KR102594803B1 - syntax-based method of searching RE-ID in compressed video - Google Patents

Abstract

The present invention generally relates to a technology in which a video analysis device effectively performs a same-person (RE-ID) search on compressed video (e.g., CCTV footage). In particular, the present invention performs playback video analysis by selecting image frames in which moving objects are detected based on the syntax information of the compressed video, and provides object characteristic information (RE-ID Feature) for the representative thumbnail (best shot image) of the moving object. This relates to a technology that can achieve high-speed same-person search by extracting and constructing a database, and then providing a target image that is the subject of a same-person (RE-ID) search, by querying the database for the object characteristic information of the target image. According to the present invention, syntax obtained from bitstream parsing for compressed video without performing complex and system resource-demanding processes such as compressed video decoding, downscale resizing, and playback video analysis as in the prior art. Since the best shot thumbnail of the object is obtained based on the information and the same person (RE-ID) search is performed based on this best shot thumbnail, there is an advantage in improving the processing efficiency of the same person search.

Description

Syntax-based method of searching RE-ID in compressed video}

The present invention generally relates to a technology in which a video analysis device effectively performs a same-person (RE-ID) search on compressed video (e.g., CCTV footage).

In particular, the present invention performs playback video analysis by selecting image frames in which moving objects are detected based on the syntax information of the compressed video, and provides object characteristic information (RE-ID Feature) for the representative thumbnail (best shot image) of the moving object. This relates to a technology that can achieve high-speed same-person search by extracting and constructing a database, and then providing a target image that is the subject of a same-person (RE-ID) search, by querying the database for the object characteristic information of the target image.

Recently, it is common to build a video control system using CCTV to prevent crime, safety accidents, and secure post-mortem evidence. There is active research on technology to analyze CCTV footage to increase the control efficiency of video control systems.

[Figure 1] is a general configuration diagram of a video control system.

CCTV cameras (10) are distributedly installed at various points and provide real-time captured images to the video control device (20). The CCTV camera 10 generally generates images with high resolution (e.g., Full HD) and high frames (e.g., 24 frames per second), compresses the images with H.264 AVC and H.265 HEVC, and transmits them.

The video control device 20 provides real-time monitoring of the video captured by the CCTV camera 10 to the control personnel and stores it in the storage device 30 for later confirmation. In addition, the video control device 20 transmits the CCTV captured video to the video analysis device 40 and instructs video analysis according to a specific purpose for video control.

The video analysis device 40 cooperates with the video control device 20 to analyze captured images and provides various information (e.g., presence of objects, occurrence of unusual situations, object tracking, etc.) to the video control device 20. [Figure 2] is a general configuration diagram of video decoding performed by the video analysis device 40 in the case of H.264 AVC. To decode compressed video, a parser (41), an entropy decoder (42), an inverse converter (43), a motion vector operator (44), a predictor (45), and a deblocking filter (46) are provided, and these modules are used to capture CCTV footage. Decompresses the video and restores the original video data. At this time, the parser 41 parses the motion vector and coding type for the coding unit of the compressed video. These coding units are generally image blocks such as macroblocks or subblocks.

Object tracking is an important function in CCTV video surveillance. Object tracking is a function that detects objects (e.g. people, cars) in CCTV images and tracks the detected objects through video analysis. To this end, the video analysis device 40 must detect objects, set tracking objects, confirm object identity, and track objects through CCTV video analysis. At this time, to confirm object identity, predict the position and size according to the frame progress of the tracked object, extract and compare object features, detect movement of feature points of the tracked object (e.g. Meanshift, Optical flow, Camshift), and compare identity. do.

As a special form of object tracking, same-person (RE-ID) search has been discussed recently. In the field of video surveillance, RE-ID search refers to finding the relevant object (same object) in the captured image or other CCTV captured images when a target image is queried in the CCTV captured image, searching for a missing person or It is useful for tracking down people of interest.

[Figure 3] is a general flow chart for performing a search for the same person in a compressed video.

Referring to [Figure 3], the compressed video is decoded according to the video compression standard (e.g. H.264 AVC, H.265 HEVC) to obtain the playback video (S10), and the frame images that make up the playback video are converted into small images. , For example, down-scale resizing to about 320x240 pixels (S20). Then, after obtaining differentials for the resized frame images and identifying objects through playback image analysis (S30), object characteristic information (RE- ID Feature) and perform identical object analysis and search (S40).

However, in order to search for the same person in this way, compressed video decoding, downscale resizing, and playback video analysis must be performed. These are processes that have a high complexity and require a lot of system resources. As a result, in conventional video control systems, the number of channels that one video analysis device can simultaneously process is significantly limited. Currently, the RE-ID simultaneous processing capacity of the high-performance video analysis device 40 is up to 10 channels. As more and more CCTV cameras are installed to prevent crime and safety accidents, a large number of video analysis devices are required, which causes problems such as increased costs and difficulty in securing physical space.

Republic of Korea Patent No. 10-2374776 “Target object re-identification system and method based on CCTV location information and object movement information” Republic of Korea Patent No. 10-2416620 “Method and system for analyzing local facility use behavior based on object tracking using CCTV” Republic of Korea Patent No. 10-1519235 “Wireless sensor network-based smart security CCTV system and its control method” Republic of Korea Patent No. 10-2470837 “Spatial analysis method using CCTV image range in 3D space” Republic of Korea Patent No. 10-2315371 “Smart CCTV control and alarm system”

The purpose of the present invention is to provide a technology for a general video analysis device to effectively perform a same-person (RE-ID) search on compressed video (e.g., CCTV footage).

In particular, the purpose of the present invention is to perform playback video analysis by selecting image frames in which moving objects are detected based on the syntax information of the compressed video, and to obtain object characteristic information (RE-ID Feature) for the representative thumbnail (best shot image) of the moving object. ) is extracted and a database is constructed, and when a target image that is the subject of a same person (RE-ID) search is given, the object characteristic information of the target image is configured to be searched in the database, providing a technology that can achieve high-speed same person search. .

The problem to be solved by the present invention is not limited to this matter, and other problems to be solved can be understood from the description in this specification.

In order to achieve the above object, the present invention proposes a method in which a video analysis device processes identical person search based on syntax for compressed video.

The syntax-based identical person search method for compressed video according to the present invention includes a first step of extracting syntax information by parsing a bitstream of compressed video; A second step of detecting a moving object in a series of image frames constituting a compressed video based on syntax information; A third step of obtaining a thumbnail for the moving object by analyzing the image frame in which the moving object is detected; A fourth step of tracking the moving object in the compressed video and obtaining the best shot image of the moving object; A fifth step of extracting object characteristic information (RE-ID Feature) from the best shot image of the moving object; A sixth step of database management of object characteristic information of moving objects; A seventh step of receiving a target image for identical person (RE-ID) search; An eighth step of performing a search for the same person for the target image by querying the database for object characteristic information of the target image.

In the present invention, the fourth step is preferably configured to track the moving object in the compressed video and set the thumbnail with the largest area of the moving object among the thumbnails of the moving object as the best shot image.

The syntax-based identical person search method for compressed video according to the present invention is performed between the second and third steps, and if the image frame in which the moving object is detected is a P frame located within a preset number at the rear of the GOP, the corresponding image The method may further include skipping the playback video analysis for the frame and proceeding to the first step.

In the present invention, syntax information may include a motion vector and coding type for a coding unit.

In the present invention, the second step includes calculating a motion vector accumulation value for a specific time for each video block for a series of image frames constituting the compressed video; Marking an image block with a motion vector accumulation value exceeding a preset first threshold among a plurality of image blocks as a moving object area - Hereinafter, an image block marked as a moving object area is referred to as a 'marked image block'. )', and an image block that is not marked as a moving object area is called an 'unmarked image block' --; Identifying one or more image blocks (hereinafter referred to as 'neighboring blocks') adjacent to the marking image block; Marking a neighboring block with an intra picture coding type as a moving object area; Marking a neighboring block with a motion vector value exceeding a preset second threshold as a moving object area; Marking a preset number or less of unmarked image blocks surrounded by marked image blocks as a moving object area; It may be configured to include the step of setting a mass of marking image blocks connected to each other in a series of image frames constituting the compressed image as a moving object.

Meanwhile, the computer program according to the present invention is stored in a non-volatile storage medium in order to execute the syntax-based identical person (RE-ID) search method for the compressed video described above on the computer.

According to the present invention, syntax obtained from bitstream parsing for compressed video without performing complex and system resource-demanding processes such as compressed video decoding, downscale resizing, and playback video analysis as in the prior art. Since the best shot thumbnail of the object is obtained based on the information and the same person (RE-ID) search is performed based on this best shot thumbnail, there is an advantage in improving the processing efficiency of the same person search.

[Figure 1] is a general configuration diagram of a video control system.
[Figure 2] is a general configuration diagram of video decoding performed by a video analysis device.
[Figure 3] is a general flowchart of performing a search for the same person in a compressed video.
[Figure 4] is a flowchart of a syntax-based identical search process for compressed video according to the present invention.
[Figure 5] is a conceptual diagram of decoding skip in the present invention.
[Figure 6] is an example of obtaining a best shot image in the present invention.
[Figure 7] is a flowchart of detecting a moving object based on syntax from compressed video in the present invention.
[Figure 8] is an example of detecting an effective motion area for CCTV compressed video.
[FIG. 9] is an example of detecting a boundary area for the video image of [FIG. 8].
[FIG. 10] is an example of interpolation applied to the video image of [FIG. 9].

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

In describing the present invention, detailed descriptions of parts that overlap with the prior art may be omitted.

[Figure 4] is a flowchart of a syntax-based identical search process for compressed video according to the present invention. This same person search process can be performed by a computer device, for example, the video analysis device 40 in the video control system of [Figure 1], and compressed video is video data compressed by, for example, H.264 AVC, H.265 HEVC, etc. am.

Steps (S1000, S1100): First, the image analysis device 40 parses the bitstream of the compressed video and extracts syntax information. At this time, the syntax information may be composed of a motion vector and a coding type for the coding unit of the compressed video according to the video compression standard. The size of the coding unit is generally 64x64. It ranges from pixel to 4x4 pixel and can be set in various ways depending on the designer's choice.

The image analysis device 40 detects moving objects based on syntax information in a series of image frames constituting compressed images. The process of detecting a moving object based on syntax information will be described later with reference to [FIGS. 7] to [FIG. 10].

Step (S1200): Next, the image analysis device 40 analyzes the image frame in which the moving object is detected based on syntax information and obtains a thumbnail for the moving object through this. Among the series of image frames that make up the compressed video, image frames in which no moving object was detected in (S1100) are skipped, and preferably, compressed video decoding, downscale resizing, and By analyzing the playback video, a thumbnail for the moving object is obtained as shown in [Figure 6].

At this time, video analysis is performed on the image frame in which the moving object is detected to determine whether a valid object has been detected. It is determined whether an object exists in the image frame through decoding, downscale resizing, difference image acquisition, and playback image analysis as seen in [Figure 2]. In this specification, the object extracted through image analysis is called an ‘effective object’. Depending on the embodiment, the moving object detected in (S1100) may be used as a valid object.

As an example, the image analysis of (S1200) may be configured to be performed only on the moving object portion (blue area in [FIG. 10]) detected in (S1100). In another embodiment, the image analysis of (S1200) may be configured to be performed on the entire corresponding image frame. In other words, if a moving object, even a very small one, is discovered based on syntax information, the entire image frame is closely examined through video analysis. In the former case, the video analysis speed will be faster.

The present invention can be configured to skip the image analysis process by determining that there is no important information in the image frame in which the moving object is not detected in (S1100). In this case, the video analysis process, that is, decoding, downscale resizing, difference image acquisition, and playback video analysis, is not performed on the image frame, but only bitstream parsing and some operations are performed, so it uses fewer computing resources.

Meanwhile, if the corresponding image frame is a P frame located within a preset number at the rear of the GOP, the video analysis for the corresponding image frame may be skipped and proceed to (S1000).

[Figure 5] is a diagram showing a series of image frames constituting compressed video and the concept of decoding skip in the present invention. I (Intra) frames can be decoded as a single frame, and P (Prediction or Inter) frames can only be decoded if the video from the last I frame to the immediately preceding P frame has been decoded. GOP (Group of Pictures) is a bundle of one I frame (Key Frame) and P frames, and is the minimum unit for completely decoding compressed video.

Video analysis generally does not analyze all frames, but only analyzes about 3 to 4 frames per second. Assuming that the input video per second is 30 frames per second and analysis is performed once every 10 frames, within the current GOP(t), video analysis is performed every 10 P frames after the "I frame video analysis. At this time, the next GOP(t+ Before analyzing the I frame in 1), there is no need to analyze video for P frames located within 9 of the rear end of the GOP, that is, the 21st to 29th P frames. Computing resources can also be saved through this video analysis skip operation. Meanwhile, depending on the implementation example, B frames may be mixed in the P frame portion in [FIG. 5].

Step (S1300): Next, the image analysis device 40 tracks the moving object in the compressed image and obtains the best shot image of the moving object. At this time, when a moving object is identified through video analysis, video analysis is continuously performed until the object disappears from the video, and through this, multiple thumbnails for one moving object are obtained. Among these thumbnails, the thumbnail that is judged to be the best according to preset standards is set as the best shot image for the moving object.

[Figure 6] is an example of obtaining a best shot image in the present invention. Object tracking is performed for a specific moving object by comparing its features or predicting its moving position, and multiple thumbnails are obtained from the time the moving object appears in the image until it disappears. After a moving object disappears from the video, the thumbnail with the largest area of the moving object among the thumbnails of the moving object can be set as the best shot image.

Steps (S1400, S1500): Next, the image analysis device 40 extracts object characteristic information (RE-ID Feature) for the best shot image of the moving object and manages the object characteristic information of the best shot image in a database. Because object characteristic information (RE-ID Feature) is already a known technology in the field of identical person (RE-ID) search, a detailed description of it will be omitted. There are many implementation examples using machine learning (deep learning) in the same person (RE-ID) search technology, and in the present invention, object characteristic information extracted from the best shot image of a moving object can be used for machine learning (deep learning).

Steps (S1600, S1700): Then, the video analysis device 40 receives a target image for searching for the same person (RE-ID) from the outside. For example, in order to search for a missing person, an image of the missing person is provided. The image analysis device 40 searches the database for object characteristic information of the target image (e.g., image of a missing person) and performs a search for the same person for the target image. Since the same person search itself based on object characteristic information is already a known technology, a detailed description of it will be omitted.

In the above description, (S1000) to (S1500) are processes for building a database of object characteristic information from compressed images, and (S1600) and (S1700) are processes for processing the same person search for a specific target image using this database. It's a process. If these two processes are implemented as separate tasks or threads and executed simultaneously, real-time search of the same person can be achieved in the CCTV video control system.

[Figure 7] is a flowchart of detecting a moving object based on syntax from compressed video in the present invention. The process in [FIG. 7] corresponds to (S1100) in [FIG. 4].

In the present invention, a method is used to identify parts that appear to be moving objects from compressed video based on syntax information and set them as object unit areas. This is a method of extracting image chunks presumed to contain moving objects based on syntax information without any knowledge of the video content. To achieve this, the bitstream of the compressed video is parsed and the moving object area is quickly extracted through syntax information for each video block. As a video block, either a macro block or a sub block or a combination thereof can be used, and as syntax information, a motion vector and a coding type are preferable. do. As can be seen in [FIGS. 8] to [FIG. 10], the moving object area obtained in this way does not accurately reflect the boundaries of the moving object existing in the image, but has the advantage of fast processing speed and high reliability.

Steps (S100, S110): First, calculate the cumulative motion vector for a specific time for each video block for a series of image frames constituting the compressed video, and calculate the motion vector that exceeds the first preset threshold among the plurality of video blocks. Image blocks with accumulated values are marked as moving object areas. This process detects effective motion that can actually be recognized as meaningful from the compressed video based on the motion vector, and marks the portion of the image where valid motion is detected as a moving object area.

Through bitstream parsing, the motion vector for the coding unit of the compressed video can be obtained. The motion vector value of the video block is stored for a certain period of time (e.g. 500 msec) preset for each video block. accumulates. In the case of compressed video at 30 frames per second, the motion vectors generated from each video block are accumulated over 15 frames (i.e., 500 msec). Then, it is checked whether the resulting motion vector accumulation value exceeds a preset first threshold (eg, 20). If such a video block is found, it is considered that valid movement has been found in the video block and it is marked as a moving object area. On the other hand, even if a motion vector occurs in a certain image block, if the cumulative value over a certain period of time does not exceed the first threshold, the image change is assumed to be minimal and the image block is not marked as a moving object area.

[Figure 8] is an example of detecting an effective motion area using (S100) and (S110) for CCTV compressed video. In [Figure 8], image blocks with motion vector accumulation values greater than the first threshold are marked as moving object areas and displayed in red. Looking at [Figure 8], sidewalk blocks, roads, and shadowed areas are not displayed as moving object areas, while people walking or driving cars are displayed as moving object areas. In this specification, for convenience, an image block marked as a moving object area is called a 'marked image block', and an image block not marked as a moving object area is called an 'unmarked image block'.

Steps (S120 ~ S140): Next, identify one or more video blocks (hereinafter referred to as 'neighboring blocks') adjacent to the marking video block, and these neighboring blocks have a coding type of intra picture or a preset second video block. If the motion vector value exceeds the threshold, it is marked as a moving object area. This process extends the boundaries of these moving object areas to approximately how far they are by examining the surrounding areas of the moving object areas detected in (S100). Through this process, the moving object areas detected in the form of fragmented image blocks in (S100) are connected to each other to create a meaningful chunk shape.

In the above (S100), image blocks that clearly appear to correspond to a moving object in the compressed image were detected and marked as a moving object area by selecting image blocks according to strict judgment criteria. In (S120), unmarked image blocks located around the marked image blocks extracted in (S100) are inspected. For convenience, these are called ‘neighboring blocks’. These neighboring blocks are parts that are not marked as a moving object area in (S100), and in (S120) to (S140), they are examined further to check whether any of these neighboring blocks can be included in the moving object area.

In compressed video, macroblocks and subblocks are very small in size. Therefore, if it is a video of people, cars, animals, etc., such as CCTV video, it is expected that moving objects will not appear in only one video block due to their nature, but will appear across multiple video blocks. In other words, it is expected that the likelihood that a moving object is captured in an image block that exists near an image block in which a moving object is captured is relatively higher than in an image block that does not. Reflecting such technical ideas, in (S120), it is determined whether non-marking image blocks existing around the marking image block correspond to the moving object area according to relatively relaxed judgment criteria.

Preferably, each neighboring block is inspected, and if the motion vector value detected in the current frame is more than a preset second threshold (e.g. 0) or the coding type is Intra Picture, the corresponding video block is also a moving object. Mark as an area. In another embodiment, if the motion vector accumulation value previously calculated in (S100) for a neighboring block is greater than or equal to the second threshold (e.g., 5) or the coding type is intra picture, the corresponding video block may also be marked as a moving object area. You can. At this time, it is logically appropriate to set the second threshold to a smaller value than the first threshold.

Conceptually, if effective motion is found and an image block with some degree of movement is found near the moving object area, it is marked as a moving object area because it is highly likely to be part of the previous moving object area. Additionally, in the case of intra pictures, since there is no motion vector, it is impossible to determine whether or not it is a moving object area based on the motion vector. Accordingly, if an intra picture is located adjacent to an image block that has already been detected as a moving object area, it is assumed to form a lump together with the previously extracted moving object area. This is because the loss when one video block that is not in the moving object area is included in the moving object area is not very large, whereas the loss when the moving object area is fragmented is large.

Looking at the previous [Figure 8], it can be seen that the video blocks corresponding to the moving objects were not properly marked and that only some were marked. In other words, if you look at a walking person or a driving car, you can find that not all of the object is marked, but only some of the image blocks are marked. Additionally, it is often found that multiple moving object areas are formed for one moving object. This means that the criteria for judging the moving object area adopted in (S100) above are very useful for filtering out the general area, but are quite strict. Therefore, focusing on the moving object area marked in (S100), the surrounding image blocks are reviewed, and if a certain standard is met, the moving object area is additionally marked, resulting in a process of detecting the boundary of the moving object area. need.

[Figure 9] is a diagram showing an example of the results of applying boundary area detection to the CCTV video image of [Figure 8]. Through the above process, a number of video blocks marked as moving object areas are displayed in blue. Comparing [FIG. 8] and [FIG. 9], the blue moving object area has expanded near the moving object area previously shown in red in [FIG. 8], and this can be compared with the actual image captured by CCTV. You can find that it has reached the point where all moving objects are covered.

Step (S150): Next, unmarked video blocks of less than a preset number surrounded by marked video blocks are marked as a moving object area. This process applies interpolation to the moving object area detected in (S100) to (S140) to organize fragmentation of the moving object area.

In (S100) to (S140), the presence or absence of a moving object area is determined on a video block basis, so there are video blocks in the middle that are not marked as a moving object area even though they actually belong to one moving object (e.g., a person). As a result, division into multiple moving object areas may occur. If unmarked video blocks exist in the middle like this, they can be considered as if they are multiple individual moving objects. If the moving object area is fragmented in this way, identification of the moving object may become inaccurate. Accordingly, if there is one or a few unmarked image blocks surrounded by a plurality of marked image blocks, they are additionally marked as a moving object area, which is called interpolation. Through this, moving object areas that are divided into several parts can be united into one.

[FIG. 10] is an example of interpolation applied to the video image of [FIG. 9]. Comparing [FIG. 8] to [FIG. 10], it can be found that the moving object area properly reflects the actual image situation through the boundary area detection process and the interpolation process. If we judge it as a lump marked in red in [Figure 8], it will be treated as if there are many very small objects moving in the image, which does not correspond to reality. On the other hand, if it is judged as a lump marked in blue in [Figure 10], it will be treated as the presence of several moving objects with a certain volume, thus similarly reflecting the actual scene.

Step (S160): Next, in a series of image frames constituting the compressed video, the marking video blocks are connected to each other and the lumped mass is set as a moving object. One or more moving object areas were obtained from the compressed image through (S100) to (S150). In each step (S100) to (S150), each video block was judged and marked as to whether it belonged to the moving object area, but ultimately, the mass of video blocks that were connected to each other was called the region of moving object. is dealt with. The moving object area is a mass of multiple image blocks as shown in [Figure 10].

Meanwhile, the present invention can be implemented in the form of computer-readable code on a computer-readable non-volatile recording medium. These non-volatile recording media include various types of storage devices, such as hard disks, SSDs, CD-ROMs, NAS, magnetic tapes, web disks, and cloud disks. Additionally, the present invention may be implemented in the form of a computer program stored on a medium in order to execute a specific procedure in combination with hardware.

10: CCTV camera
20: Video control device
30: storage device
40: video analysis device

Claims (5)

In a CCTV video control system that processes a large amount of compressed video, the video analysis device 40 processes a syntax-based search for the same person (RE-ID) for compressed video,
A first step of extracting syntax information including a motion vector and coding type for a coding unit according to a video compression standard by parsing the bitstream of a plurality of compressed images provided from the CCTV camera 10;
Based on the syntax information including the extracted motion vector and coding type, parts that appear to be moving objects are identified in a series of image frames constituting the compressed video, and these are set as object unit areas to perform compressed video decoding, downscale resizing, A second step of detecting a moving object without performing difference image acquisition and playback image analysis;
Among the series of image frames constituting the plurality of compressed images, an image frame in which a moving object was not detected in the second step (hereinafter referred to as 'first image frame') and an image frame in which a moving object was detected (hereinafter referred to as ' A 3a step of classifying and identifying a second image frame (referred to as a 'second image frame');
A 3b step of skipping image analysis by compressed image decoding, downscale resizing, difference image acquisition, and playback image analysis for the first image frame among the series of image frames constituting the plurality of compressed images;
Among the series of image frames that make up the plurality of compressed images, compressed image decoding, downscale resizing, difference image acquisition, and playback image analysis are performed on the second image frame, which is a P frame located within a preset number at the rear of the GOP (Group of Pictures). A 3c step of skipping image analysis by;
A 3d step of performing image analysis by compressed image decoding, downscale resizing, difference image acquisition, and playback image analysis only on the remaining second image frames among the series of image frames constituting the plurality of compressed images;
A 3e step of acquiring a thumbnail for the moving object detected in the second step by analyzing the image in the 3d step;
A 4a step of obtaining a plurality of thumbnails for the moving object by performing object tracking on the compressed image until the moving object detected in the second step disappears from the image;
A 4b step of acquiring a best shot image for the moving object from among the obtained plurality of thumbnails;
A fifth step of extracting object characteristic information (RE-ID Feature) from the best shot image of the moving object;
A sixth step of managing a database of object characteristic information of the best shot image of the moving object;
A seventh step of receiving a target image for identical person (RE-ID) search; and
An eighth step of querying the database for object characteristic information of the target image and performing a search for the same person for the target image based on the object characteristic information;
It is composed including,
The first to sixth steps are configured as a first task or thread, the seventh step and the eighth step are implemented as a second task or thread, and the first and second tasks or threads are executed simultaneously. A syntax-based identical person search method for compressed video that is configured to perform real-time identical person search in a CCTV video control system.
In claim 1,
The 4b step is configured to set the thumbnail with the largest area of the moving object among the obtained plurality of thumbnails as the best shot image.
delete In claim 1,
The second step is,
Calculating a motion vector accumulation value for a specific time for each video block for a series of image frames constituting the compressed video;
Marking an image block with a motion vector accumulation value exceeding a preset first threshold among the plurality of image blocks as a moving object area - Hereinafter, the image block marked as a moving object area is referred to as a 'marked image block (marked image block)'. block)', and an image block that is not marked as a moving object area is called an 'unmarked image block'--;
Identifying one or more image blocks (hereinafter referred to as 'neighboring blocks') adjacent to the marking image block;
Marking a neighboring block with an intra picture coding type as a moving object area;
Marking a neighboring block with a motion vector value exceeding a preset second threshold as a moving object area;
Marking a preset number or less of unmarked image blocks surrounded by marked image blocks as a moving object area;
In a series of image frames constituting a compressed image, marking image blocks are connected to each other and set as a moving object;
A syntax-based identical person search method for compressed video, comprising:
A computer program stored in a storage medium to execute a syntax-based same-person search method for compressed video according to any one of claims 1, 2, and 4 on a computer.

Images