WO2022156234A1 - 一种目标重识别方法、装置及计算机可读存储介质 - Google Patents
一种目标重识别方法、装置及计算机可读存储介质 Download PDFInfo
- Publication number
- WO2022156234A1 WO2022156234A1 PCT/CN2021/117512 CN2021117512W WO2022156234A1 WO 2022156234 A1 WO2022156234 A1 WO 2022156234A1 CN 2021117512 W CN2021117512 W CN 2021117512W WO 2022156234 A1 WO2022156234 A1 WO 2022156234A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- target
- global
- image
- identification
- target image
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 78
- 238000003860 storage Methods 0.000 title claims abstract description 23
- 238000001514 detection method Methods 0.000 claims abstract description 148
- 238000012544 monitoring process Methods 0.000 claims abstract description 47
- 230000000717 retained effect Effects 0.000 claims description 13
- 230000009466 transformation Effects 0.000 claims description 10
- 238000006243 chemical reaction Methods 0.000 claims description 9
- 238000009432 framing Methods 0.000 claims description 7
- 230000036961 partial effect Effects 0.000 claims description 7
- 238000012216 screening Methods 0.000 claims description 6
- 238000010586 diagram Methods 0.000 description 14
- 238000004422 calculation algorithm Methods 0.000 description 8
- 238000004590 computer program Methods 0.000 description 7
- 230000009286 beneficial effect Effects 0.000 description 6
- 230000008901 benefit Effects 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000012545 processing Methods 0.000 description 4
- 241001465754 Metazoa Species 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000002085 persistent effect Effects 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000003247 decreasing effect Effects 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000000670 limiting effect Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000003908 quality control method Methods 0.000 description 1
- 230000002829 reductive effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/40—Scenes; Scene-specific elements in video content
- G06V20/41—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N7/00—Television systems
- H04N7/18—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast
- H04N7/181—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast for receiving images from a plurality of remote sources
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Definitions
- the invention belongs to the field of identification, and in particular relates to a target re-identification method, device and computer-readable storage medium.
- the present invention provides the following solutions.
- a first aspect provides a target re-identification method, comprising: acquiring a plurality of current frames collected by a plurality of cameras arranged in a monitoring area; performing target detection according to the plurality of current frames, and determining a target image captured by each camera; Carry out quantity detection according to the target image captured by each camera to obtain the global target quantity; carry out target re-identification according to the target image and target recognition library, the target recognition library includes the identity and feature data of at least one target; when the global target quantity is detected When the preset addition conditions are met, at least one unrecognized target image is determined according to the result of the target re-identification, and a new identification mark is created to mark the at least one unrecognized target image; according to the new identification mark and the characteristics of the at least one unrecognized target image Data update target recognition library.
- performing target detection according to a plurality of current frames further includes: inputting the plurality of current frames into a trained target detection model to extract target images captured by each camera; wherein the target detection
- the model is a human detection model created based on the YOLOv4-tiny network.
- the method further includes: training the YOLOv4-tiny network according to the real collected images in the monitoring area to obtain a target detection model.
- the target image is a partial image containing target features in the current frame
- the number detection is performed according to the target image captured by each camera, and further includes: according to the viewing position of each camera, the captured target The image is converted into position to obtain the global position corresponding to the target image captured by each camera; the global position coincidence degree of the target images captured by different cameras is determined, and the target images captured by different cameras are screened according to the global position coincidence degree. Number of target images retained after detection screening.
- the method further includes: when the result of the quantity detection is less than the number of previous global targets, determining whether to There is a target that leaves the monitoring area from the predetermined area; if there is no target, the number of previous global targets is still retained as the number of global targets determined this time; if there are targets, the result of quantity detection is used as the number of global targets determined this time ; wherein, the number of prior global targets is obtained by performing target detection and quantity detection on the previous frame of multiple current frames.
- performing position transformation on the captured target image according to the framing position of each camera further comprising: performing a projective transformation on the bottom center point of the target image in the current frame according to the framing position of each camera , thereby determining the ground coordinates of each target image.
- the method further includes: inputting multiple current frames into a trained target quantity detection model to perform target detection and quantity detection to obtain a global target quantity; wherein, the target quantity detection model is based on YOLOv4- Pedestrian detection model created by tiny network.
- the target re-identification is performed according to the target image and the target recognition library, further comprising: calculating the similarity between the target image and the feature data in the target recognition library, and according to the calculated similarity, The target image is subjected to target re-identification; when the result of the target re-identification indicates that the first target image matches the first target in the target recognition library, the first target image is marked according to the identity of the first target.
- the method further includes: if the current frame is not the first frame and the number of global targets corresponding to the current frame increases compared to the number of global targets corresponding to the previous frame, then the number of global targets conforms to the preset number Increase conditions; if the current frame is the first frame, the default global target number meets the preset increase conditions.
- updating the target recognition library according to the new identity identifier and the feature data of the at least one unrecognized target image further includes: judging whether the at least one unrecognized target image satisfies a preset image quality condition; The identity identifier and the unrecognized target image satisfying the preset image quality condition are stored in the target recognition library correspondingly.
- the method further includes: according to the first target image or the feature value of the first target image, the characteristics of the first target in the target recognition library are compared. Data is dynamically updated.
- the method further includes replacing and updating the target recognition database, which specifically includes: according to the comparison result of the source time corresponding to the feature data of each target in the target recognition database and the current time, updating the target recognition database Carry out replacement and update; And/or, according to the comparison result of the global position corresponding to the characteristic data of each target in the target recognition library and the current global position of each target, the target recognition library is replaced and updated; And/or, according to the target Identify the feature similarity between multiple feature data of each target in the library, and replace and update the target identification library.
- the method further includes: after the quantity of characteristic data of any one target exceeds a preset threshold, starting a replacement update.
- a target re-identification device comprising: an acquisition module for acquiring a plurality of current frames collected by a plurality of cameras arranged in a monitoring area; a target detection module for performing target detection according to the plurality of current frames , determine the target image captured by each camera; the quantity detection module is used for quantity detection according to the target image captured by each camera to obtain the global target quantity; Target re-identification, the target recognition library includes at least one target's identity and feature data; the identity module is used to determine at least one unrecognized target image according to the result of target re-identification when it is detected that the number of global targets meets the preset increase condition , creating a new identity mark to mark at least one unrecognized target image; a target recognition library updating module for updating the target recognition library according to the new identity mark and feature data of at least one unrecognized target image.
- the target detection module is further configured to: input multiple current frames into the trained target detection model to extract the target image captured by each camera; wherein, the target detection model is based on YOLOv4 - Human detection model created by tiny network.
- the target detection module is further configured to: train the YOLOv4-tiny network according to the real collected images in the monitoring area to obtain a target detection model.
- the target image is a partial image containing target features in the current frame
- the quantity detection module is further configured to: perform position conversion on the captured target image according to the viewing position of each camera to obtain each camera The global position corresponding to the captured target image; determine the global position coincidence of the target images captured by different cameras, filter the target images captured by different cameras according to the global position coincidence, and detect the number of target images retained after screening .
- the quantity detection module is further configured to: when the result of quantity detection is less than the number of previous global targets, then according to the multiple current frames collected by multiple cameras and the previous frame of multiple current frames , to judge whether there is a target that leaves the monitoring area from the predetermined area; if there is no target, the number of previous global targets is still retained as the number of global targets determined this time; The number of global targets; wherein, the number of previous global targets is obtained by performing target detection and quantity detection on the previous frame of multiple current frames.
- the quantity detection module is further configured to: perform projective transformation on the bottom center point of the target image in the current frame according to the viewing position of each camera, so as to determine the ground coordinates of each target image.
- the device is further configured to: input multiple current frames into a trained target quantity detection model to perform target detection and quantity detection to obtain a global target quantity; wherein, the target quantity detection model is based on YOLOv4 -Pedestrian number detection model created by tiny network.
- the target re-identification module is further used to: calculate the similarity between the target image and the feature data in the target recognition library, and perform target re-identification on the target image according to the calculated similarity; When the result of the target re-identification indicates that the first target image matches the first target in the target recognition library, the first target image is marked according to the identity of the first target.
- the identity identification module is further configured to: if the current frame is not the first frame, and the number of global targets corresponding to the current frame increases compared to the number of global targets corresponding to the previous frame, the number of global targets The preset increase conditions are met; if the current frame is the first frame, the default global target number meets the preset increase conditions.
- the target recognition library update module is further configured to: determine whether at least one unrecognized target image satisfies the preset image quality condition; Correspondingly stored in the target recognition library.
- the target recognition library updating module is further configured to: dynamically update the feature data of the first target in the target recognition library according to the first target image or the feature value of the first target image.
- the target recognition library updating module is further configured to: replace and update the target recognition library according to the comparison result between the source time corresponding to the characteristic data of each target in the target recognition library and the current time; and /or, according to the comparison result of the global position corresponding to the feature data of each target in the target recognition library and the current global position of each target, the target recognition library is replaced and updated; and/or, according to each target recognition library The feature similarity between multiple feature data of a target is replaced and updated to the target recognition library.
- the target identification library update module is further configured to: start the replacement update after the quantity of characteristic data of any target exceeds a preset threshold.
- a target re-identification device comprising: one or more multi-core processors; a memory for storing one or more programs; when the one or more programs are processed by the one or more multi-core processors When the processor is executed, the one or more multi-core processors are made to implement: the method of the first aspect.
- a computer-readable storage medium stores a program, and when the program is executed by a multi-core processor, the multi-core processor causes the multi-core processor to perform the method of the first aspect.
- the above-mentioned at least one technical solution adopted in this embodiment of the present application can achieve the following beneficial effects: in this embodiment, the number of global targets in the monitoring area is detected, and the creation and distribution of new identities are controlled by the detected number of global targets, It can well ensure the accurate allocation of identity identifiers, and ensure the accuracy and stability of target re-identification. .
- FIG. 1 is a schematic flowchart of a target re-identification method according to an embodiment of the present invention
- FIG. 2 is a ground schematic diagram of a monitoring area according to an embodiment of the present invention.
- FIG. 3 is a schematic diagram of a viewfinder screen of a plurality of cameras according to an embodiment of the present invention
- FIG. 4 is a schematic diagram of a current frame of a plurality of cameras according to an embodiment of the present invention.
- FIG. 5 is a schematic diagram of a target image captured by a plurality of cameras according to an embodiment of the present invention.
- FIG. 6 is a schematic diagram of a global position of a target image captured by a plurality of cameras according to an embodiment of the present invention
- FIG. 7 is a schematic structural diagram of a target re-identification apparatus according to an embodiment of the present invention.
- FIG. 8 is a schematic structural diagram of a target re-identification apparatus according to another embodiment of the present invention.
- A/B can mean A or B; "and/or” in this text is only an association relationship to describe related objects, indicating that there can be three kinds of relationships, For example, A and/or B can mean that A exists alone, A and B exist at the same time, and B exists alone.
- first”, “second”, etc. are used for descriptive purposes only, and should not be construed as indicating or implying relative importance or implying the number of indicated technical features. Thus, a feature defined as “first”, “second”, etc., may expressly or implicitly include one or more of that feature. In the description of the embodiments of the present application, unless otherwise specified, "plurality" means two or more.
- a method for real-time target tracking comprising: acquiring a plurality of current frames collected by a plurality of cameras arranged in a monitoring area; performing target detection according to the plurality of current frames, and determining a target image captured by each camera; The target image captured by each camera is subjected to quantity detection to obtain the global target quantity; target re-identification is carried out according to the target image and a target recognition library, wherein the target recognition library includes at least one target identification and feature data; When it is detected that the number of global targets meets the preset increase condition, at least one unrecognized target image is determined according to the result of the target re-identification, and a new identity is created to mark the at least one unrecognized target image; The target recognition library is updated with the new identity and feature data of the at least one unrecognized target image.
- the execution subject may be one or more electronic devices ; From a program point of view, the execution subject can be a program mounted on these electronic devices accordingly.
- the method 100 includes:
- Step 101 acquiring multiple current frames collected by multiple cameras arranged in the monitoring area
- the monitoring area refers to the sum of the viewing areas of multiple cameras
- the multiple cameras include at least two cameras, and the viewing areas of the multiple cameras are adjacent to each other or at least partially overlap, so that the target to be tracked can be Movement in the monitoring area and then appearing in the viewing area of any one or more cameras.
- the current frames of the multiple cameras are respectively extracted from the surveillance videos of the multiple cameras, wherein the current frames of each camera have the same acquisition time.
- the tracking target in the present disclosure is preferably a pedestrian, and those skilled in the art can understand that the above-mentioned tracking target may also be other movable objects, such as animals, vehicles, etc., which is not specifically limited in the present disclosure.
- FIG. 2 shows a schematic monitoring scene, in which a camera 201 and a camera 202 are set
- FIG. 3 shows the viewfinder pictures of the camera 201 and the camera 202 .
- the surveillance video collected by the camera 201 can be parsed into a sequence of image frames (A 1 , A 2 , . . . , A N ), and the surveillance video collected by the camera 202 can be parsed into a sequence of image frames (B 1 , B 2 , . . . .,B N ), wherein the above analysis can be performed online or offline in real time.
- the method 100 may include:
- Step 102 perform target detection according to multiple current frames, and determine the target image captured by each camera;
- the target image may be a partial image containing target features in the current frame.
- the current frames A n and B n of the camera 201 and the camera 202 are shown .
- Detect output a series of pedestrian images (an example of a target image) for each camera.
- the target detection model may be, for example, a YOLO (Unified Real-Time Object Detection, You Only Look Once) model, which is not specifically limited in this disclosure.
- FIG. 5 shows multiple target (pedestrian) detection frames obtained by detecting multiple current frames A n and B n . It can be understood that the target can be intercepted from the current frame according to the target (pedestrian) detection frame.
- the intercepted target (pedestrian) image can be normalized to facilitate subsequent tracking display.
- step 102 may further include: inputting multiple current frames into the trained target detection model to extract the target captured by each camera Image; among them, the target detection model is a human detection model created based on the YOLOv4-tiny network.
- the deep learning-based real-time target detection algorithm YOLOV4-TINY can be improved to obtain YOLOV4-TINY-P (YOLOv4-tiny-People), and trained to generate a human detection model, which can be used to identify the overall characteristics of pedestrians , and is not affected by face coverings such as wearing masks.
- the above-mentioned target detection can be directly completed by using a plurality of ordinary surveillance cameras without a professional face camera.
- target detection algorithms such as faster-rcnn target detection algorithm, yolov4 target detection algorithm, etc., which is not specifically limited in this application.
- target detection models may be correspondingly adopted, which are not specifically limited in this application.
- the following steps can also be performed to obtain the above-mentioned target detection model: YOLOv4-tiny
- the network is trained to obtain a target detection model.
- the pedestrians in the actual scene such as the computer room can be trained in a targeted manner, and the target positive and negative samples can be added based on the actual scene.
- the target positive and negative samples can be added based on the actual scene.
- chairs, backpacks, servers and other items are negative samples
- pedestrians are positive samples.
- the training data can be jointly trained with actual computer room scene data, PASCAL VOC2007 and VOC2012 and other target detection data sets to further improve the model detection ability.
- the method 100 further includes:
- Step 103 Perform quantity detection according to the target images captured by each camera to obtain the global target quantity.
- the above-mentioned quantity detection can be performed using any possible target statistical method, which is not specifically limited in this application.
- the number of local targets in each camera can be detected individually first, and then the number of local targets can be accumulated, and then the overlapping target images captured by different cameras can be analyzed, and corresponding deletions are performed.
- the camera 201 captures three target (pedestrian) images, including (a 1 , a 2 , a 3 ), and the camera 202 captures one target (pedestrian) image, including (b), the number of local targets Accumulation, due to the intersection of the framing range between different cameras, there must be a situation where different cameras have captured target images of the same target from different angles. It can be determined through position analysis that a 3 captured by the camera 201 and the image captured by the camera 202. (b) Coincidence, the number of coincidences is deleted from the accumulated result of the number of local targets, so that the number of global targets can be obtained as 3.
- step 103 may further include the following steps: performing position conversion on the captured target image according to the viewing position of each camera to obtain the target captured by each camera.
- the global position corresponding to the image determine the global position coincidence of the target images captured by different cameras, and filter the target images according to the global position coincidence; determine the number of global targets in the monitoring area according to the number of target images retained after screening.
- the target image is a local image containing the target features in the current frame.
- the target image (a 1 , a 2 , a 3 ) is captured by the camera 201 , the target image (b) captured by the camera 202 , and the target image (a 1 , a) captured according to the framing position of the camera 201 2 , a 3 ) Perform position conversion, and perform position conversion on the captured target image (b) according to the framing position of the camera 202 to obtain the global position of each target image shown in FIG. 6.
- the camera 201 captures
- the global position of the target image a 3 and the target image b captured by the camera 202 has a high degree of coincidence. Assuming that it exceeds the preset coincidence degree threshold, it can be considered that the target images a 3 and b are actually the same target, and only one can be reserved. , and then it can be determined that the number of global targets in the monitoring area is 3.
- the number of detected global targets is reduced compared to the actual number. Based on this, the following steps may also be performed: when the number of detected targets is detected When the result is less than the number of previous global targets, then according to the multiple current frames collected by multiple cameras and the previous frame of multiple current frames, it is judged whether there is a target leaving the monitoring area from the predetermined area; If the target from the predetermined area leaves the monitoring area, the number of previous global targets is still retained as the number of global targets determined this time; if there is a target that leaves the monitoring area from the predetermined area, the result of the quantity detection will be used as the number of global targets determined this time. ;
- the number of prior global targets is obtained by performing target detection and quantity detection on multiple previous frames. Specifically, by replacing the multiple current frames in steps 101 to 103 with the previous frames of the multiple current frames, the same scheme can be used to obtain the prior global target number, which is not repeated in this application.
- the exit area such as the monitoring area, can be divided into predetermined areas to determine whether there is a target. According to the previous frame of multiple current frames, it can be determined that the target is located in the exit area, and according to multiple current frames, it can be determined that the target is located in the exit area. The exit area disappears.
- the target has actually left the monitoring area, and the result of the above-mentioned quantity detection can be used as the global target quantity.
- the result of the above-mentioned quantity detection can be used as the global target quantity.
- the target occlusion, etc. it can be considered that there is target occlusion, etc., and the number of previous global targets is still retained as the number of global targets determined this time.
- performing position transformation on the captured target image according to the viewing position of each camera further comprising: performing a projective transformation on the bottom center point of the target image in the current frame according to the viewing position of each camera, so as to determine Ground coordinates for each target image.
- performing a projective transformation on the bottom center point of the target image in the current frame according to the viewing position of each camera so as to determine Ground coordinates for each target image.
- the position of the bottom center point of each target image captured by each camera in Figure 5 can be obtained, and the bottom center point of each target image can be converted to obtain the actual ground position of the target to be identified in the monitoring scene, Figure 6
- the ground coordinates corresponding to each target image obtained by projection transformation are shown.
- the ground aisle under the viewing angle of each camera is an approximate trapezoid area, so for the target image captured by each camera, the bottom center point of each target image can be obtained through trapezoid-rectangle transformation.
- Coordinates in the rectangular area secondly, rotate the standard rectangular area according to the actual layout of the monitoring scene, calculate the rotated coordinates of the bottom center point of each target image through the rotation matrix, and finally perform the rotated coordinates according to the actual layout of the monitoring scene. Pan and zoom to get the final ground coordinate position.
- a target quantity detection model can be pre-trained to detect the global target quantity in the monitoring area in real time, and when executing the target real-time tracking method, multiple current frames are input into the trained target quantity.
- the detection model is performed to perform object detection and quantity detection, and the global target quantity is directly obtained.
- an improved people counting algorithm YOLOv4-TINY-PC (YOLOv4-tiny-People Counting) can be proposed by improving the deep learning-based real-time target detection algorithm YOLOV4-TINY.
- the YOLOV4-TINY algorithm does not have the ability to count people with multiple cameras.
- YOLOv4-TINY-PC can get the number of people in the monitoring area in real time to count the flow of people.
- the target number detection model can obtain the target image recognized by each camera through the pedestrian detection algorithm (YOLOv4-TINY-P), and perform position conversion on the target image to obtain the global position coordinates in the overall monitoring area.
- the target number detection model is a pedestrian number detection model created based on the YOLOv4-tiny network.
- pedestrian number detection models can also be created based on other networks such as faster-rcnn, yolov4.
- a target number detection model such as a vehicle number detection model and an animal number detection model can also be created for other application scenarios.
- the method further includes:
- Step 104 perform target re-identification according to the target image and the target recognition library.
- the target identification library includes the identification and characteristic data of at least one target.
- the target recognition library may include ⁇ target 1: feature data 1, ..., feature data N ⁇ ; ⁇ target 2: feature data 1, ..., feature data N ⁇ , and so on.
- the method may further include: calculating the similarity between the target image and the feature data in the target recognition library, and performing the calculation on the target image according to the calculated similarity.
- Target re-identification when the result of the target re-identification indicates that the first target image matches the first target in the target recognition library, identifying the identity of the first target for the first target image.
- the similarity between the shown pedestrian image b and the feature data of each target contained in the target recognition library is calculated, assuming that the similarity between the pedestrian image b and the feature data of target 1 is the highest, and the If the similarity exceeds the preset matching threshold, it can be considered that the result of target re-identification indicates that the pedestrian image b matches the target 1 in the target recognition library, and the pedestrian image b can be further identified as the target 1.
- the pedestrian image a 2 is matched to the target 2 in the target recognition library and identified.
- the pedestrian image a 3 also matches and identifies the target 1 in the target recognition library.
- the method further includes:
- Step 105 when it is detected that the number of global targets meets the preset increase condition, determine at least one unrecognized target image according to the result of target re-identification, and create a new identity identifier (hereinafter referred to as a new ID) to mark at least one unrecognized target image. .
- a new ID a new identity identifier
- a new ID needs to be assigned to the new target.
- the industry usually adopts the method of calculating the feature similarity between the target image and the feature data in the target recognition library to determine whether to create a new target. Assign a new ID, and in some scenarios, problems such as target occlusion and shooting angle will have a greater impact on the accuracy of the above judgment. For example, when a target already in the monitoring area cannot match the corresponding feature data in the target recognition library due to the poor shooting quality of the target image, it is easy to be mistaken as a new target.
- the target recognition library may include ⁇ target 1: feature data 1, ..., feature data N ⁇ ; ⁇ target 2: feature data 1, ..., feature data N ⁇ , and the result of the target re-identification indicates a pedestrian Image b is matched to target 1 and identified.
- Pedestrian image a 2 is matched to target 2 and identified.
- Pedestrian image a 3 is also matched to target 1 and identified.
- the pedestrian image a 1 does not match any target in the target recognition library, that is, the pedestrian image a 1 is the unrecognized target image determined in the above-mentioned target re-identification process. target 3 ) and mark the pedestrian image a1. This makes it possible to assign a new ID to the new target.
- the above-mentioned new target refers to the target that does not store the matching identification and feature data in the current target recognition library.
- a pedestrian may still be used as a new target when entering the monitoring area next time, and the new target needs to be re-assigned with a newly created identity and correspondingly stored in feature data.
- step 105 may further include a step of detecting whether the number of global targets meets the preset increase condition, specifically including: if the current frame is not the first frame, and the number of global targets corresponding to the current frame is the same When the number of prior global objects corresponding to the previous frame of the plurality of current frames increases, the number of global objects complies with the preset increase condition. If the current frame is the first frame, the default global target number meets the preset increase conditions. Specifically, the preceding global target number has been described above, and will not be repeated here.
- the method further includes:
- Step 106 update the target recognition library according to the new identity identifier and the feature data of at least one unrecognized target image.
- step 106 may specifically include: judging whether at least one unrecognized target image satisfies the preset image quality condition; The unrecognized target images are correspondingly stored in the target recognition library.
- the target recognition library since there is less feature data corresponding to the new ID in the target recognition library, in order to ensure the accuracy of subsequent target re-identification involving the new ID, it is necessary to carry out stricter quality control on the first feature data corresponding to the new ID.
- at least one unrecognized target image corresponding to a new ID comes from a different camera. It is possible that some unrecognized target images have a small size of the original image, image quality problems such as blurred acquisition and environmental occlusion. Identify whether the target image satisfies the preset image quality conditions, so as to comprehensively judge whether it satisfies the first characteristic data of the new ID. In this way, incomplete shooting, occlusion, etc. can be filtered out, and the accuracy of new ID recognition can be improved.
- the above method may further include: according to the first target image or the feature value pair of the first target image
- the feature data of the first target in the target recognition library is dynamically updated. In this way, feature data with high real-time performance can be used for feature matching, which is beneficial to improve the recognition accuracy.
- the feature value of the target image is used to replace the target image for updating, the feature value can be directly used in the subsequent calculation to avoid repeated calculation, greatly reduce the operation time, and ensure the real-time effect.
- the method further includes replacing and updating the target recognition library, specifically including the following three scenarios of replacement and updating: (1) According to each target in the target recognition library The comparison result between the source time and the current time corresponding to the characteristic data of , replace and update the target recognition library. For example, all feature data acquired before a specified time length of the current time can be deleted. It is also possible to delete all the feature data acquired before another specified length of time for one or more targets whose amount of feature data exceeds a threshold. Therefore, the real-time performance of the target recognition library can be ensured, which is beneficial to the subsequent target re-identification.
- the target recognition library is replaced and updated.
- the source of the feature data is the previously obtained target image, so the feature data can correspond to a certain global position according to the target image from which it is derived.
- the distance from the current global position of the target can exceed a certain range.
- feature data is deleted.
- replace and update the target recognition database For example, for each target in the target recognition library, delete two or more feature data whose feature similarity range is higher than a preset value, so as to reduce the feature duplication in the target recognition library.
- the method further includes: after the quantity of characteristic data of any one target exceeds a preset threshold, starting a replacement update. For example, if the preset threshold is set to 100, in the target recognition library, after the number of characteristic data of each target exceeds 100, the replacement and update described in the above embodiment will be started, so as to effectively avoid redundancy while ensuring sufficient characteristic data. .
- an embodiment of the present invention further provides a target re-identification apparatus, which is used to execute the target re-identification method provided by any of the above embodiments.
- FIG. 7 is a schematic structural diagram of a target re-identification apparatus according to an embodiment of the present invention.
- the target re-identification apparatus 700 includes:
- an acquisition module 701 configured to acquire a plurality of current frames collected by a plurality of cameras arranged in the monitoring area;
- a target detection module 702 configured to perform target detection according to multiple current frames, and determine the target image captured by each camera;
- the quantity detection module 703 is used for quantity detection according to the target image captured by each camera to obtain the global target quantity;
- the target re-identification module 704 is used to perform target re-identification according to the target image and the target recognition library, and the target recognition library includes the identity identification and characteristic data of at least one target;
- An identification module 705 configured to determine at least one unrecognized target image according to the result of target re-identification when it is detected that the number of global targets meets the preset increase condition, and create a new identification mark to mark the at least one unrecognized target image;
- the target recognition library updating module 706 is configured to update the target recognition library according to the new identity identifier and the characteristic data of at least one unrecognized target image.
- the target detection module is further configured to: input multiple current frames into the trained target detection model to extract the target image captured by each camera; wherein, the target detection model is based on YOLOv4 - Human detection model created by tiny network.
- the target detection module is further configured to: train the YOLOv4-tiny network according to the real collected images in the monitoring area to obtain a target detection model.
- the target image is a partial image containing target features in the current frame
- the quantity detection module is further configured to: perform position conversion on the captured target image according to the viewing position of each camera to obtain each camera The global position corresponding to the captured target image; determine the global position coincidence of the target images captured by different cameras, filter the target images captured by different cameras according to the global position coincidence, and detect the number of target images retained after screening .
- the quantity detection module is further configured to: when the result of quantity detection is less than the number of previous global targets, then according to the multiple current frames collected by multiple cameras and the previous frame of multiple current frames , to judge whether there is a target that leaves the monitoring area from the predetermined area; if there is no target, the number of previous global targets is still retained as the number of global targets determined this time; The number of global targets; wherein, the number of previous global targets is obtained by performing target detection and quantity detection on the previous frame of multiple current frames.
- the quantity detection module is further configured to: perform projective transformation on the bottom center point of the target image in the current frame according to the viewing position of each camera, so as to determine the ground coordinates of each target image.
- the device is further configured to: input multiple current frames into a trained target quantity detection model to perform target detection and quantity detection to obtain a global target quantity; wherein, the target quantity detection model is based on YOLOv4 -Pedestrian number detection model created by tiny network.
- the target re-identification module is further used to: calculate the similarity between the target image and the feature data in the target recognition library, and perform target re-identification on the target image according to the calculated similarity; When the result of the target re-identification indicates that the first target image matches the first target in the target recognition library, the first target image is marked according to the identity of the first target.
- the identity identification module is further configured to: if the current frame is not the first frame, and the number of global targets corresponding to the current frame increases compared to the number of global targets corresponding to the previous frame, the number of global targets The preset increase conditions are met; if the current frame is the first frame, the default global target number meets the preset increase conditions.
- the target recognition library update module is further configured to: determine whether at least one unrecognized target image satisfies the preset image quality condition; Correspondingly stored in the target recognition library.
- the target recognition library updating module is further configured to: dynamically update the feature data of the first target in the target recognition library according to the first target image or the feature value of the first target image.
- the target recognition library updating module is further configured to: replace and update the target recognition library according to the comparison result between the source time corresponding to the characteristic data of each target in the target recognition library and the current time; and /or, according to the comparison result of the global position corresponding to the feature data of each target in the target recognition library and the current global position of each target, the target recognition library is replaced and updated; and/or, according to each target recognition library The feature similarity between multiple feature data of a target is replaced and updated to the target recognition library.
- the target identification library update module is further configured to: start the replacement update after the quantity of characteristic data of any target exceeds a preset threshold.
- the target re-identification apparatus in this embodiment of the present application can implement each process of the foregoing embodiments of the target re-identification method, and achieve the same effects and functions, which will not be repeated here.
- the apparatus includes: at least one processor; and a memory communicatively connected to the at least one processor;
- the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can execute the method described in the above embodiments.
- a non-volatile computer storage medium of a method for object re-identification having computer-executable instructions stored thereon, the computer-executable instructions being arranged to be executed when executed by a processor: the above-mentioned embodiments the method described.
- the apparatuses, devices, and computer-readable storage media and methods provided in the embodiments of the present application are in one-to-one correspondence. Therefore, the apparatuses, devices, and computer-readable storage media also have beneficial technical effects similar to those of the corresponding methods.
- the beneficial technical effects of the method have been described in detail, therefore, the beneficial technical effects of the apparatus, equipment and computer-readable storage medium will not be repeated here.
- embodiments of the present invention may be provided as a method, system or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.
- computer-usable storage media including, but not limited to, disk storage, CD-ROM, optical storage, etc.
- These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions
- the apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.
- a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.
- processors CPUs
- input/output interfaces network interfaces
- memory volatile and non-volatile memory
- Memory may include forms of non-persistent memory, random access memory (RAM) and/or non-volatile memory in computer readable media, such as read only memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.
- RAM random access memory
- ROM read only memory
- flash RAM flash memory
- Computer-readable media includes both persistent and non-permanent, removable and non-removable media, and storage of information may be implemented by any method or technology.
- Information may be computer readable instructions, data structures, modules of programs, or other data.
- Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Flash Memory or other memory technology, Compact Disc Read Only Memory (CD-ROM), Digital Versatile Disc (DVD) or other optical storage, Magnetic tape cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device.
- PRAM phase-change memory
- SRAM static random access memory
- DRAM dynamic random access memory
- RAM random access memory
- ROM read only memory
- EEPROM Electrically Erasable Programmable Read Only Memory
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Human Computer Interaction (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Signal Processing (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
- Closed-Circuit Television Systems (AREA)
- Transition And Organic Metals Composition Catalysts For Addition Polymerization (AREA)
Abstract
本发明提供了一种目标重识别方法、装置及计算机可读存储介质,该方法包括:获取设置于监控区域内的多个摄像头采集的多个当前帧;根据多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;根据每个摄像头捕获到的目标图像进行数量检测,得到全局目标数量;根据目标图像和目标识别库进行目标重识别,目标识别库包括至少一个目标的身份标识和特征数据;当检测到全局目标数量符合预设增加条件时,根据目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识对至少一个未识别目标图像进行标记;根据新的身份标识和至少一个未识别目标图像的特征数据更新目标识别库。利用上述方法,能够提供目标重识别的准确度和稳定性。
Description
本申请要求于2021年01月25日提交的、申请号为202110095415.1、标题为“一种目标重识别方法、装置及计算机可读存储介质”的中国专利申请的优先权,该中国专利申请的公开内容以引用的方式并入本文。
本发明属于识别领域,具体涉及一种目标重识别方法、装置及计算机可读存储介质。
本部分旨在为权利要求书中陈述的本发明的实施方式提供背景或上下文。此处的描述不因为包括在本部分中就承认是现有技术。
目前,随着视频监控技术的普及以及不断提升的安防需求,应用于视频监控领域中的目标重识别逐渐成为计算机视觉研究领域的热点之一。
在诸如数据中心、商场等的安全需求较高的监控场所中实现跨摄像头的目标重识别显得非常重要。在目标重识别过程中,当新的目标进入监控区域时,需要为该新的目标分配一个新ID以便后续进行识别,业界通常采用计算目标图像和目标识别库中的特征数据之间的特征相似度的方法来判断是否创建并分配新ID,而有些场景下,由于目标遮挡、拍摄角度等问题会对上述判断的准确度造成较大的影响,进而可能造成目标重识别不准确的问题。
发明内容
针对上述现有技术中存在的问题,提出了一种目标重识别方法、装置及计算机可读存储介质,利用这种方法、装置及计算机可读存储介质,能够解决上述问题。
本发明提供了以下方案。
第一方面,提供一种目标重识别方法,包括:获取设置于监控区域内的多个摄像头采集的多个当前帧;根据多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;根据每个摄像头捕获到的目标图像进行数量检测,得到全局目标数量;根据目标图像和目标识别库进行目标重识别,目标识别库包括至少一个目标的身份标识和特征数据;当检测到全局目标 数量符合预设增加条件时,根据目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识对至少一个未识别目标图像进行标记;根据新的身份标识和至少一个未识别目标图像的特征数据更新目标识别库。
在一种可能的实施方式中,根据多个当前帧进行目标检测,还包括:将多个当前帧输入经训练的目标检测模型,以提取出每个摄像头捕获到的目标图像;其中,目标检测模型为基于YOLOv4-tiny网络创建的人体检测模型。
在一种可能的实施方式中,方法还包括:根据监控区域内的真实采集图像对YOLOv4-tiny网络进行训练,得到目标检测模型。
在一种可能的实施方式中,目标图像为当前帧中包含目标特征的局部图像,根据每个摄像头捕获到的目标图像进行数量检测,还包括:根据每个摄像头的取景位置对捕获到的目标图像进行位置转换,得到每个摄像头捕获到的目标图像对应的全局位置;确定由不同摄像头各自捕获的目标图像的全局位置重合度,根据全局位置重合度对不同摄像头各自捕获的目标图像进行筛选,检测筛选后保留的目标图像的数量。
在一种可能的实施方式中,方法还包括:当数量检测的结果少于在先全局目标数量时,则根据多个摄像头采集的多个当前帧和多个当前帧的上一帧,判断是否存在从预定区域离开监控区域的目标;若不存在目标,则仍然保留在先全局目标数量作为本次确定的全局目标数量;若存在目标,则将数量检测的结果作为本次确定的全局目标数量;其中,在先全局目标数量根据对多个当前帧的上一帧进行目标检测和数量检测得到。
在一种可能的实施方式中,根据每个摄像头的取景位置对捕获到的目标图像进行位置转换,还包括:根据每个摄像头的取景位置对当前帧中的目标图像的底部中心点进行投影变换,从而确定每个目标图像的地面坐标。
在一种可能的实施方式中,方法还包括:将多个当前帧输入经训练的目标数量检测模型,以执行目标检测和数量检测,得到全局目标数量;其中,目标数量检测模型为基于YOLOv4-tiny网络创建的行人数量检测模型。
在一种可能的实施方式中,根据目标图像和目标识别库进行目标重识别,还包括:计算目标图像与目标识别库中的特征数据之间的相似度,并依据计算得到的相似度,对目标图像进行目标重识别;当目标重识别的结果指示第一目标图像与目标识别库中的第一目标匹配时,根据第一目标的身份标识对第一目标图像进行标记。
在一种可能的实施方式中,方法还包括:若当前帧为非首帧,且当前帧对应的全局目标数量相较于上一帧对应的全局目标数量增加时,则全局目标数量符合预设增加条件;若当前帧为首帧时,默认全局目标数量符合预设增加条件。
在一种可能的实施方式中,根据新的身份标识和至少一个未识别目标图像的特征数据更新目标识别库,还包括:判断至少一个未识别目标图像是否满足预设图像质量条件;将新的身份标识和满足预设图像质量条件的未识别目标图像对应存入目标识别库。
在一种可能的实施方式中,根据目标图像和目标识别库进行目标重识别之后,方法还包括:根据第一目标图像或第一目标图像的特征值对目标识别库中的第一目标的特征数据进行动态更新。
在一种可能的实施方式中,方法还包括对目标识别库进行替换更新,具体包括:根据目标识别库中的每个目标的特征数据对应的来源时间和当前时间的比较结果,对目标识别库进行替换更新;和/或,根据目标识别库中的每个目标的特征数据对应的全局位置和每个目标的当前全局位置的比较结果,对目标识别库进行替换更新;和/或,根据目标识别库中的每个目标的多个特征数据之间的特征相似度,对目标识别库进行替换更新。
在一种可能的实施方式中,方法还包括:任意一个目标的特征数据的数量超过预设阈值之后,启动替换更新。
第二方面,提供一种目标重识别装置,包括:获取模块,用于获取设置于监控区域内的多个摄像头采集的多个当前帧;目标检测模块,用于根据多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;数量检测模块,用于根据每个摄像头捕获到的目标图像进行数量检测,得到全局目标数量;目标重识别模块,用于根据目标图像和目标识别库进行目标重识别,目标识别库包括至少一个目标的身份标识和特征数据;身份标识模块,用于当检测到全局目标数量符合预设增加条件时,根据目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识对至少一个未识别目标图像进行标记;目标识别库更新模块,用于根据新的身份标识和至少一个未识别目标图像的特征数据更新目标识别库。
在一种可能的实施方式中,目标检测模块,还用于:将多个当前帧输入经训练的目标检测模型,以提取出每个摄像头捕获到的目标图像;其中,目标检测模型为基于YOLOv4-tiny网络创建的人体检测模型。
在一种可能的实施方式中,目标检测模块,还用于:根据监控区域内的真实采集图像对YOLOv4-tiny网络进行训练,得到目标检测模型。
在一种可能的实施方式中,目标图像为当前帧中包含目标特征的局部图像,数量检测模块还用于:根据每个摄像头的取景位置对捕获到的目标图像进行位置转换,得到每个摄像头捕获到的目标图像对应的全局位置;确定由不同摄像头各自捕获的目标图像的全局位置重合度,根据全局位置重合度对不同摄像头各自捕获的目标图像进行筛选,检测筛选后保留的目标图像的数量。
在一种可能的实施方式中,数量检测模块还用于:当数量检测的结果少于在先全局目标数量时,则根据多个摄像头采集的多个当前帧和多个当前帧的上一帧,判断是否存在从预定区域离开监控区域的目标;若不存在目标,则仍然保留在先全局目标数量作为本次确定的全局目标数量;若存在目标,则将数量检测的结果作为本次确定的全局目标数量;其中,在先全局目标数量根据对多个当前帧的上一帧进行目标检测和数量检测得到。
在一种可能的实施方式中,数量检测模块还用于:根据每个摄像头的取景位置对当前帧中的目标图像的底部中心点进行投影变换,从而确定每个目标图像的地面坐标。
在一种可能的实施方式中,装置还用于:将多个当前帧输入经训练的目标数量检测模型,以执行目标检测和数量检测,得到全局目标数量;其中,目标数量检测模型为基于YOLOv4-tiny网络创建的行人数量检测模型。
在一种可能的实施方式中,目标重识别模块还用于:计算目标图像与目标识别库中的特征数据之间的相似度,并依据计算得到的相似度,对目标图像进行目标重识别;当目标重识别的结果指示第一目标图像与目标识别库中的第一目标匹配时,根据第一目标的身份标识对第一目标图像进行标记。
在一种可能的实施方式中,身份标识模块还用于:若当前帧为非首帧,且当前帧对应的全局目标数量相较于上一帧对应的全局目标数量增加时,则全局目标数量符合预设增加条件;若当前帧为首帧时,默认全局目标数量符合预设增加条件。
在一种可能的实施方式中,目标识别库更新模块还用于:判断至少一个未识别目标图像是否满足预设图像质量条件;将新的身份标识和满足预设图像质量条件的未识别目标图像对应存入目标识别库。
在一种可能的实施方式中,目标识别库更新模块还用于:根据第一目标图像或第一目标图像的特征值对目标识别库中的第一目标的特征数据进行动态更新。
在一种可能的实施方式中,目标识别库更新模块还用于:根据目标识别库中的每个目标的特征数据对应的来源时间和当前时间的比较结果,对目标识别库进行替换更新;和/或,根据目标识别库中的每个目标的特征数据对应的全局位置和每个目标的当前全局位置的比较结果,对目标识别库进行替换更新;和/或,根据目标识别库中的每个目标的多个特征数据之间的特征相似度,对目标识别库进行替换更新。
在一种可能的实施方式中,目标识别库更新模块还用于:任意一个目标的特征数据的数量超过预设阈值之后,启动替换更新。
第三方面,提供一种目标重识别装置,包括:一个或者多个多核处理器;存储器,用于存储一个或多个程序;当所述一个或多个程序被所述一个或者多个多核处理器执行时,使得所述一个或多个多核处理器实现:如第一方面的方法。
第四方面,提供一种计算机可读存储介质,所述计算机可读存储介质存储有程序,当所述程序被多核处理器执行时,使得所述多核处理器执行如第一方面的方法。
本申请实施例采用的上述至少一个技术方案能够达到以下有益效果:本实施例中,检测监控区域内的全局目标数量,通过检测到的全局目标数量对新的身份标识的创建和分配进行控制,能够很好地保证身份标识的分配准确,保证目标重识别的准确度和稳定性。。
应当理解,上述说明仅是本发明技术方案的概述,以便能够更清楚地了解本发明的技术手段,从而可依照说明书的内容予以实施。为了让本发明的上述和其它目的、特征和优点能够更明显易懂,以下特举例说明本发明的具体实施方式。
通过阅读下文的示例性实施例的详细描述,本领域普通技术人员将明白本文所述的优点和益处以及其他优点和益处。附图仅用于示出示例性实施例的目的,而并不认为是对本发明的限制。而且在整个附图中,用相同的标号表示相同的部件。在附图中:
图1为根据本发明一实施例的目标重识别方法的流程示意图;
图2为根据本发明一实施例的监控区域的地面示意图;
图3为根据本发明一实施例的多个摄像头的取景画面示意图;
图4为根据本发明一实施例的多个摄像头的当前帧的示意图;
图5为根据本发明一实施例的多个摄像头捕获的目标图像的示意图;
图6为根据本发明一实施例的多个摄像头捕获的目标图像的全局位置的示意图;
图7为根据本发明一实施例的目标重识别装置的结构示意图;
图8为根据本发明另一实施例的目标重识别装置的结构示意图。
在附图中,相同或对应的标号表示相同或对应的部分。
下面将参照附图更详细地描述本公开的示例性实施例。虽然附图中显示了本公开的示例性实施例,然而应当理解,可以以各种形式实现本公开而不应被这里阐述的实施例所限制。相反,提供这些实施例是为了能够更透彻地理解本公开,并且能够将本公开的范围完整的传达给本领域的技术人员。
在本申请实施例的描述中,应理解,诸如“包括”或“具有”等术语旨在指示本说明书中所公开的特征、数字、步骤、行为、部件、部分或其组合的存在,并且不旨在排除一个或多个其他特征、数字、步骤、行为、部件、部分或其组合存在的可能性。
除非另有说明,“/”表示或的意思,例如,A/B可以表示A或B;本文中的“和/或”仅仅是一种描述关联对象的关联关系,表示可以存在三种关系,例如,A和/或B,可以表示:单独存在A,同时存在A和B,单独存在B这三种情况。
术语“第一”、“第二”等仅用于描述目的,而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此,限定有“第一”、“第二”等的特征可以明示或者隐含地包括一个或者更多个该特征。在本申请实施例的描述中,除非另有说明,“多个”的含义是两个或两个以上。
本申请中的所有代码都是示例性的,本领域技术人员根据所使用的编程语言,具体的需求和个人习惯等因素会在不脱离本申请的思想的条件下想到各种变型。
目标实时跟踪方法,其特征在于,包括:获取设置于监控区域内的多个摄像头采集的多个当前帧;根据所述多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;根据每个摄像头捕获到的所述目标图像进行数量检测,得到全局目标数量;根据所述目标图像和目标识别库进行目标重识别,其中所述目标识别库包括至少一个目标的身份标识和特征数据;当检测到所述全局目标数量符合预设增加条件时,根据所述目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识对所述至少一个未识别目标图像进行标记;根据所述新的身份标识和所述至少一个未识别目标图像的特征数据更新所述目标识别库。
另外还需要说明的是,在不冲突的情况下,本发明中的实施例及实施例中的特征可以相互组合。下面将参考附图并结合实施例来详细说明本发明。
图1为根据本申请一实施例的实时目标跟踪方法的流程示意图,用于跟踪预设场景内的指定目标,在该流程中,从设备角度而言,执行主体可以是一个或者多个电子设备;从程序角度而言,执行主体相应地可以是搭载于这些电子设备上的程序。
如图1所示,该方法100包括:
步骤101、获取设置于监控区域内的多个摄像头采集的多个当前帧;
具体地,监控区域是指多个摄像头的取景区域的总和,多个摄像头包括至少两个摄像头,并且上述多个摄像头的取景区域彼此相邻接或至少部分地重叠,从而待跟踪的目标能够在监控区域中移动进而出现在任意一个或多个摄像头的取景区域内。其中,从多个摄像头的监控视频中分别提取多个摄像头的当前帧,其中每个摄像头的当前帧具有相同的采集时间。 可选地,本公开中的跟踪目标优选为行人,本领域技术人员可以理解,上述跟踪目标也可以是其他可移动的物体,比如动物、车辆等,本公开对此不作具体限制。
例如,在复杂监控场景下,比如在楼道、大型商场、机房等场所,通常会使用大量的摄像头对各个区域进行监控,并得到多路监控视频。图2示出一种示意性监控场景,在该监控场景中设置有摄像头201和摄像头202,如图3示出上述摄像头201和摄像头202的取景画面。其中,摄像头201采集的监控视频可解析为图像帧序列(A
1,A
2,...,A
N),摄像头202采集的监控视频可解析为图像帧序列(B
1,B
2,...,B
N),其中上述解析可以实时在线进行或离线进行。基于此,可以按时序从上述多个图像帧序列中依次提取两个摄像头的当前帧A
n和B
n以进行本公开所示出的实时目标跟踪,其中,下标n的取值可以是n=1,2,…,N。
如图1所示,该方法100可以包括:
步骤102、根据多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;
具体地,目标图像可以是当前帧中包含目标特征的局部图像。例如,如图4所示,示出了摄像头201和摄像头202的当前帧A
n和B
n,然后,在任意基于深度学习的目标检测模型中输入预处理后的当前帧A
n和B
n进行检测,输出针对每个摄像头的一系列行人图像(目标图像的一种示例)。目标检测模型比如可以是YOLO(统一实时目标检测,You Only Look Once)模型等,本公开对此不作具体限制。如图5所示,示出了对多个当前帧A
n和B
n进行检测得到的多个目标(行人)检测框,可以理解,根据目标(行人)检测框从当前帧中可以截取到目标(行人)图像,其中摄像头201捕获到的目标(行人)图像包括(a
1,a
2,a
3),摄像头202捕获到的目标(行人)包括图像(b)。可以将截取出的目标(行人)图像进行归一化处理,以便于后续的跟踪展示。
进一步地,在一种可能的实施方式中,为了更准确地检测到目标图像,步骤102还可以包括:将多个当前帧输入经训练的目标检测模型,以提取出每个摄像头捕获到的目标图像;其中,目标检测模型为基于YOLOv4-tiny网络创建的人体检测模型。
具体地,可以基于深度学习的实时目标检测算法YOLOV4-TINY进行改进得到YOLOV4-TINY-P(YOLOv4-tiny-People),并训练生成人体检测模型,利用该人体检测模型可以针对行人整体特征进行识别,且不受佩戴口罩等脸部遮挡的影响。此外,无需专业的人脸摄像头,利用多个普通的监控摄像头就可以直接完成上述目标检测。
可选地,也可以采用基于其他目标检测算法,比如faster-rcnn目标检测算法、yolov4目标检测算法等,本申请对此不作具体限制。
可选地,针对诸如车辆检测、动物检测等其他目标检测场景,可以对应采用其他目标检测模型,本申请对此不作具体限制。
进一步地,在一些实施方式中,为了使目标检测模型针对具体监控场景时仍可以保持高准确度,还可以执行以下步骤以获取上述目标检测模型:根据监控区域内的真实采集图像对YOLOv4-tiny网络进行训练,得到目标检测模型。
例如,当应用于机房场景时,可以对诸如机房的实际场景中的行人进行针对性训练,基于实际场景增设目标正负样本,例如,椅子、背包、服务器等物品为负样本,行人为正样本,从而避免由于光线原因将远处的背包和椅子杂物等物体误识别成不同形态的行人的情况。训练数据采用可以采用实际机房场景数据、PASCAL VOC2007和VOC2012等目标检测数据集联合训练,进一步提升模型检测能力。
如图1所示,该方法100还包括:
步骤103、根据每个摄像头捕获到的目标图像进行数量检测,得到全局目标数量。
可以利用任何可能的目标统计方法进行上述数量检测,本申请对此不作具体限制。
比如,可以先单独检测每个摄像头中的局部目标数量,将多个局部目标数量累加后再分析出不同摄像头中捕获到得到重合的目标图像,并进行对应删减。参考图5,在摄像头201捕获到三个目标(行人)图像,包括(a
1,a
2,a
3),摄像头202捕获到的一个目标(行人)图像,包括(b),将局部目标数量累加,由于不同摄像头之间存在取景范围的交叉,因此必然存在由不同摄像头拍摄到了同一目标的不同角度的目标图像的情况,可以通过位置分析判断摄像头201捕获到的a
3与摄像头202捕获到的(b)重合,从局部目标数量累加结果中删减重合的数量,由此可以得到全局目标数量为3个。
在一些实施方式中,为了准确获取监控区域内的全局目标数量,步骤103还可以包括以下步骤:根据每个摄像头的取景位置对捕获到的目标图像进行位置转换,得到每个摄像头捕获到的目标图像对应的全局位置;确定由不同摄像头各自捕获的目标图像的全局位置重合度,根据全局位置重合度对目标图像进行筛选;根据筛选后保留的目标图像的数量确定监控区域内的全局目标数量。
可以理解,目标图像为当前帧中包含目标特征的局部图像,通过局部图像和当前帧的位置关系以及对应摄像头的取景范围进行简单的位置计算,即可获知目标图像的全局位置和监控区域内的全局目标数量。
参考图5,在摄像头201捕获到目标图像(a
1,a
2,a
3),摄像头202捕获到的目标图像(b),根据摄像头201的取景位置对捕获到的目标图像(a
1,a
2,a
3)进行位置转换,根据摄像头202的取景位置对捕获到的目标图像(b)进行位置转换,得到图6示出的每个目标图像的全局位置,可以看出,摄像头201捕获到的目标图像a
3和摄像头202捕获到的目标图像b的全局位 置重合度很高,假设其超过预设的重合度阈值,即可认为目标图像a
3和b实际为同一目标,可以仅保留一个,进而可以判断监控区域中的全局目标数量为3个。
在一些实施方式中,进一步地,由于监控区域中可能出现背景遮挡目标的情况,从而导致检测到的全局目标数量相较于实际数量减少的情况,基于此,还可以执行以下步骤:当数量检测的结果少于在先全局目标数量时,则根据多个摄像头采集的多个当前帧和多个当前帧的上一帧,判断是否存在从预定区域离开监控区域的目标;其中,若不存在从预定区域离开监控区域的目标,则仍然保留在先全局目标数量作为本次确定的全局目标数量;若存在从预定区域离开监控区域的目标,则将数量检测的结果作为本次确定的全局目标数量;
其中,在先全局目标数量是根据对多个上一帧进行目标检测和数量检测得到。具体地,将步骤101-步骤103中的多个当前帧替换为多个当前帧的上一帧,即可利用相同的方案获得该在先全局目标数量,本申请不再赘述。
例如,假设在对多个摄像头采集的多个当前帧的上一帧进行目标实时跟踪时,检测到监控区域中的全局目标数量为5,也即总共包括5个目标对象。在对多个摄像头采集的多个当前帧进行目标实时跟踪时,数量检测的结果指示监控区域内仅包含4个目标对象,相较于上一帧发生了数量减少,则需要考虑是否存在暂时的目标遮挡情况。具体可以将诸如监控区域的出口区域划分为预定区域,判断是否存在一个目标,根据多个当前帧的上一帧能够确定该目标位于该出口区域,且根据多个当前帧能够确定该目标从该出口区域消失。如果存在这样的目标,则可以认为真实发生了目标离开监控区域的情况,可以将上述数量检测的结果作为全局目标数量。相反,如果不存在这样的目标对象,则可以认为存在目标遮挡等情况,仍然保留在先全局目标数量作为本次确定的全局目标数量。
在一些实施方式中,根据每个摄像头的取景位置对捕获到的目标图像进行位置转换,还包括:根据每个摄像头的取景位置对当前帧中的目标图像的底部中心点进行投影变换,从而确定每个目标图像的地面坐标。这样,可以将每个摄像头取景范围内捕获的待识别目标组合到统一的坐标系中。
例如,可以获取图5中每个摄像头捕获的每个目标图像的底部中心点位置,对该每个目标图像的底部中心点进行转换,得到待识别目标在监控场景中的实际地面位置,图6示出了通过投影转换获得的每个目标图像对应的地面坐标。具体而言,可以看出,每个摄像头视角下的地面过道是一个近似梯形区域,因此针对每个摄像头捕获的目标图像,首先可以通过梯形-矩形转换得到每个目标图像的底部中心点在标准矩形区域中的坐标,其次根据监控场景的实际布局对标准矩形区域进行旋转,通过旋转矩阵计算得到每个目标图像的底部中心点 的旋转后坐标,最后根据监控场景的实际布局对旋转后坐标进行平移和缩放,得到最终的地面坐标位置。
进一步地,在一些实施方式中,可以预先训练获得目标数量检测模型,以用于实时检测监控区域的全局目标数量的,在执行目标实时跟踪方法时,将多个当前帧输入经训练的目标数量检测模型以执行目标检测和数量检测,直接得到全局目标数量。
例如,可以通过改进基于深度学习的实时目标检测算法YOLOV4-TINY,提出改进的人数统计算法YOLOv4-TINY-PC(YOLOv4-tiny-People Counting),其中YOLOV4-TINY算法不具备多摄像头的人数统计能力,YOLOv4-TINY-PC可以实时得到监控区域内人数信息以统计人流量。具体地,该目标数量检测模型可以通过行人检测算法(YOLOv4-TINY-P)得到各个摄像头识别的目标图像,对目标图像进行位置转换,得到在整体监控区域内的全局位置坐标。对机房内各个摄像头区域进行划分,对机房摄像头分为主摄像头和辅摄像头,对各个摄像头的数量检测结果进行筛选,使得彼此无重合,得到最终当前帧的所有摄像头中的目标数量,即为全局目标数量。
在本实施例中,目标数量检测模型为基于YOLOv4-tiny网络创建的行人数量检测模型。可选地,也可以基于诸如faster-rcnn、yolov4其他网络创建行人数量检测模型。可选地,也可以针对其他应用场景创建诸如车辆数量检测模型、动物数量检测模型的目标数量检测模型。
如图1所示,该方法还包括:
步骤104、根据目标图像和目标识别库进行目标重识别。
其中,目标识别库包括至少一个目标的身份标识和特征数据。例如,目标识别库可以包括{目标1:特征数据1,…,特征数据N};{目标2:特征数据1,…,特征数据N},诸如此类。
进一步地,在一种可能的实施方式中,在步骤104之后,还可以包括:计算目标图像与目标识别库中的特征数据之间的相似度,并依据计算得到的相似度,对目标图像进行目标重识别;当目标重识别的结果指示第一目标图像与目标识别库中的第一目标匹配时,对第一目标图像标识第一目标的身份标识。
例如,参考图5,计算示出的行人图像b和目标识别库包含的每个目标的特征数据之间的相似度,假设行人图像b和目标1的特征数据之间的相似度最高,且该相似度超过预设匹配阈值,则可以认为目标重识别的结果指示行人图像b与目标识别库中的目标1匹配,进一步可以将该行人图像b标识为目标1。基于相似的做法,行人图像a
2匹配到目标识别库中的目标2并进行标识。行人图像a
3同样匹配目标识别库中的目标1并进行标识。
如图1所示,该方法还包括:
步骤105、当检测到全局目标数量符合预设增加条件时,根据目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识(以下简称新ID)对至少一个未识别目标图像进行标记。
可以理解,当新的目标进入监控区域时,需要为该新的目标分配一个新ID,业界通常采用计算目标图像和目标识别库中的特征数据之间的特征相似度的方法来判断是否创建并分配新ID,而有些场景下,由于目标遮挡、拍摄角度等问题会对上述判断的准确度造成较大的影响。比如,当已经处于监控区域内的某个目标由于其目标图像的拍摄质量较差,导致其目标识别库中的对应特征数据无法匹配,则容易将其误认为是新的目标。与本实施例中,只有当检测到的全局目标数量符合预设增长条件时,比如当全局目标数量相较于根据多个当前帧的前一帧进行目标数量检测得到的先前目标数量增加时,才会生成新的ID,通过全局目标数量对新ID的创建和分配进行控制,能够很好保证身份标识数量的增长准确,保证稳定性。
基于此,当检测到的全局目标数量符合预设增长条件时,进一步根据目标重识别的结果确定至少一个未识别目标图像。例如,参考图5,假设目标识别库可以包括{目标1:特征数据1,…,特征数据N};{目标2:特征数据1,…,特征数据N},该目标重识别的结果指示行人图像b匹配到目标1并进行标识。行人图像a
2匹配到目标2并进行标识。行人图像a
3同样匹配到目标1并进行标识。此时,行人图像a
1并未匹配到目标识别库中的任何一个目标,也即行人图像a
1为上述目标重识别过程中确定的未识别目标图像,进一步地,可以根据创建新ID(比如目标3)并对该行人图像a
1进行标记。由此可以实现为新的目标分配一个新ID。
值得注意的是,由于目标识别库是不断进行淘汰更新的,上述新的目标是指当前的目标识别库中未存放与其相匹配的身份标识和特征数据的目标。换言之,如果一个行人在先前进入过该监控区域并离开,仍然可能在下次进入该监控区域时作为新的目标,需要重新为该新的目标分配新创建的身份标识并相应存入特征数据。
在一种可能的实施方式中,进一步地,步骤105还可以包括检测全局目标数量是否符合预设增加条件的步骤,具体包括:若当前帧为非首帧,且当前帧对应的全局目标数量相较于多个当前帧的前一帧对应的在先全局目标数量增加时,则全局目标数量符合预设增加条件。若当前帧为首帧时,默认全局目标数量符合预设增加条件。具体地,已经在前文中描述了该在先全局目标数量,此处不再赘述。
如图1所示,该方法还包括:
步骤106、根据新的身份标识和至少一个未识别目标图像的特征数据更新目标识别库。
在一种实施方式中,为了提升新ID的识别准确性,步骤106可以具体包括:判断至少一个未识别目标图像是否满足预设图像质量条件;将新的身份标识和满足预设图像质量条件的未识别目标图像对应存入目标识别库。
可以理解,由于在目标识别库中,新ID对应的特征数据较少,为了保证后续涉及新ID的目标重识别准确度,需要对新ID对应的首项特征数据进行较为严格的质量控制。比如,某个新ID对应的至少一个未识别目标图像来源自不同的摄像头,可能某些未识别目标图像存在原始图像尺寸较小,采集模糊、环境遮挡等图像质量问题,判断新ID对应的未识别目标图像是否满足预设图像质量条件,从而综合判断其是否满足成为新ID的首个特征数据。这样,可以过滤掉拍摄不完整,遮挡等情况,提升新ID识别的准确性。
在一种实施方式中,进一步地,在上述步骤103之后,为了保证该目标识别库的实时性和避免冗余,上述方法还可以包括:根据第一目标图像或第一目标图像的特征值对目标识别库中的第一目标的特征数据进行动态更新。这样可以利用实时性较高的特征数据进行特征匹配,有利于提升识别准确度。
可以理解,当采用目标图像的特征值替代目标图像进行更新后,在后续计算中可以直接采用特征值,避免重复计算,极大减少了运算时间,保证了实时效果。
在一种实施方式中,为避免目标识别库产生特征冗余,方法还包括对目标识别库进行替换更新,具体包括以下三种替换更新的场景:(1)根据目标识别库中的每个目标的特征数据对应的来源时间和当前时间的比较结果,对目标识别库进行替换更新。比如,可以将当前时间的指定时间长度之前获取的全部特征数据予以删除。也可以针对特征数据数量超过阈值的一个或多个目标,将其在另一指定时间长度之前获取的全部特征数据予以删除。由此可以保证目标识别库的实时性,有利于后续的目标重识别。(2)根据目标识别库中的每个目标的特征数据对应的全局位置和每个目标的当前全局位置的比较结果,对目标识别库进行替换更新。比如,可以理解,特征数据的来源为先前获得的目标图像,因此特征数据可以根据其来源的目标图像对应到某个全局位置,针对一个或多个目标,可以将距离目标当前全局位置超过一定范围的特征数据进行删除。(3)根据目标识别库中的每个目标的多个特征数据之间的特征相似度,对目标识别库进行替换更新。比如,针对目标识别库中的每个目标,对特征相似程高于预设值的两个或以上特征数据进行删减,以减少目标识别库中的特征重复。
在一种实施方式中,方法还包括:任意一个目标的特征数据的数量超过预设阈值之后,启动替换更新。例如,设置该预设阈值为100,在目标识别库中,每个目标的特征数据数量超过100之后,开始进行上述实施例描述的替换更新,在保证特征数据足量的同时有效避免出现冗余。
关于本申请实施例的方法流程图,将某些操作描述为以一定顺序执行的不同的步骤。这样的流程图属于说明性的而非限制性的。可以将在本文中所描述的某些步骤分组在一起并且在单个操作中执行、可以将某些步骤分割成多个子步骤、并且可以以不同于在本文中所示出的顺序来执行某些步骤。可以由任何电路结构和/或有形机制(例如,由在计算机设备上运行的软件、硬件(例如,处理器或芯片实现的逻辑功能)等、和/或其任何组合)以任何方式来实现在流程图中所示出的各个步骤。
基于相同的技术构思,本发明实施例还提供一种目标重识别装置,用于执行上述任一实施例所提供的目标重识别方法。图7为本发明实施例提供的一种目标重识别装置结构示意图。
如图7所示,目标重识别装置700包括:
获取模块701,用于获取设置于监控区域内的多个摄像头采集的多个当前帧;
目标检测模块702,用于根据多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;
数量检测模块703,用于根据每个摄像头捕获到的目标图像进行数量检测,得到全局目标数量;
目标重识别模块704,用于根据目标图像和目标识别库进行目标重识别,目标识别库包括至少一个目标的身份标识和特征数据;
身份标识模块705,用于当检测到全局目标数量符合预设增加条件时,根据目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识对至少一个未识别目标图像进行标记;
目标识别库更新模块706,用于根据新的身份标识和至少一个未识别目标图像的特征数据更新目标识别库。
在一种可能的实施方式中,目标检测模块,还用于:将多个当前帧输入经训练的目标检测模型,以提取出每个摄像头捕获到的目标图像;其中,目标检测模型为基于YOLOv4-tiny网络创建的人体检测模型。
在一种可能的实施方式中,目标检测模块,还用于:根据监控区域内的真实采集图像对YOLOv4-tiny网络进行训练,得到目标检测模型。
在一种可能的实施方式中,目标图像为当前帧中包含目标特征的局部图像,数量检测模块还用于:根据每个摄像头的取景位置对捕获到的目标图像进行位置转换,得到每个摄像头捕获到的目标图像对应的全局位置;确定由不同摄像头各自捕获的目标图像的全局位置重 合度,根据全局位置重合度对不同摄像头各自捕获的目标图像进行筛选,检测筛选后保留的目标图像的数量。
在一种可能的实施方式中,数量检测模块还用于:当数量检测的结果少于在先全局目标数量时,则根据多个摄像头采集的多个当前帧和多个当前帧的上一帧,判断是否存在从预定区域离开监控区域的目标;若不存在目标,则仍然保留在先全局目标数量作为本次确定的全局目标数量;若存在目标,则将数量检测的结果作为本次确定的全局目标数量;其中,在先全局目标数量根据对多个当前帧的上一帧进行目标检测和数量检测得到。
在一种可能的实施方式中,数量检测模块还用于:根据每个摄像头的取景位置对当前帧中的目标图像的底部中心点进行投影变换,从而确定每个目标图像的地面坐标。
在一种可能的实施方式中,装置还用于:将多个当前帧输入经训练的目标数量检测模型,以执行目标检测和数量检测,得到全局目标数量;其中,目标数量检测模型为基于YOLOv4-tiny网络创建的行人数量检测模型。
在一种可能的实施方式中,目标重识别模块还用于:计算目标图像与目标识别库中的特征数据之间的相似度,并依据计算得到的相似度,对目标图像进行目标重识别;当目标重识别的结果指示第一目标图像与目标识别库中的第一目标匹配时,根据第一目标的身份标识对第一目标图像进行标记。
在一种可能的实施方式中,身份标识模块还用于:若当前帧为非首帧,且当前帧对应的全局目标数量相较于上一帧对应的全局目标数量增加时,则全局目标数量符合预设增加条件;若当前帧为首帧时,默认全局目标数量符合预设增加条件。
在一种可能的实施方式中,目标识别库更新模块还用于:判断至少一个未识别目标图像是否满足预设图像质量条件;将新的身份标识和满足预设图像质量条件的未识别目标图像对应存入目标识别库。
在一种可能的实施方式中,目标识别库更新模块还用于:根据第一目标图像或第一目标图像的特征值对目标识别库中的第一目标的特征数据进行动态更新。
在一种可能的实施方式中,目标识别库更新模块还用于:根据目标识别库中的每个目标的特征数据对应的来源时间和当前时间的比较结果,对目标识别库进行替换更新;和/或,根据目标识别库中的每个目标的特征数据对应的全局位置和每个目标的当前全局位置的比较结果,对目标识别库进行替换更新;和/或,根据目标识别库中的每个目标的多个特征数据之间的特征相似度,对目标识别库进行替换更新。
在一种可能的实施方式中,目标识别库更新模块还用于:任意一个目标的特征数据的数量超过预设阈值之后,启动替换更新。
需要说明的是,本申请实施例中的目标重识别装置可以实现前述目标重识别方法的实施例的各个过程,并达到相同的效果和功能,这里不再赘述。
图8为根据本申请一实施例的目标重识别装置,用于执行图1所示出的目标重识别方法,该装置包括:至少一个处理器;以及,与至少一个处理器通信连接的存储器;其中,存储器存储有可被至少一个处理器执行的指令,指令被至少一个处理器执行,以使至少一个处理器能够执行上述实施例所述的方法。
根据本申请的一些实施例,提供了目标重识别方法的非易失性计算机存储介质,其上存储有计算机可执行指令,该计算机可执行指令设置为在由处理器运行时执行:上述实施例所述的方法。
本申请中的各个实施例均采用递进的方式描述,各个实施例之间相同相似的部分互相参见即可,每个实施例重点说明的都是与其他实施例的不同之处。尤其,对于装置、设备和计算机可读存储介质实施例而言,由于其基本相似于方法实施例,所以其描述进行了简化,相关之处可参见方法实施例的部分说明即可。
本申请实施例提供的装置、设备和计算机可读存储介质与方法是一一对应的,因此,装置、设备和计算机可读存储介质也具有与其对应的方法类似的有益技术效果,由于上面已经对方法的有益技术效果进行了详细说明,因此,这里不再赘述装置、设备和计算机可读存储介质的有益技术效果。
本领域内的技术人员应明白,本发明的实施例可提供为方法、系统或计算机程序产品。因此,本发明可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且,本发明可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本发明是参照根据本发明实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器,使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中,使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品,该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上,使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理,从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
在一个典型的配置中,计算设备包括一个或多个处理器(CPU)、输入/输出接口、网络接口和内存。
内存可能包括计算机可读介质中的非永久性存储器,随机存取存储器(RAM)和/或非易失性内存等形式,如只读存储器(ROM)或闪存(flash RAM)。内存是计算机可读介质的示例。
计算机可读介质包括永久性和非永久性、可移动和非可移动媒体可以由任何方法或技术来实现信息存储。信息可以是计算机可读指令、数据结构、程序的模块或其他数据。计算机的存储介质的例子包括,但不限于相变内存(PRAM)、静态随机存取存储器(SRAM)、动态随机存取存储器(DRAM)、其他类型的随机存取存储器(RAM)、只读存储器(ROM)、电可擦除可编程只读存储器(EEPROM)、快闪记忆体或其他内存技术、只读光盘只读存储器(CD-ROM)、数字多功能光盘(DVD)或其他光学存储、磁盒式磁带,磁带磁磁盘存储或其他磁性存储设备或任何其他非传输介质,可用于存储可以被计算设备访问的信息。此外,尽管在附图中以特定顺序描述了本发明方法的操作,但是,这并非要求或者暗示必须按照该特定顺序来执行这些操作,或是必须执行全部所示的操作才能实现期望的结果。附加地或备选地,可以省略某些步骤,将多个步骤合并为一个步骤执行,和/或将一个步骤分解为多个步骤执行。
虽然已经参考若干具体实施方式描述了本发明的精神和原理,但是应该理解,本发明并不限于所公开的具体实施方式,对各方面的划分也不意味着这些方面中的特征不能组合以进行受益,这种划分仅是为了表述的方便。本发明旨在涵盖所附权利要求的精神和范围内所包括的各种修改和等同布置。
Claims (28)
- 一种目标重识别方法,其特征在于,包括:获取设置于监控区域内的多个摄像头采集的多个当前帧;根据所述多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;根据每个摄像头捕获到的所述目标图像进行数量检测,得到全局目标数量;根据所述目标图像和目标识别库进行目标重识别,所述目标识别库包括至少一个目标的身份标识和特征数据;当检测到所述全局目标数量符合预设增加条件时,根据所述目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识对所述至少一个未识别目标图像进行标记;根据所述新的身份标识和所述至少一个未识别目标图像的特征数据更新所述目标识别库。
- 根据权利要求1所述的方法,其特征在于,根据所述多个当前帧进行目标检测,还包括:将所述多个当前帧输入经训练的目标检测模型,以提取出每个摄像头捕获到的所述目标图像;其中,所述目标检测模型为基于YOLOv4-tiny网络创建的人体检测模型。
- 根据权利要求2所述的方法,其特征在于,所述方法还包括:根据所述监控区域内的真实采集图像对所述YOLOv4-tiny网络进行训练,得到所述目标检测模型。
- 根据权利要求1-3中任意一项所述的方法,其特征在于,所述目标图像为所述当前帧中包含目标特征的局部图像,根据每个摄像头捕获到的所述目标图像进行数量检测,还包括:根据每个摄像头的取景位置对捕获到的所述目标图像进行位置转换,得到每个摄像头捕获到的所述目标图像对应的全局位置;确定由不同摄像头各自捕获的所述目标图像的全局位置重合度,根据所述全局位置重合度对不同摄像头各自捕获的所述目标图像进行筛选,检测筛选后保留的所述目标图像的数量。
- 根据权利要求4所述的方法,其特征在于,所述方法还包括:当所述数量检测的结果少于在先全局目标数量时,则根据所述多个摄像头采集的所述多个当前帧和所述多个当前帧的上一帧,判断是否存在从预定区域离开所述监控区域的目标;若不存在所述目标,则仍然保留所述在先全局目标数量作为本次确定的所述全局目标数量;若存在所述目标,则将所述数量检测的结果作为本次确定的所述全局目标数量;其中,所述在先全局目标数量根据对所述多个当前帧的上一帧进行所述目标检测和所述数量检测得到。
- 根据权利要求4或5所述的方法,其特征在于,以及,根据每个摄像头的取景位置对捕获到的所述目标图像进行位置转换,还包括:根据每个摄像头的取景位置对所述当前帧中的所述目标图像的底部中心点进行投影变换,从而确定所述每个所述目标图像的地面坐标。
- 根据权利要求1-6中任意一项所述的方法,其特征在于,所述方法还包括:将所述多个当前帧输入经训练的目标数量检测模型,以执行所述目标检测和所述数量检测,得到所述全局目标数量;其中,所述目标数量检测模型为基于YOLOv4-tiny网络创建的行人数量检测模型。
- 根据权利要求1-7中任意一项所述的方法,其特征在于,根据所述目标图像和目标识别库进行目标重识别,还包括:计算所述目标图像与所述目标识别库中的特征数据之间的相似度,并依据计算得到的相似度,对所述目标图像进行目标重识别;当所述目标重识别的结果指示第一目标图像与所述目标识别库中的第一目标匹配时,根据所述第一目标的身份标识对所述第一目标图像进行标记。
- 根据权利要求1-8中任意一项所述的方法,其特征在于,所述方法还包括:若所述当前帧为非首帧,且当前帧对应的所述全局目标数量相较于上一帧对应的所述全局目标数量增加时,则所述全局目标数量符合所述预设增加条件;若所述当前帧为首帧时,默认所述全局目标数量符合所述预设增加条件。
- 根据权利要求1-9中任意一项所述的方法,其特征在于,根据所述新的身份标识和所述至少一个未识别目标图像的特征数据更新所述目标识别库,还包括:判断所述至少一个未识别目标图像是否满足预设图像质量条件;将所述新的身份标识和满足所述预设图像质量条件的所述未识别目标图像对应存入所述目标识别库。
- 根据权利要求8所述的方法,其特征在于,根据所述目标图像和目标识别库进行目标重识别之后,所述方法还包括:根据所述第一目标图像或所述第一目标图像的特征值对所述目标识别库中的所述第一目标的特征数据进行动态更新。
- 根据权利要求1-11中任意一项所述的方法,其特征在于,所述方法还包括对所述目标识别库进行替换更新,具体包括:根据所述目标识别库中的每个目标的所述特征数据对应的来源时间和当前时间的比较结果,对所述目标识别库进行替换更新;和/或,根据所述目标识别库中的每个目标的所述特征数据对应的全局位置和每个所述目标的当前全局位置的比较结果,对所述目标识别库进行替换更新;和/或,根据所述目标识别库中的每个目标的多个特征数据之间的特征相似度,对所述目标识别库进行替换更新。
- 根据权利要求12所述的方法,其特征在于,所述方法还包括:任意一个所述目标的所述特征数据的数量超过预设阈值之后,启动所述替换更新。
- 一种目标重识别装置,其特征在于,包括:获取模块,用于获取设置于监控区域内的多个摄像头采集的多个当前帧;目标检测模块,用于根据所述多个当前帧进行目标检测,确定每个摄像头捕获到的目标图像;数量检测模块,用于根据每个摄像头捕获到的所述目标图像进行数量检测,得到全局目标数量;目标重识别模块,用于根据所述目标图像和目标识别库进行目标重识别,所述目标识别库包括至少一个目标的身份标识和特征数据;身份标识模块,用于当检测到所述全局目标数量符合预设增加条件时,根据所述目标重识别的结果确定至少一个未识别目标图像,创建新的身份标识对所述至少一个未识别目标图像进行标记;目标识别库更新模块,用于根据所述新的身份标识和所述至少一个未识别目标图像的特征数据更新所述目标识别库。
- 根据权利要求14所述的装置,其特征在于,所述目标检测模块,还用于:将所述多个当前帧输入经训练的目标检测模型,以提取出每个摄像头捕获到的所述目标图像;其中,所述目标检测模型为基于YOLOv4-tiny网络创建的人体检测模型。
- 根据权利要求15所述的装置,其特征在于,所述目标检测模块,还用于:根据所述监控区域内的真实采集图像对所述YOLOv4-tiny网络进行训练,得到所述目标检测模型。
- 根据权利要求16所述的装置,其特征在于,所述目标图像为所述当前帧中包含目标特征的局部图像,所述数量检测模块还用于:根据每个摄像头的取景位置对捕获到的所述目标图像进行位置转换,得到每个摄像头捕获到的所述目标图像对应的全局位置;确定由不同摄像头各自捕获的所述目标图像的全局位置重合度,根据所述全局位置重合度对不同摄像头各自捕获的所述目标图像进行筛选,检测筛选后保留的所述目标图像的数量。
- 根据权利要求17所述的装置,其特征在于,所述数量检测模块还用于:当所述数量检测的结果少于在先全局目标数量时,则根据所述多个摄像头采集的所述多个当前帧和所述多个当前帧的上一帧,判断是否存在从预定区域离开所述监控区域的目标;若不存在所述目标,则仍然保留所述在先全局目标数量作为本次确定的所述全局目标数量;若存在所述目标,则将所述数量检测的结果作为本次确定的所述全局目标数量;其中,所述在先全局目标数量根据对所述多个当前帧的上一帧进行所述目标检测和所述数量检测得到。
- 根据权利要求17或18所述的装置,其特征在于,所述数量检测模块还用于:根据每个摄像头的取景位置对所述当前帧中的所述目标图像的底部中心点进行投影变换,从而确定所述每个所述目标图像的地面坐标。
- 根据权利要求14-19中任意一项所述的装置,其特征在于,所述装置还用于:将所述多个当前帧输入经训练的目标数量检测模型,以执行所述目标检测和所述数量检测,得到所述全局目标数量;其中,所述目标数量检测模型为基于YOLOv4-tiny网络创建的行人数量检测模型。
- 根据权利要求14-20中任意一项所述的装置,其特征在于,所述目标重识别模块还用于:计算所述目标图像与所述目标识别库中的特征数据之间的相似度,并依据计算得到的相似度,对所述目标图像进行目标重识别;当所述目标重识别的结果指示第一目标图像与所述目标识别库中的第一目标匹配时,根据所述第一目标的身份标识对所述第一目标图像进行标记。
- 根据权利要求14-21中任意一项所述的装置,其特征在于,所述身份标识模块还用于:若所述当前帧为非首帧,且当前帧对应的所述全局目标数量相较于上一帧对应的所述全局目标数量增加时,则所述全局目标数量符合所述预设增加条件;若所述当前帧为首帧时,默认所述全局目标数量符合所述预设增加条件。
- 根据权利要求14-22中任意一项所述的装置,其特征在于,所述目标识别库更新模块还用于:判断所述至少一个未识别目标图像是否满足预设图像质量条件;将所述新的身份标识和满足所述预设图像质量条件的所述未识别目标图像对应存入所述目标识别库。
- 根据权利要求21所述的装置,其特征在于,所述目标识别库更新模块还用于:根据所述第一目标图像或所述第一目标图像的特征值对所述目标识别库中的所述第一目标的特征数据进行动态更新。
- 根据权利要求14-25中任意一项所述的装置,其特征在于,所述目标识别库更新模块还用于:根据所述目标识别库中的每个目标的所述特征数据对应的来源时间和当前时间的比较结果,对所述目标识别库进行替换更新;和/或,根据所述目标识别库中的每个目标的所述特征数据对应的全局位置和每个所述目标的当前全局位置的比较结果,对所述目标识别库进行替换更新;和/或,根据所述目标识别库中的每个目标的多个特征数据之间的特征相似度,对所述目标识别库进行替换更新。
- 根据权利要求25所述的装置,其特征在于,所述目标识别库更新模块还用于:任意一个所述目标的所述特征数据的数量超过预设阈值之后,启动所述替换更新。
- 一种目标重识别装置,其特征在于,包括:一个或者多个多核处理器;存储器,用于存储一个或多个程序;当所述一个或多个程序被所述一个或者多个多核处理器执行时,使得所述一个或多个多核处理器实现:如权利要求1-13中任一项所述的方法。
- 一种计算机可读存储介质,所述计算机可读存储介质存储有程序,当所述程序被多核处理器执行时,使得所述多核处理器执行如权利要求1-13中任一项所述的方法。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110095415.1A CN112906483B (zh) | 2021-01-25 | 2021-01-25 | 一种目标重识别方法、装置及计算机可读存储介质 |
CN202110095415.1 | 2021-01-25 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022156234A1 true WO2022156234A1 (zh) | 2022-07-28 |
Family
ID=76118765
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2021/117512 WO2022156234A1 (zh) | 2021-01-25 | 2021-09-09 | 一种目标重识别方法、装置及计算机可读存储介质 |
Country Status (3)
Country | Link |
---|---|
CN (1) | CN112906483B (zh) |
TW (1) | TWI798815B (zh) |
WO (1) | WO2022156234A1 (zh) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117423051A (zh) * | 2023-10-18 | 2024-01-19 | 广州元沣智能科技有限公司 | 一种基于场所活动对象的信息监测分析方法 |
CN117974718A (zh) * | 2024-02-02 | 2024-05-03 | 北京视觉世界科技有限公司 | 目标检测追踪方法、装置、设备及存储介质 |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112906483B (zh) * | 2021-01-25 | 2024-01-23 | 中国银联股份有限公司 | 一种目标重识别方法、装置及计算机可读存储介质 |
CN113723361A (zh) * | 2021-09-18 | 2021-11-30 | 西安邮电大学 | 一种基于深度学习的视频监测方法及设备 |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9607402B1 (en) * | 2016-05-09 | 2017-03-28 | Iteris, Inc. | Calibration of pedestrian speed with detection zone for traffic intersection control |
CN110309717A (zh) * | 2019-05-23 | 2019-10-08 | 南京熊猫电子股份有限公司 | 一种基于深度神经网络的行人计数方法 |
CN112183431A (zh) * | 2020-10-12 | 2021-01-05 | 上海汉时信息科技有限公司 | 实时行人数量统计方法、装置、相机和服务器 |
CN112906483A (zh) * | 2021-01-25 | 2021-06-04 | 中国银联股份有限公司 | 一种目标重识别方法、装置及计算机可读存储介质 |
Family Cites Families (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105518744B (zh) * | 2015-06-29 | 2018-09-07 | 北京旷视科技有限公司 | 行人再识别方法及设备 |
GB2545900B (en) * | 2015-12-21 | 2020-08-12 | Canon Kk | Method, device, and computer program for re-identification of objects in images obtained from a plurality of cameras |
WO2017150899A1 (ko) * | 2016-02-29 | 2017-09-08 | 광주과학기술원 | 전역적 다중 객체 추적을 위한 객체 재식별 방법 |
CN107346409B (zh) * | 2016-05-05 | 2019-12-17 | 华为技术有限公司 | 行人再识别方法和装置 |
US10395385B2 (en) * | 2017-06-27 | 2019-08-27 | Qualcomm Incorporated | Using object re-identification in video surveillance |
CN109697391A (zh) * | 2017-10-23 | 2019-04-30 | 北京京东尚科信息技术有限公司 | 封闭场所内人物再识别的方法、系统及终端设备 |
CN111652197B (zh) * | 2018-02-08 | 2023-04-18 | 创新先进技术有限公司 | 入离场状态检测方法和装置 |
CN108399381B (zh) * | 2018-02-12 | 2020-10-30 | 北京市商汤科技开发有限公司 | 行人再识别方法、装置、电子设备和存储介质 |
CN108875588B (zh) * | 2018-05-25 | 2022-04-15 | 武汉大学 | 基于深度学习的跨摄像头行人检测跟踪方法 |
CN110008799A (zh) * | 2018-11-09 | 2019-07-12 | 阿里巴巴集团控股有限公司 | 指定区域的人流状况估算方法和装置 |
CN109740413B (zh) * | 2018-11-14 | 2023-07-28 | 平安科技(深圳)有限公司 | 行人重识别方法、装置、计算机设备及计算机存储介质 |
CN109902573B (zh) * | 2019-01-24 | 2023-10-31 | 中国矿业大学 | 面向矿井下视频监控的多摄像机无标注行人重识别方法 |
CN110175527B (zh) * | 2019-04-29 | 2022-03-25 | 北京百度网讯科技有限公司 | 行人再识别方法及装置、计算机设备及可读介质 |
CN110826415A (zh) * | 2019-10-11 | 2020-02-21 | 上海眼控科技股份有限公司 | 一种场景图像中车辆的重识别方法及设备 |
TWI705383B (zh) * | 2019-10-25 | 2020-09-21 | 緯創資通股份有限公司 | 人物追蹤系統及人物追蹤方法 |
CN110991283A (zh) * | 2019-11-21 | 2020-04-10 | 北京格灵深瞳信息技术有限公司 | 重识别及训练数据获取方法、装置、电子设备及存储介质 |
CN111159475B (zh) * | 2019-12-06 | 2022-09-23 | 中山大学 | 基于多摄像头视频图像的行人再识别路径生成方法 |
CN111145213A (zh) * | 2019-12-10 | 2020-05-12 | 中国银联股份有限公司 | 一种目标跟踪方法、装置、系统及计算机可读存储介质 |
CN111160275B (zh) * | 2019-12-30 | 2023-06-23 | 深圳元戎启行科技有限公司 | 行人重识别模型训练方法、装置、计算机设备和存储介质 |
CN111274992A (zh) * | 2020-02-12 | 2020-06-12 | 北方工业大学 | 一种跨摄像头的行人重识别方法及系统 |
CN111382751B (zh) * | 2020-03-11 | 2023-04-18 | 西安应用光学研究所 | 一种基于颜色特征的目标再识别方法 |
CN111680551B (zh) * | 2020-04-28 | 2024-06-11 | 平安国际智慧城市科技股份有限公司 | 监控牲畜数量的方法、装置、计算机设备及存储介质 |
CN111623791A (zh) * | 2020-05-28 | 2020-09-04 | 识加科技(上海)有限公司 | 在公共区域进行导航的方法、装置、设备和介质 |
CN111783570A (zh) * | 2020-06-16 | 2020-10-16 | 厦门市美亚柏科信息股份有限公司 | 一种目标重识别的方法、装置、系统及计算机存储介质 |
CN111882586B (zh) * | 2020-06-23 | 2022-09-13 | 浙江工商大学 | 一种面向剧场环境的多演员目标跟踪方法 |
-
2021
- 2021-01-25 CN CN202110095415.1A patent/CN112906483B/zh active Active
- 2021-09-07 TW TW110133172A patent/TWI798815B/zh active
- 2021-09-09 WO PCT/CN2021/117512 patent/WO2022156234A1/zh active Application Filing
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9607402B1 (en) * | 2016-05-09 | 2017-03-28 | Iteris, Inc. | Calibration of pedestrian speed with detection zone for traffic intersection control |
CN110309717A (zh) * | 2019-05-23 | 2019-10-08 | 南京熊猫电子股份有限公司 | 一种基于深度神经网络的行人计数方法 |
CN112183431A (zh) * | 2020-10-12 | 2021-01-05 | 上海汉时信息科技有限公司 | 实时行人数量统计方法、装置、相机和服务器 |
CN112906483A (zh) * | 2021-01-25 | 2021-06-04 | 中国银联股份有限公司 | 一种目标重识别方法、装置及计算机可读存储介质 |
Non-Patent Citations (1)
Title |
---|
WANG YONGKANG: "Research on Trajectory Recovery of Indoor Dynamic Object (Pedestrian) Based on Binocular Vision", MASTER THESIS, TIANJIN POLYTECHNIC UNIVERSITY, CN, no. 9, 15 September 2019 (2019-09-15), CN , XP055952109, ISSN: 1674-0246 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117423051A (zh) * | 2023-10-18 | 2024-01-19 | 广州元沣智能科技有限公司 | 一种基于场所活动对象的信息监测分析方法 |
CN117423051B (zh) * | 2023-10-18 | 2024-03-26 | 广州元沣智能科技有限公司 | 一种基于场所活动对象的信息监测分析方法 |
CN117974718A (zh) * | 2024-02-02 | 2024-05-03 | 北京视觉世界科技有限公司 | 目标检测追踪方法、装置、设备及存储介质 |
Also Published As
Publication number | Publication date |
---|---|
TWI798815B (zh) | 2023-04-11 |
TW202230215A (zh) | 2022-08-01 |
CN112906483A (zh) | 2021-06-04 |
CN112906483B (zh) | 2024-01-23 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2022156234A1 (zh) | 一种目标重识别方法、装置及计算机可读存储介质 | |
WO2021043073A1 (zh) | 基于图像识别的城市宠物活动轨迹监测方法及相关设备 | |
WO2019218824A1 (zh) | 一种移动轨迹获取方法及其设备、存储介质、终端 | |
US9210385B2 (en) | Method and system for metadata extraction from master-slave cameras tracking system | |
US8634601B2 (en) | Surveillance-based high-resolution facial recognition | |
CN107016322B (zh) | 一种尾随人员分析的方法及装置 | |
CN110866466B (zh) | 一种人脸识别方法、装置、存储介质和服务器 | |
CN107545256B (zh) | 一种结合时空和网络一致性的摄像机网络行人重识别方法 | |
WO2020258978A1 (zh) | 对象检测方法和装置 | |
US20220301317A1 (en) | Method and device for constructing object motion trajectory, and computer storage medium | |
WO2022213540A1 (zh) | 目标检测、属性识别与跟踪方法及系统 | |
Wiratma et al. | An experimental evaluation of grouping definitions for moving entities | |
CN112562315A (zh) | 一种获取车流信息的方法、终端及存储介质 | |
CN111708907A (zh) | 一种目标人员的查询方法、装置、设备及存储介质 | |
CN112001280A (zh) | 一种实时、可在线优化的人脸识别系统和方法 | |
WO2020232697A1 (zh) | 一种在线人脸聚类的方法及系统 | |
De Marsico et al. | ES-RU: an e ntropy based rule to s elect r epresentative templates in face su rveillance | |
US20230055581A1 (en) | Privacy preserving anomaly detection using semantic segmentation | |
CN115272967A (zh) | 一种跨摄像机行人实时跟踪识别方法、装置及介质 | |
TWI728655B (zh) | 應用於動物的卷積神經網路偵測方法及系統 | |
CN110956644B (zh) | 一种运动轨迹确定方法及系统 | |
CN110751065B (zh) | 训练数据的采集方法及装置 | |
RU2694139C1 (ru) | Способ определения девиантного поведения человека в режиме одновременной работы группы видеокамер | |
Zhang et al. | What makes for good multiple object trackers? | |
Piccinini et al. | SIFT-based segmentation of multiple instances of low-textured objects |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21920612 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21920612 Country of ref document: EP Kind code of ref document: A1 |