. If it is determined that the image of the second moving body is similar, causing the processor to execute a procedure for integrating information related to the moving path of the second moving body into information related to the moving path of the first moving body. A non-transitory computer-readable storage medium. A non-transitory computer readable storage medium according to claim 11, comprising:
The program further includes an image of the selected moving object and an image of the moving object detected from the second video both including a predetermined first type area and included in the selected image. The image of the first type area is similar to the image of the first type area included in the moving body image detected from the second video, and from the second video A non-transitory computer readable method characterized by causing the processor to execute a procedure for outputting information relating to the second type area when the extracted moving body image further includes a second type area. Storage medium. A non-transitory computer readable storage medium according to claim 11, comprising:
The program further selects one of the plurality of images of each moving body when a plurality of moving body images are detected from any of the videos and a plurality of images of each of the moving bodies are detected. And a non-transitory computer readable program that causes the processor to execute a procedure for displaying the selected image of each mobile object so as not to overlap with the selected image of another mobile object. Storage medium.