WO2019237536A1 - Procédé et appareil de suivi d'une cible en temps réel, dispositif informatique et support d'informations - Google Patents

Procédé et appareil de suivi d'une cible en temps réel, dispositif informatique et support d'informations Download PDF

Info

Publication number
WO2019237536A1
WO2019237536A1 PCT/CN2018/105404 CN2018105404W WO2019237536A1 WO 2019237536 A1 WO2019237536 A1 WO 2019237536A1 CN 2018105404 W CN2018105404 W CN 2018105404W WO 2019237536 A1 WO2019237536 A1 WO 2019237536A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
tracking
tracking target
similarity
current
Prior art date
Application number
PCT/CN2018/105404
Other languages
English (en)
Chinese (zh)
Inventor
马潜
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2019237536A1 publication Critical patent/WO2019237536A1/fr

Links

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
    • G06V20/42—Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items of sport video content
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/70—Determining position or orientation of objects or cameras
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/30—Scenes; Scene-specific elements in albums, collections or shared content, e.g. social network photos or video
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the present application relates to the technical field of target tracking, and in particular, to a method, a device, a computer device, and a storage medium for real-time target tracking.
  • Target tracking is one of the hot topics in the field of computer vision research. Over the past few decades, the research on target tracking has made great progress.
  • the current target tracking field mainly adopts a single target long-term tracking (Tracking-Learning-Detection (hereinafter referred to as TLD)) algorithm and a kernel correlation filtering (Kernel Correlation Filter (hereinafter referred to as KCF)) algorithm for target tracking processing.
  • TLD Tracker
  • KCF Kernelation Filter
  • KCF kernel correlation filtering
  • the TLD algorithm is mainly composed of a tracker, a detector, and a machine learning module.
  • the tracking process is: the tracker predicts the position of the tracked object, and the detector detects the position of all objects in the image and sends the most trusted position to the machine learning module and updates The position of the object in the tracker.
  • the machine learning module trains a classifier based on the position sent by the tracker and the detector to improve the accuracy of the detector.
  • the KCF algorithm trains a target detector during the tracking process, uses the target detector to detect whether the predicted position of the next frame is the target, obtains a new detection result, and then uses the new detection result to update the target detector.
  • the embodiments of the present application provide a method, a device, a computer device, and a storage medium for real-time tracking of targets to solve the problems of low efficiency and incomplete processing of lost targets in current target tracking.
  • a method for real-time target tracking includes:
  • the target similarity is less than a preset similarity, determining that the reference tracking target corresponding to the target similarity is a missing tracking target in the current image
  • the reference tracking target is released.
  • a target real-time tracking device includes:
  • An original video data acquisition module configured to acquire original video data, where the original video data includes at least two frames of images;
  • a reference tracking target detection module configured to select any frame image in the original video data as a reference image, perform target detection on the reference image, and obtain at least one reference tracking target and a corresponding reference target feature vector;
  • a current tracking target detection module configured to perform target detection on a current image in the original video data, and obtain at least one current tracking target and a corresponding current target feature vector;
  • a feature similarity acquisition module configured to calculate feature similarities between any of the reference target feature vectors and all current target feature vectors to determine a target similarity corresponding to the reference target feature vector
  • a missing tracking target determination module configured to determine that the reference tracking target corresponding to the target similarity is a missing tracking target in the current image if the target similarity is less than a preset similarity
  • the reference tracking target release module is configured to release the reference tracking target if consecutive N frames of images after the current image in the reference tracking target are lost tracking targets.
  • a computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor.
  • the processor executes the computer-readable instructions, the following steps are implemented:
  • the target similarity is less than a preset similarity, determining that the reference tracking target corresponding to the target similarity is a missing tracking target in the current image
  • the reference tracking target is released.
  • One or more non-volatile readable storage media storing computer readable instructions, where the computer readable storage medium stores computer readable instructions, and when the computer readable instructions are executed by one or more processors, Cause the one or more processors to perform the following steps:
  • the target similarity is less than a preset similarity, determining that the reference tracking target corresponding to the target similarity is a missing tracking target in the current image
  • the reference tracking target is released.
  • FIG. 1 is an application environment diagram of a method for real-time target tracking in an embodiment of the present application
  • FIG. 2 is a flowchart of a method for real-time tracking of a target in an embodiment of the present application
  • FIG. 3 is a flowchart of step S20 in FIG. 2;
  • FIG. 4 is a flowchart of step S40 in FIG. 2;
  • FIG. 5 is a flowchart of a method for real-time tracking of a target in an embodiment of the present application
  • step S82 in FIG. 5 is a flowchart of step S82 in FIG. 5;
  • FIG. 7 is a schematic diagram of a target real-time tracking device in an embodiment of the present application.
  • FIG. 8 is a schematic diagram of a computer device according to an embodiment of the present application.
  • the target real-time tracking method provided in the embodiment of the present application can be applied in the application environment shown in FIG. 1, that is, the target real-time tracking method is applied in the monitoring system shown in FIG. 1.
  • the monitoring system includes a server and is connected to the server through a network. At least one monitoring terminal and at least one camera terminal connected to the server through a network.
  • the monitoring terminal is a terminal used by the monitoring personnel to track and process the target. Any monitoring terminal can communicate with the server through any of wireless networks such as WiFi, 3G, 4G, and 5G or wired networks.
  • the camera terminal is a terminal for collecting video data, and any camera terminal communicates with a server through any of a wireless network such as WiFi, 3G, 4G, and 5G or a wired network.
  • the monitoring terminal may be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server can be implemented by an independent server or a server cluster composed of multiple servers.
  • the camera terminal may include one camera
  • the target real-time tracking method is applied to the server in FIG. 1 as an example for description. As shown in FIG. 2, the target real-time tracking method includes the following steps:
  • the original video data includes at least two frames of images.
  • Raw video data refers to unprocessed video data, specifically video data collected by any camera terminal and uploaded to the server.
  • Video data refers to a continuous image sequence, and its essence is composed of continuous frames of images.
  • the original video data obtained by the server includes at least two frames of images, where a frame is the smallest visual unit of the video and is a static image. Time-series frames are synthesized together to form dynamic video data.
  • each camera terminal carries a camera ID
  • the camera ID is a unique identifier for identifying the camera terminal.
  • the camera ID may specifically be a factory identifier of the camera terminal.
  • the uploaded video data carries a corresponding camera ID in order to identify the source of the video data.
  • monitoring personnel perform target monitoring through the monitoring terminal, they can send corresponding monitoring requests to the server according to the monitoring requirements.
  • the monitoring request carries a camera ID, which is the identifier of the camera terminal in the area to be monitored.
  • the server after acquiring video data sent by at least one camera terminal, stores all video data in association with its corresponding camera ID, so as to subsequently find corresponding video data based on the camera ID.
  • the server obtains the monitoring request sent by the monitoring terminal, it can query the corresponding original video data based on the camera ID in the monitoring request, so as to perform target tracking processing based on the original video data.
  • the server performs frame processing on the original video data by using a video editing tool to obtain at least two frames of images.
  • Video editing tools are used for editing operations such as framing, editing, and merging of videos. Common video editing tools include VideoStudio, Pr, Vegas, and Edius. For example, after the server obtains the original video data with a duration of 10 seconds, it can use the video editing tool Pr to perform frame processing on the original video data, divide the original video data into 30 frames per second, and obtain a total of 300 frames.
  • S20 Select a reference image from the original video data, perform target detection on the reference image, and obtain at least one reference tracking target and a corresponding reference target feature vector.
  • the reference image refers to the image used to detect and determine the reference tracking target when starting target tracking.
  • the reference image can be the first frame of at least two frames obtained by framing the original video data, or it can be any frame selected by the monitoring staff through all the frames processed by the monitoring terminal.
  • Target detection also called target extraction, is an image segmentation based on the geometric and statistical characteristics of the target. It combines target segmentation and recognition into one. Its accuracy and real-time performance is an important capability of the entire system. Object detection is to identify the frame image and detect a specific object (such as a pedestrian or an animal) in the image.
  • the reference tracking target refers to a tracking target determined after target detection is performed on a reference image, and the reference tracking target is a tracking target used as a reference object.
  • the first frame of the original video data may be used as a reference image, and a specific target obtained after subject detection is performed on the reference image as a reference tracking target. Understandably, after the target detection is performed on the reference image, the server may obtain at least one reference tracking target, that is, may obtain one reference tracking target, or may obtain multiple reference tracking targets.
  • the server creates a pair of monitoring threads for each benchmark tracking target, and the monitoring thread includes a tracking thread and a verification thread.
  • the monitoring thread is a thread for monitoring the benchmark tracking target in real time.
  • the tracking thread is a thread for extracting the benchmark target feature vector corresponding to the benchmark tracking target and maintaining the benchmark target feature vector queue.
  • the verification thread is a thread for performing comparison between feature vectors and updating comparison results.
  • the extraction and comparison of the reference target feature vector are monitored by the tracking thread and the verification thread, respectively, which can help improve the efficiency of target tracking.
  • the reference target feature vector in the tracking thread is updated according to the comparison result. Improved the accuracy of target tracking.
  • the reference target feature vector is a row vector that characterizes the basic features of the reference tracking target in the reference image.
  • a row vector is a 1 ⁇ n matrix, that is, a matrix consisting of a vector of n elements.
  • the reference target feature vector may be obtained by extracting pixel values in the reference image.
  • the reference image may be a digital image, that is, a two-dimensional image uses a finite digital value to represent its corresponding pixel value.
  • a digital computer or a digital image may be used.
  • An image stored and processed by a circuit represented by an array or matrix.
  • a frame of image can be seen as a matrix where each pixel is a color value, which means that the basic feature is to connect each row of the matrix into a row vector.
  • each frame of image is composed of 8 ⁇ 8 pixels, and the value of each pixel is 0-16, white is 0, and black is 16.
  • the 8 ⁇ 8 matrix is converted into a 64-dimensional row vector to create a feature vector.
  • the server may obtain a reference target feature vector in the reference image according to the reference tracking target.
  • S30 Perform target detection on the current image in the original video data, and obtain at least one current tracking target and a corresponding current target feature vector.
  • the current image refers to the image used to detect the target and track the current position of the target.
  • the current image is the next frame image or the Nth frame image after the reference frame image selected in the original video data, and N is an integer greater than 1.
  • the current tracking target is a tracking target determined after target detection is performed on the current image.
  • the current tracking target uses the existing reference tracking target as a reference, and compares whether the current tracking target and the existing reference tracking target are the same tracking target; if the comparison is successful, the current tracking target and the existing tracking target are compared.
  • the reference tracking target is the same tracking target; if the comparison is unsuccessful, the current tracking target may be a new tracking target or an obstructed reference tracking target.
  • the first frame image of the original video data may be used as the reference image
  • the second frame image may be used as the current image
  • the specific target obtained after subject detection is performed on the current image as the current tracking target.
  • the server may obtain at least one current tracking target, that is, may obtain one current tracking target, or may obtain multiple current tracking targets.
  • the current target feature vector is a row vector that characterizes the basic features of the current tracking target in the current image.
  • a row vector is a 1 ⁇ n matrix, that is, a matrix consisting of a vector of rows with n elements.
  • the current target feature vector can be obtained by extracting pixel values in the current image.
  • the reference image can be a digital image, that is, a two-dimensional image can represent its corresponding pixel value with a finite numerical value, which can be digital computer or digital.
  • An image stored and processed by a circuit represented by an array or matrix.
  • An image is composed of 8 ⁇ 8 pixels. Each pixel has a value of 0-16, white is 0, and black is 16.
  • the 8 ⁇ 8 matrix is converted into a 64-dimensional row vector to create a feature vector.
  • the server may obtain a corresponding current target feature vector in the current image according to the current tracking target.
  • Feature similarity refers to the degree of similarity between two feature vectors of any reference target feature vector and any current target feature vector.
  • Feature similarity can generally be obtained by calculating the distance between two feature vectors. By calculating the distance, the closer the two feature vectors are, the greater the similarity is.
  • Euclidean distance algorithm, Manhattan distance algorithm, cosine similarity algorithm, etc. may be used to implement the calculation of feature similarity.
  • the server may calculate any existing reference target feature vector and the current target feature vectors corresponding to all current tracking targets detected in the current image, obtain multiple feature similarities, and then calculate the feature similarity from the calculated The most similar one among multiple feature similarities is selected as the target similarity corresponding to the reference target feature vector.
  • a rule for selecting a target similarity from a plurality of feature similarities is: selecting a feature similarity corresponding to a maximum value of the feature similarities from the multiple feature similarities as the target similarity.
  • the current tracking target A1, the current tracking target A3, the current target feature vector A1 (x1, y1, z1), and the current target feature vector A3 (x3, y3, z3) are detected; and Based on the reference tracking target B1 and the reference target feature vector B1 (x2, y2, z2) detected in the reference image.
  • the distance between the current target feature vector A1 (x1, y1, z1) and the reference target feature vector B1 (x2, y2, z2) can be calculated using the Euclidean distance algorithm. The calculation formula is specifically: Then use the calculated Euclidean distance d 11 to calculate the feature similarity sim 11.
  • the formula for calculating the feature similarity is as follows: Then use the Euclidean distance algorithm to calculate the distance between the current target feature vector A3 (x3, y3, z3) and the reference target feature vector B1 (x2, y2, z2). The calculation formula is as follows: Then use the calculated Euclidean distance d 13 to calculate the feature similarity sim 13.
  • the formula for calculating the feature similarity is as follows: From the calculation formula of feature similarity, it can be known that the smaller the Euclidean distance calculated, the greater the similarity between the two feature vectors. In this embodiment, the sizes of the feature similarity sim 11 and the feature similarity sim 13 are compared, and the larger one is selected, such as sim 11 as the target similarity corresponding to the reference tracking target B1.
  • the preset similarity is an index set in advance for evaluating whether or not the similarity for the same tracking target is reached.
  • the missing tracking target refers to a reference tracking target that appears in the reference image but does not appear in the current image.
  • the missing tracking target may be a reference tracking target that appears to be blocked or disappears in the current image. If the feature similarity of any reference tracking target is less than the preset similarity, it means that the reference tracking target is not the same tracking target as all current tracking targets detected in the current image, and the reference tracking target is determined in the current image. Tracking targets for loss.
  • the Euclidean distance algorithm is used to calculate the Euclidean distance between two feature vectors, and the corresponding feature similarity sim is calculated based on the Euclidean distance.
  • the formula for calculating the feature similarity is as follows: Among them, d is the Euclidean distance between two feature vectors, and the range of sim is [0,1]. When d is smaller, sim is larger, that is, when the distance is closer, the feature similarity is larger. Understandably, the greater the feature similarity, the more similar the two feature vectors, and the greater the likelihood that the corresponding two tracking targets are the same tracking target; the smaller the feature similarity, the more dissimilar the two feature vectors are, and the corresponding The less likely that two tracking targets are the same tracking target.
  • the preset similarity may be set to 0.8. If the target similarity corresponding to any reference tracking target is less than 0.8, it means that the reference tracking target is not the same as all current tracking targets detected in the current image. Tracking target, the reference tracking target is a missing tracking target in the current image.
  • the continuous N frames of images refer to the continuous N frames of images after the current image with the missing tracking target is determined, where N is an integer greater than 1, and the value of N can be set according to actual needs.
  • N can be set to 500, and monitoring of the reference tracking target for 10 seconds can be implemented under the condition that video data is divided into 50 frames of images per second.
  • the release of the benchmark tracking target by the server refers to deleting the monitoring thread of the benchmark tracking target that is determined to be a missing tracking target, and no longer tracking this benchmark tracking target. If the reference tracking target is not detected again within a preset number of frames (that is, consecutive N frames of the reference tracking target after the current image are lost tracking targets), the server deletes the monitoring thread of the reference tracking target. Abandon the tracking of the benchmark tracking target, achieve reasonable handling of the missing tracking target, avoid the monitoring thread that has not been detected for a long time, but keep running the benchmark tracking target, reducing the occupation of system resources and improving target tracking effectiveness. If any baseline tracking target is determined to be a lost tracking target, it will be released and its monitoring thread will be deleted.
  • the target similarity of the reference tracking target is not less than the preset similarity, it means that there is one current tracking target whose reference tracking target is the same tracking target among all current tracking targets detected in the current image
  • the current tracking target for is updated to the new baseline tracking target. Updating the current tracking target to a new reference tracking target refers to updating the monitoring thread of the reference tracking target, replacing the reference target feature vector with the current target feature vector, and updating the feature maintenance queue of the reference tracking target. That is, after the current image, target detection is performed with the updated reference tracking target as a new reference object, which can accurately locate and track each reference tracking target and improve the accuracy of target tracking.
  • the Euclidean distance algorithm is used to calculate the Euclidean distance between two feature vectors, and the corresponding feature similarity sim is calculated based on the Euclidean distance.
  • the formula for calculating the feature similarity is as follows: Where d is the Euclidean distance between two feature vectors, and the range of sim is [0,1]. The smaller d is, the larger sim is, that is, the closer the Euclidean distance is, the more similar the feature is. Big.
  • the preset similarity may be set to 0.8.
  • the target similarity of any reference tracking target is not less than 0.8, it means that one of the current tracking targets detected in the current image has the reference tracking target as The current tracking target of the same tracking target, updates the current tracking target to a new reference tracking target in the current image, updates the monitoring thread of the reference tracking target, and updates the current target feature vector to the tracking thread.
  • the server obtains original video data, selects a reference image and performs target detection, obtains a reference tracking target and a reference target feature vector, and can independently select a monitoring starting point, thereby enhancing monitoring flexibility. .
  • the server can allocate a corresponding monitoring thread for each benchmark tracking target to improve the target monitoring accuracy and accuracy.
  • the server performs target detection on the current image, obtains the current tracking target and the current target feature vector, determines the target similarity corresponding to each reference target feature vector based on the reference target feature vector and the current target feature vector, and sets the target similarity Compare with the preset similarity to determine whether the reference tracking target is a lost tracking target and determine whether it is a lost tracking target through feature similarity.
  • the calculation process is simple and convenient, which helps improve tracking efficiency.
  • the server releases the reference tracking target when the consecutive N frames of any reference tracking target behind the current image are lost tracking targets, which can determine and release the lost tracking reference tracking target, reducing the use of system resources and improving target tracking. Efficiency and accuracy.
  • the server After determining that the baseline tracking target is not a lost tracking target, the server updates the current tracking target to a new baseline tracking target to improve the accuracy of target tracking.
  • the server uses a target detection algorithm to detect the reference image to obtain at least one reference tracking target and a corresponding reference target feature vector from the reference image. That is, the reference tracking target may be one or more, and each reference tracking target corresponds to a reference target feature vector.
  • target detection is performed on the reference image to obtain at least one reference tracking target and a corresponding reference target feature vector, which specifically includes the following steps:
  • S21 Use a target detection algorithm to perform target detection on the reference image, and obtain at least one detection window.
  • the target detection algorithm is an algorithm for quickly and accurately detecting a specific target in an image.
  • object detection algorithms are YOLO (You Only Look, Once, Unified Real-time Object Detection) algorithm, SSD (Single Shot Multibox Detector, single deep neural network detection) algorithm, R-CNN (Regions with CNN features), based on convolutional neural network detection ) Algorithm and Fast R-CNN (Fast Regions with CNN, fast detection based on convolutional neural network) algorithms, etc.
  • the detection window is an area that contains a specific target and is determined when the target image is detected using the target detection algorithm.
  • the detection window is generally a rectangular window.
  • the target detection of the reference image by the server requires the use of computer vision tools combined with the target detection algorithm.
  • Common computer vision tools include OPENCV and MATLAB. When a computer vision tool is selected, the target detection algorithm can be selected according to personal use proficiency or combined with actual needs. Computer vision tools are generally compatible with most mainstream target detection algorithms.
  • OPENCV combined with the YOLO algorithm can be used to implement target detection on the reference image, that is, the computer vision tool OPENCV is used to perform target detection on the reference image.
  • YOLO algorithm is an algorithm that solves object detection as a regression problem. After one inference of the input image, the position of all objects in the image, their category and the corresponding confidence probability can be obtained.
  • YOLO includes 24 convolutional layers and 2 fully connected layers. Among them, the convolutional layer is used to extract image features, and the fully connected layer is used to predict image position and class probability values.
  • the YOLO algorithm can divide the input image into S ⁇ S grids, and each grid is responsible for detecting objects that 'fall' into the grid.
  • Each grid outputs B bounding boxes (including rectangular areas of objects).
  • the Bounding boxes contain 5 data values: x, y, w, h, and confidence.
  • x and y refer to the objects predicted by the current grid.
  • the coordinates of the center position of the bounding box, w and h are the width and height of the bounding box, and the confidence reflects whether the current bounding box contains an object and the accuracy of the object position.
  • the intersection area can be normalized to the [0,1] interval by using pixels as the unit of the intersection area with the real area of the object. determine. For example, divide the reference image into 7 ⁇ 7 grids, and each grid outputs 2 bounding boxes (rectangular areas containing objects).
  • Each bounding box information corresponds to an original window, and a total of 7 ⁇ 7 ⁇ 2 ⁇ 98 originals are obtained. Window, first compare the confidence of each original window with a preset threshold to remove the original windows that are less likely, and then use NMS (non maximum suppression) to remove redundant original windows, and the remaining The original window is the detection window.
  • the server obtains a reference tracking target included in the detection window based on each detection window detected in the reference image, and extracts basic features of the reference target image in the reference image to generate a reference target feature vector.
  • OPENCV combined with the YOLO algorithm can be used to extract the hog feature vector as the reference target feature vector, and the detection window is divided into cells of several pixels. Each cell is divided into 9 bins according to the gradient direction.
  • each cell perform histogram statistics on all bins in the gradient direction of all pixels to obtain a 9-dimensional feature vector; each adjacent 4 cells form a block, and a block (block)
  • the feature vectors in) are combined to obtain a 36-dimensional feature vector, and the detection window is scanned with a block, and the scanning step is a unit; finally, the feature vectors of all the blocks are connected in series to obtain the reference target feature vector.
  • OPENCV is combined with the YOLO algorithm to perform target detection, to obtain the reference tracking target and the reference target feature vector, which can quickly perform target detection and extract the reference target feature vector, which helps improve the efficiency of target tracking.
  • step S30 target detection is performed on the current image in the original video data, and at least one current tracking target and a corresponding current target feature vector are obtained.
  • a target detection algorithm is used to perform a current detection on the current image in the original video data.
  • For target detection at least one detection window is obtained, and based on each detection window, a corresponding current tracking target and a current target feature vector are obtained.
  • the specific implementation process is similar to steps S21-S22. To avoid repetition, details are not repeated here.
  • the target real-time tracking method further includes the following steps:
  • the server creates a pair of monitoring threads for each benchmark tracking target, and the pair of monitoring threads includes a tracking thread and a verification thread.
  • a monitoring thread is a thread used to monitor benchmark tracking targets in real time.
  • the tracking thread is used to extract the reference target feature vector corresponding to the reference tracking target and maintain the reference target feature vector queue. That is, the tracking thread extracts and stores the reference target feature vector in real time, and updates the latest benchmark according to a certain frequency.
  • Target feature vector is a thread for performing comparison between feature vectors and updating comparison results.
  • the tracking thread is a thread for extracting the reference target feature vector corresponding to the reference tracking target and maintaining the reference target feature vector queue.
  • the tracking thread performs the extraction and storage of the reference target feature vector in real time, and updates the latest reference within a certain frequency.
  • the target feature vector is pushed to the corresponding verification thread so that the verification thread compares based on the latest benchmark target feature vector.
  • the target real-time tracking method further includes: in a verification thread, determining whether the target similarity is less than a preset similarity, and obtaining a judgment result. And update the tracking thread based on the judgment result.
  • the verification thread After the verification thread calculates the target similarity corresponding to the reference target feature vector, it compares the target similarity with the preset similarity, that is, determines whether the target similarity is less than the preset similarity. When the target similarity is less than the preset similarity, the reference tracking target is a missing tracking target in the current image. The verification thread will send the original reference target feature vector to the tracking thread for labeling. When consecutive N frames of images after the image are marked as missing tracking targets, the tracking thread corresponding to the reference tracking target is released to save system resources. When the target similarity is not less than the preset similarity, the reference target is the same tracking target as the current tracking target, and the verification thread will send the current target feature vector to the tracking thread to update the reference target feature vector.
  • the extraction and comparison of feature vectors are monitored by the tracking thread and the verification thread respectively, which can improve the efficiency of target tracking.
  • the reference target feature vector in the tracking thread is updated according to the comparison result to improve the target tracking. accuracy.
  • step S40 the feature similarity between any reference target feature vector and all current target feature vectors is calculated to determine the target similarity corresponding to the reference target feature vector, which specifically includes the following steps:
  • the distance algorithm is used to calculate the measured distances between any reference target feature vector and all current target feature vectors, and the corresponding feature similarity is obtained based on the measured distances.
  • the distance algorithm is an algorithm used to calculate the distance between two or more points. Common distance algorithms include Euclidean distance algorithm (ie, European distance algorithm), Manhattan distance algorithm, or cosine distance algorithm.
  • the measured distance refers to the actual distance between the reference target feature vector and the current target feature vector.
  • the server uses the distance algorithm for distance calculation, it first selects a reference target feature vector, and then calculates the distance between all current target feature vectors and the reference target feature vector, and obtains the reference target feature vector and at least one current target. The measured distance between the feature vectors, and the feature similarity is calculated based on the measured distance.
  • obtaining the corresponding feature similarity based on the measured distance specifically refers to: using a formula for calculating feature similarity Calculate the measured distance to obtain the corresponding feature similarity, where d is the measured distance between any reference target feature vector and a current target feature vector, and sim is the feature similarity corresponding to the measured distance.
  • S42 Select the maximum value from all feature similarities and determine the target similarity corresponding to the reference target feature vector.
  • the target similarity is to select the most similar feature similarity from the feature similarities between the reference target feature vector and all current target feature vectors, that is, the feature similarity corresponding to the maximum feature similarity.
  • determining the target similarity is to determine the current target feature vector which is the same tracking target as the reference target feature vector, compare all the calculated feature similarities, and determine the largest feature similarity as the reference target feature. The target similarity corresponding to the vector.
  • the server detects that there are 3 reference tracking targets in the reference image and obtains the corresponding 3 reference target feature vectors, D, E, and F, respectively. It detects 3 current tracking targets in the current image and obtains The corresponding three current target tracking vectors are G, H, and I; the server can select the reference target feature vector D and use the Euclidean distance algorithm to separate the reference target feature vector D from the current target feature vectors G, H, and I, respectively. Calculate the distance, calculate the three measured distances to be 0.8, 0.5, and 0.1, and use the feature similarity calculation formula based on the three measured distances The three feature similarities are calculated to be 0.56, 0.67, and 0.91. These three feature similarities are compared and sorted, and the ranking is 0.91>0.67> 0.56. The maximum value of 0.91 is selected as the target similarity corresponding to the reference target feature vector D.
  • the distance algorithm is used to calculate the measured distances between each reference target feature vector and all current target feature vectors, and the corresponding feature similarity is obtained based on the measured distances. The largest one among the calculated feature similarities is selected.
  • a target similarity corresponding to the reference target feature vector realizes the fast calculation of the target similarity corresponding to the reference target feature vector and improves the efficiency of target tracking.
  • the target real-time tracking method further includes the following steps:
  • the new tracking target refers to the current tracking target in the current image that does not match the same tracking target as all the reference tracking targets. That is, in the current image, the target similarity between the current target feature vector corresponding to the newly added tracking target and all reference target feature vectors is less than a preset similarity.
  • a new tracking target may be a specific target that appears in the current image, such as a person or an animal; or it may be a baseline tracking target that has appeared before but is blocked.
  • the occluded reference tracking target refers to the reference tracking target whose reference tracking target is partially blocked by a building or by other objects.
  • To determine whether the current tracking target is an occluded reference tracking target it can be determined by calculating the coordinate distances between the current target feature vector corresponding to the new tracking target and all reference target feature vectors. If all coordinate distances are greater than a preset distance threshold , The newly added tracking target is determined as a reference tracking target that is not occluded; if a coordinate distance exists that is smaller than a preset distance threshold, the newly added tracking target is determined as the occluded reference tracking target.
  • the server uses the newly added tracking target to update the reference tracking target of the occlusion, which refers to the reference target feature vector of the updated tracking target of the occlusion. That is, if the server determines that the newly added tracking target is an occluded reference tracking target, it replaces the current target feature vector corresponding to the newly added tracking target with the reference target feature vector corresponding to the occluded reference tracking target. Specifically, it updates the monitoring thread of the benchmark tracking target, and updates the feature vector of the benchmark target in the tracking thread.
  • Setting a new tracking target as the baseline tracking target by the server means adding a new baseline tracking target. That is, if the server determines that the new tracking target is not an occluded reference tracking target, determines the new tracking target as a new reference tracking target, and sets the current target feature vector corresponding to the new tracking target as the new reference target feature vector, A new monitoring thread is created for the new baseline tracking target, that is, a tracking thread and a verification thread are allocated.
  • a new tracking target is determined in the current image, and it is determined whether the new tracking target is an occluded reference tracking target. If it is determined that it is an occluded reference tracking target, then the Monitoring thread; if it is determined to be a new tracking target, add a baseline tracking target and add a new monitoring thread.
  • the server can determine whether the newly added tracking target is an occluded reference tracking target and separately process it according to the judgment result, so that it can re-track the occluded reference tracking target and create new monitoring for the newly added tracking target. Thread to monitor the newly added tracking target and improve the accuracy and efficiency of target tracking.
  • step S82 determines whether the newly added tracking target is a reference tracking target for occlusion, which specifically includes the following steps:
  • S821 Acquire a target position coordinate corresponding to the newly added tracking target and a reference position coordinate corresponding to each reference target feature vector.
  • a rectangular coordinate system is constructed in the current image, so as to obtain the target position coordinates of the current target feature vector corresponding to the newly added tracking target in the current image and the reference of the reference target feature vector corresponding to the reference target feature vector in the current image.
  • Position coordinates The target position coordinate is the position information of the newly added tracking target in the current image
  • the reference position coordinate is the position information of the reference tracking target in the reference image.
  • the position coordinates of the center point of the detection window may be adopted as the position coordinates corresponding to the target feature vector obtained based on the detection window.
  • S822 Calculate the corresponding position distance by using the target position coordinates and the reference position coordinates.
  • the position distance is a distance between the target position coordinates and the reference position coordinates.
  • the position distance can be calculated by using a distance algorithm to calculate the target position coordinates and the reference position coordinates.
  • the server uses a distance algorithm to calculate a target position coordinate corresponding to the newly added tracking target and a reference position coordinate of each reference position coordinate, thereby obtaining a position distance between the newly added tracking target and each reference tracking target. Understandably, if there are N reference tracking targets, there are also N position distances between the target position coordinates of any newly added tracking target and the reference position coordinates of the N reference tracking targets.
  • the tracking threshold is an index used to determine whether the newly added tracking target is a reference tracking target for occlusion.
  • the position distances of the target position coordinates of a newly added tracking target and the reference position coordinates of all the reference tracking targets are greater than the distance threshold, it indicates that the new tracking target is a newly appeared target in the current image and does not belong to the occluded reference. Tracking target.
  • the new tracking target can be added as a new reference tracking target, and the corresponding current target feature vector can be added as the reference target feature vector.
  • the corresponding position distance is calculated based on the target position coordinates of the newly added tracking target and the reference position coordinates of each reference tracking target, and the position distance is compared with a preset distance threshold to determine the new tracking. Whether the target is a baseline tracking target for occlusion. If all the position distances are greater than the preset distance threshold, it is determined that the newly added tracking target is not the occluded reference tracking target, and judging the newly added tracking target can realize multi-target tracking and improve target tracking efficiency; if it exists When a position distance is less than a preset distance threshold, the newly added tracking target is determined as the blocked reference tracking target, and the monitoring thread of the blocked reference tracking target is updated to judge the blocked reference tracking target to prevent the reference tracking target. Lost tracking improves the accuracy of target tracking. This real-time target tracking method solves the problem that the tracking target loses tracking due to occlusion and improves the accuracy of target tracking.
  • FIG. 7 shows a principle block diagram of a target real-time tracking device that corresponds to the target real-time tracking method in the one-to-one embodiment.
  • the target real-time tracking device includes an original video data acquisition module 10, a reference tracking target detection module 20, a current tracking target detection module 30, a target similarity acquisition module 40, a missing tracking target determination module 50, and a reference tracking target.
  • the release module 60 is described in detail as follows:
  • the original video data obtaining module 10 is configured to obtain original video data, where the original video data includes at least two frames of images.
  • the reference tracking target detection module 20 is configured to select any frame image in the original video data as a reference image, perform target detection on the reference image, and obtain at least one reference tracking target and a corresponding reference target feature vector.
  • the current tracking target detection module 30 is configured to perform target detection on the current image in the original video data, and obtain at least one current tracking target and a corresponding current target feature vector.
  • the target similarity acquisition module 40 is configured to calculate feature similarities between any reference target feature vector and all current target feature vectors to determine a target similarity corresponding to the reference target feature vector.
  • the lost tracking target determination module 50 is configured to determine that the reference tracking target corresponding to the target similarity is a lost tracking target in the current image if the target similarity is less than a preset similarity.
  • the reference tracking target release module 60 is configured to release the reference tracking target if consecutive N frames of images after the current image in the reference tracking target are lost tracking targets.
  • the target real-time tracking device further includes: a reference tracking target update module 70, configured to: if the feature similarity is not less than a preset similarity, a current tracking target that is the same tracking target as the reference tracking target exists in the current image, and The tracking target is updated with the new baseline tracking target.
  • a reference tracking target update module 70 configured to: if the feature similarity is not less than a preset similarity, a current tracking target that is the same tracking target as the reference tracking target exists in the current image, and The tracking target is updated with the new baseline tracking target.
  • the reference tracking target detection module 20 includes a detection window acquisition unit 21 and a reference target feature vector acquisition unit 22.
  • the detection window acquisition unit 21 is configured to perform target detection on the reference image by using a target detection algorithm, and acquire at least one detection window.
  • the reference target feature vector obtaining unit 22 is configured to obtain a corresponding reference tracking target and a reference target feature vector based on each detection window.
  • the target real-time tracking device further includes: a thread allocation unit 201, a first thread processing unit 202, and a second thread processing unit 203.
  • the thread allocation unit 201 is configured to allocate a tracking thread and a verification thread to each reference tracking target.
  • the first thread processing unit 202 is configured to acquire and store a reference target feature vector corresponding to a reference tracking target in real time in a tracking thread.
  • the second thread processing unit 203 is configured to determine whether the target similarity is less than a preset similarity in the verification thread, obtain a judgment result, and update the tracking thread according to the judgment result.
  • the feature similarity acquisition module 40 includes a feature similarity acquisition unit 41 and a target similarity selection unit 42.
  • a feature similarity obtaining unit 41 is configured to calculate a measured distance between any reference target feature vector and all current target feature vectors by using a distance algorithm, and obtain a corresponding feature similarity based on the measured distance.
  • the target similarity selection unit 42 is configured to select a maximum value from all feature similarities and determine the target similarity corresponding to the reference target feature vector.
  • the target real-time tracking device further includes: a new tracking target determination unit 81, an occlusion target determination unit 82, a first occlusion target update unit 83, and a second occlusion target update unit 84.
  • a new tracking target determination unit 81 is configured to determine, in the current image, a current tracking target that is not the same tracking target as all of the reference tracking targets as a new tracking target.
  • the occlusion target determination unit 82 is configured to determine whether the newly added tracking target is a reference tracking target for occlusion.
  • the first occlusion target update unit 83 is configured to update the occluded reference tracking target by using the newly added tracking target if the new tracking target is an occluded reference tracking target.
  • the second occlusion target updating unit 84 is configured to set the newly added tracking target as the reference tracking target if the newly added tracking target is not the occluded reference tracking target.
  • the occlusion target determination unit 82 includes a position coordinate acquisition subunit 821, a position distance acquisition subunit 822, an occlusion target non-stator unit 823, and an occlusion target determination subunit 824.
  • the position coordinate acquisition subunit 821 is configured to acquire a target position coordinate corresponding to the newly added tracking target and a reference position coordinate corresponding to each reference target feature vector.
  • the position and distance acquisition subunit 822 is configured to calculate a corresponding position distance by using the target position coordinates and the reference position coordinates.
  • the occlusion target non-stator unit 823 is used to add a new tracking target that is not a reference tracking target for occlusion if the distance between all positions is greater than the tracking threshold.
  • the occlusion target determination sub-unit 824 is configured to add a new tracking target as a reference tracking target for occlusion if any one of all position distances is not greater than a tracking threshold.
  • Each module in the above-mentioned target real-time tracking device may be implemented in whole or in part by software, hardware, and a combination thereof.
  • the above-mentioned modules may be embedded in the hardware in or independent of the processor in the computer device, or may be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure diagram may be as shown in FIG. 8.
  • the computer device includes a processor, a memory, a network interface, and a database connected through a system bus.
  • the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium and an internal memory.
  • the non-volatile storage medium stores an operating system, computer-readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer-readable instructions in a non-volatile storage medium.
  • the database of the computer equipment is used to store data called or generated during the execution of computer-readable instructions.
  • the network interface of the computer device is used to communicate with an external terminal through a network connection.
  • the computer-readable instructions are executed by a processor to implement a target real-time tracking method.
  • a computer device including a memory, a processor, and computer-readable instructions stored on the memory and executable on the processor.
  • the processor executes the computer-readable instructions, the following steps are implemented: Video data, the original video data includes at least two frames of images; select a reference image from the original video data, perform target detection on the reference image, and obtain at least one reference tracking target and the corresponding reference target feature vector; the current image in the original video data Perform target detection to obtain at least one current tracking target and the corresponding current target feature vector; calculate the feature similarity between any reference target feature vector and all current target feature vectors to determine the target similarity corresponding to the reference target feature vector; if the target If the similarity is less than the preset similarity, the reference tracking target corresponding to the target similarity is determined to be a lost tracking target in the current image; if the consecutive N frames of the reference tracking target after the current image are lost tracking targets, the reference tracking is released aims.
  • the processor executes the computer-readable instructions, the following steps are further implemented: if the target similarity is not less than a preset similarity, a current tracking target that is the same tracking target as the reference tracking target exists in the current image, and the current tracking target is The tracking target is updated with the new baseline tracking target.
  • the processor when the processor executes the computer-readable instructions, the processor further implements the following steps: performing target detection on the reference image by using a target detection algorithm to obtain at least one detection window; and acquiring a corresponding reference tracking target and reference based on each detection window Target feature vector.
  • the processor when the processor executes the computer-readable instructions, the processor further implements the following steps: allocating a tracking thread and a verification thread to each benchmark tracking target; obtaining and storing the benchmark target corresponding to the benchmark tracking target in real time in the tracking thread Feature vector; in the verification thread, determine whether the target similarity is less than a preset similarity, obtain a judgment result, and update the tracking thread according to the judgment result.
  • the processor when the processor executes the computer-readable instructions, the processor further implements the following steps: using a distance algorithm to calculate the measured distances between any reference target feature vector and all current target feature vectors, and obtaining corresponding feature similarities based on the measured distances; The maximum value of all feature similarities is selected and determined as the target similarity corresponding to the reference target feature vector.
  • the processor executes the computer-readable instructions, the following steps are further implemented: in the current image, determining a current tracking target that is not the same as all the reference tracking targets as a new tracking target; Whether the tracking target is an occluded reference tracking target; if the newly added tracking target is an occluded reference tracking target, the new tracking target is used to update the occluded reference tracking target; if the new tracking target is not an occluded reference tracking target, the Added a new tracking target as the baseline tracking target.
  • the processor executes the computer-readable instructions, the following steps are further implemented: obtaining target position coordinates corresponding to the newly added tracking target and reference position coordinates corresponding to each reference target feature vector; using target position coordinates and reference position coordinates Calculate the corresponding position distance; if all the position distances are greater than the tracking threshold, the newly added tracking target is not the reference tracking target for occlusion; if any of the position distances is not greater than the tracking threshold, the new tracking target is the occlusion reference Track your goals.
  • one or more non-volatile readable storage media storing computer readable instructions are provided.
  • the computer readable storage medium stores computer readable instructions, and the computer readable instructions are processed by one or more processors. When executed, cause one or more processors to perform the following steps: obtain original video data, the original video data includes at least two frames of image; select a reference image from the original video data, perform target detection on the reference image, and obtain at least one reference tracking target And the corresponding reference target feature vector; perform target detection on the current image in the original video data to obtain at least one current tracking target and the corresponding current target feature vector; calculate the characteristics of any reference target feature vector similar to all current target feature vectors Degree to determine the target similarity corresponding to the reference target feature vector; if the target similarity is less than a preset similarity, determine that the reference tracking target corresponding to the target similarity is a missing tracking target in the current image; if the reference tracking target is in the current image The subsequent consecutive N frames of images are discarded If the tracking target is lost, the baseline tracking target is
  • the one or more processors when the computer-readable instructions are executed by one or more processors, the one or more processors further perform the following steps: if the target similarity is not less than a preset similarity, there is a The tracking target is the current tracking target of the same tracking target, and the current tracking target is updated to a new reference tracking target.
  • the one or more processors when the computer-readable instructions are executed by one or more processors, the one or more processors further perform the following steps: using a target detection algorithm to perform target detection on the reference image to obtain at least one detection window; based on For each detection window, a corresponding reference tracking target and a reference target feature vector are obtained.
  • the one or more processors when the computer-readable instructions are executed by one or more processors, the one or more processors further perform the following steps: allocating a tracking thread and a verification thread to each benchmark tracking target; in the tracking thread The benchmark target feature vector corresponding to the benchmark tracking target is obtained and stored in real time; in the verification thread, it is determined whether the target similarity is less than a preset similarity, a judgment result is obtained, and the tracking thread is updated according to the judgment result.
  • the one or more processors when the computer-readable instructions are executed by one or more processors, the one or more processors further perform the following steps: using a distance algorithm to calculate the actual measurement of any reference target feature vector and all current target feature vectors Distance, based on the measured distance to obtain the corresponding feature similarity; select the maximum value from all feature similarities, and determine the target similarity corresponding to the reference target feature vector.
  • the one or more processors when the computer-readable instructions are executed by one or more processors, the one or more processors further perform the following steps: in the current image, all reference tracking targets that are not the same tracking target as The current tracking target is determined as the new tracking target; it is determined whether the new tracking target is the occluded reference tracking target; if the new tracking target is the occluded reference tracking target, the new tracking target is used to update the occluded reference tracking target; If the increase tracking target is not the reference tracking target for occlusion, the new tracking target is set as the reference tracking target.
  • the one or more processors when the computer-readable instructions are executed by one or more processors, the one or more processors further perform the following steps: acquiring target position coordinates corresponding to the newly added tracking target and corresponding to each reference target feature vector The reference position coordinates are calculated using the target position coordinates and the reference position coordinates. If all the position distances are greater than the tracking threshold, the newly added tracking target is not an obstructed reference tracking target; if any of the position distances is not If it is greater than the tracking threshold, the newly added tracking target is the reference tracking target for occlusion.
  • Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous chain (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

La présente invention concerne un procédé et un appareil de suivi d'une cible en temps réel, un dispositif informatique et un support d'informations. Le procédé de suivi d'une cible en temps réel comprend les étapes suivantes : acquérir des données vidéo d'origine; sélectionner une image de référence à partir des données vidéo d'origine et détecter une cible pour obtenir au moins une cible de suivi de référence et un vecteur de caractéristiques correspondant de la cible de référence; détecter une cible sur l'image actuelle pour obtenir au moins une cible de suivi actuelle et un vecteur de caractéristiques correspondant de la cible actuelle; déterminer une similarité de caractéristiques de chaque cible de suivi de référence sur la base du vecteur de caractéristiques de la cible de référence et du vecteur de caractéristiques de la cible actuelle; si la similarité des caractéristiques est inférieure à une similarité prédéfinie, déterminer la cible de suivi de référence correspondant à la similarité des caractéristiques comme étant une cible de suivi manquante dans l'image actuelle; et si la cible de suivi de référence est la cible de suivi manquante dans N trames consécutives après l'image actuelle, libérer la cible de suivi de référence. Le procédé de suivi en temps réel de la cible permet de déterminer et de libérer une cible de suivi manquante et d'améliorer la précision et l'efficacité du suivi de la cible.
PCT/CN2018/105404 2018-06-11 2018-09-13 Procédé et appareil de suivi d'une cible en temps réel, dispositif informatique et support d'informations WO2019237536A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810594665.8A CN108985162B (zh) 2018-06-11 2018-06-11 目标实时跟踪方法、装置、计算机设备及存储介质
CN201810594665.8 2018-06-11

Publications (1)

Publication Number Publication Date
WO2019237536A1 true WO2019237536A1 (fr) 2019-12-19

Family

ID=64540169

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/105404 WO2019237536A1 (fr) 2018-06-11 2018-09-13 Procédé et appareil de suivi d'une cible en temps réel, dispositif informatique et support d'informations

Country Status (2)

Country Link
CN (1) CN108985162B (fr)
WO (1) WO2019237536A1 (fr)

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111160212A (zh) * 2019-12-24 2020-05-15 浙江大学 一种基于YOLOv3-Tiny改进的跟踪学习检测系统及方法
CN111368743A (zh) * 2020-03-05 2020-07-03 北京科技大学 一种基于监控视频的游泳馆深水区早期危险行为检测方法
CN111402296A (zh) * 2020-03-12 2020-07-10 浙江大华技术股份有限公司 基于摄像机、雷达的目标跟踪方法及相关装置
CN111627043A (zh) * 2020-04-13 2020-09-04 浙江工业大学 一种基于标志物和特征筛选器的简易人体曲线获取方法
CN111862161A (zh) * 2020-07-30 2020-10-30 浙江大华技术股份有限公司 一种目标跟踪处理方法及装置
CN111986237A (zh) * 2020-09-01 2020-11-24 安徽炬视科技有限公司 一种人数无关的实时多目标跟踪算法
CN112037256A (zh) * 2020-08-17 2020-12-04 中电科新型智慧城市研究院有限公司 目标跟踪方法、装置、终端设备及计算机可读存储介质
CN112037257A (zh) * 2020-08-20 2020-12-04 浙江大华技术股份有限公司 一种目标跟踪方法、终端及其计算机可读存储介质
CN112614168A (zh) * 2020-12-21 2021-04-06 浙江大华技术股份有限公司 一种目标人脸跟踪方法、装置、电子设备及存储介质
CN112651994A (zh) * 2020-12-18 2021-04-13 零八一电子集团有限公司 地面多目标跟踪方法
CN112884810A (zh) * 2021-03-18 2021-06-01 沈阳理工大学 一种基于YOLOv3的行人跟踪方法
CN113129333A (zh) * 2020-01-16 2021-07-16 舜宇光学（浙江）研究院有限公司 多目标实时跟踪方法及其系统和电子设备
CN113450385A (zh) * 2021-06-04 2021-09-28 广州大学 一种夜间工作工程机械视觉跟踪方法、装置及存储介质
CN113554064A (zh) * 2021-06-28 2021-10-26 北京百度网讯科技有限公司 图片的标注方法、装置及电子设备
CN116309726A (zh) * 2023-04-07 2023-06-23 北京甲板智慧科技有限公司 基于ar互动游戏场景的多目标实时跟踪方法及装置

Families Citing this family (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109658434B (zh) * 2018-12-26 2023-06-16 成都纵横自动化技术股份有限公司 一种目标跟踪的方法及装置
CN111507999B (zh) * 2019-01-30 2023-07-18 北京四维图新科技股份有限公司 一种基于fdsst算法的目标追踪方法及装置
CN111696130A (zh) * 2019-03-12 2020-09-22 北京京东尚科信息技术有限公司 目标跟踪方法、目标跟踪装置及计算机可读存储介质
CN110490902B (zh) * 2019-08-02 2022-06-14 西安天和防务技术股份有限公司 应用于智慧城市的目标跟踪方法、装置、计算机设备
CN110634153A (zh) * 2019-09-19 2019-12-31 上海眼控科技股份有限公司 目标跟踪模板更新方法、装置、计算机设备和存储介质
CN110706250B (zh) * 2019-09-27 2022-04-01 广东博智林机器人有限公司 一种对象的跟踪方法、装置、系统及存储介质
CN110781824B (zh) 2019-10-25 2023-03-14 阿波罗智联(北京)科技有限公司 目标检测及跟踪方法、装置、电子设备及存储介质
CN111354023A (zh) * 2020-03-09 2020-06-30 中振同辂(江苏)机器人有限公司 一种基于摄像头的视觉多目标跟踪方法
CN111429483A (zh) * 2020-03-31 2020-07-17 杭州博雅鸿图视频技术有限公司 高速跨摄像机多目标跟踪方法、系统、装置及存储介质
CN111627045B (zh) * 2020-05-06 2021-11-02 佳都科技集团股份有限公司 单镜头下的多行人在线跟踪方法、装置、设备及存储介质
CN111738063B (zh) * 2020-05-08 2023-04-18 华南理工大学 船舶目标跟踪方法、系统、计算机设备及存储介质
CN111654668B (zh) * 2020-05-26 2021-08-13 山东摄云信息技术有限公司 监控设备同步方法、装置及计算机终端
CN112634327A (zh) * 2020-12-21 2021-04-09 合肥讯图信息科技有限公司 基于YOLOv4模型的跟踪方法
CN112651996A (zh) * 2020-12-22 2021-04-13 南京市城市与交通规划设计研究院股份有限公司 目标检测跟踪方法、装置、电子设备和存储介质
CN113014958A (zh) * 2021-03-10 2021-06-22 平安科技（深圳）有限公司 视频传输处理方法、装置、计算机设备及存储介质
CN113591904A (zh) * 2021-06-17 2021-11-02 浙江大华技术股份有限公司 逗留时间统计方法、货品调整方法及相关装置
CN114549584A (zh) * 2022-01-28 2022-05-27 北京百度网讯科技有限公司 信息处理的方法、装置、电子设备及存储介质

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106033550A (zh) * 2015-03-16 2016-10-19 北京大学 目标跟踪方法及装置
CN108053427A (zh) * 2017-10-31 2018-05-18 深圳大学 一种基于KCF与Kalman的改进型多目标跟踪方法、系统及装置

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102881022B (zh) * 2012-07-20 2015-04-08 西安电子科技大学 基于在线学习的遮挡目标跟踪方法
CN106651908B (zh) * 2016-10-13 2020-03-31 北京科技大学 一种多运动目标跟踪方法
CN106920248A (zh) * 2017-01-19 2017-07-04 博康智能信息技术有限公司上海分公司 一种目标跟踪方法及装置
CN106778712B (zh) * 2017-03-01 2020-04-14 扬州大学 一种多目标检测与跟踪方法
CN108010067B (zh) * 2017-12-25 2018-12-07 北京航空航天大学 一种基于组合判断策略的视觉目标跟踪方法

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106033550A (zh) * 2015-03-16 2016-10-19 北京大学 目标跟踪方法及装置
CN108053427A (zh) * 2017-10-31 2018-05-18 深圳大学 一种基于KCF与Kalman的改进型多目标跟踪方法、系统及装置

Cited By (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111160212A (zh) * 2019-12-24 2020-05-15 浙江大学 一种基于YOLOv3-Tiny改进的跟踪学习检测系统及方法
CN113129333A (zh) * 2020-01-16 2021-07-16 舜宇光学（浙江）研究院有限公司 多目标实时跟踪方法及其系统和电子设备
CN113129333B (zh) * 2020-01-16 2023-06-16 舜宇光学（浙江）研究院有限公司 多目标实时跟踪方法及其系统和电子设备
CN111368743A (zh) * 2020-03-05 2020-07-03 北京科技大学 一种基于监控视频的游泳馆深水区早期危险行为检测方法
CN111368743B (zh) * 2020-03-05 2023-08-04 北京科技大学 一种基于监控视频的游泳馆深水区早期危险行为检测方法
CN111402296A (zh) * 2020-03-12 2020-07-10 浙江大华技术股份有限公司 基于摄像机、雷达的目标跟踪方法及相关装置
CN111402296B (zh) * 2020-03-12 2023-09-01 浙江大华技术股份有限公司 基于摄像机、雷达的目标跟踪方法及相关装置
CN111627043A (zh) * 2020-04-13 2020-09-04 浙江工业大学 一种基于标志物和特征筛选器的简易人体曲线获取方法
CN111627043B (zh) * 2020-04-13 2023-09-19 浙江工业大学 一种基于标志物和特征筛选器的简易人体曲线获取方法
CN111862161A (zh) * 2020-07-30 2020-10-30 浙江大华技术股份有限公司 一种目标跟踪处理方法及装置
CN112037256A (zh) * 2020-08-17 2020-12-04 中电科新型智慧城市研究院有限公司 目标跟踪方法、装置、终端设备及计算机可读存储介质
CN112037257A (zh) * 2020-08-20 2020-12-04 浙江大华技术股份有限公司 一种目标跟踪方法、终端及其计算机可读存储介质
CN112037257B (zh) * 2020-08-20 2023-09-29 浙江大华技术股份有限公司 一种目标跟踪方法、终端及其计算机可读存储介质
CN111986237A (zh) * 2020-09-01 2020-11-24 安徽炬视科技有限公司 一种人数无关的实时多目标跟踪算法
CN112651994A (zh) * 2020-12-18 2021-04-13 零八一电子集团有限公司 地面多目标跟踪方法
CN112614168A (zh) * 2020-12-21 2021-04-06 浙江大华技术股份有限公司 一种目标人脸跟踪方法、装置、电子设备及存储介质
CN112614168B (zh) * 2020-12-21 2023-08-29 浙江大华技术股份有限公司 一种目标人脸跟踪方法、装置、电子设备及存储介质
CN112884810A (zh) * 2021-03-18 2021-06-01 沈阳理工大学 一种基于YOLOv3的行人跟踪方法
CN112884810B (zh) * 2021-03-18 2024-02-02 沈阳理工大学 一种基于YOLOv3的行人跟踪方法
CN113450385A (zh) * 2021-06-04 2021-09-28 广州大学 一种夜间工作工程机械视觉跟踪方法、装置及存储介质
CN113450385B (zh) * 2021-06-04 2023-08-18 广州大学 一种夜间工作工程机械视觉跟踪方法、装置及存储介质
CN113554064B (zh) * 2021-06-28 2023-07-25 北京百度网讯科技有限公司 图片的标注方法、装置及电子设备
CN113554064A (zh) * 2021-06-28 2021-10-26 北京百度网讯科技有限公司 图片的标注方法、装置及电子设备
CN116309726A (zh) * 2023-04-07 2023-06-23 北京甲板智慧科技有限公司 基于ar互动游戏场景的多目标实时跟踪方法及装置

Also Published As

Publication number Publication date
CN108985162A (zh) 2018-12-11
CN108985162B (zh) 2023-04-18

Similar Documents

Publication Publication Date Title
WO2019237536A1 (fr) Procédé et appareil de suivi d'une cible en temps réel, dispositif informatique et support d'informations
US10192107B2 (en) Object detection method and object detection apparatus
CN109035299B (zh) 目标跟踪方法、装置、计算机设备及存储介质
CN111242977B (zh) 全景视频的目标跟踪方法、可读存储介质及计算机设备
US20190362503A1 (en) Target tracking method and apparatus, electronic device, and storage medium
US8965107B1 (en) Feature reduction based on local densities for bundle adjustment of images
CN110751674A (zh) 多目标跟踪方法及相应视频分析系统
WO2020094088A1 (fr) Procédé de capture d'image, caméra de surveillance et système de surveillance
US11900676B2 (en) Method and apparatus for detecting target in video, computing device, and storage medium
CN108961316B (zh) 图像处理方法、装置及服务器
JP2020515974A (ja) 歩行者追跡の方法および電子デバイス
US20240078680A1 (en) Image segmentation method, network training method, electronic equipment and storage medium
US10936876B2 (en) Content recognition method, system and storage medium thereof
KR20210027778A (ko) 객체 검출 및 추적을 통한 이상행동 분석 장치 및 방법
WO2020024394A1 (fr) Procédé et dispositif d'élimination d'arrière-plan, dispositif informatique et support de stockage
WO2019033575A1 (fr) Dispositif électronique, procédé et système de suivi de visage, et support d'informations
JP7330708B2 (ja) 画像処理装置、画像処理方法、及びプログラム
CN111159476B (zh) 目标对象的搜索方法、装置、计算机设备及存储介质
KR102022971B1 (ko) 영상의 객체 처리 방법 및 장치
WO2022142416A1 (fr) Procédé de suivi de cible et dispositif associé
CN108564045B (zh) 增强现实的数据处理方法、装置、存储介质和计算机设备
CN114092720A (zh) 目标跟踪方法、装置、计算机设备和存储介质
JP6734820B2 (ja) 映像検索装置
CN112750145A (zh) 目标检测与跟踪的方法、装置和系统
CN117474983B (zh) 一种基于光视联动的预警方法以及相关装置

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: 18922961

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS (EPO FORM 1205A DATED 19.03.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18922961

Country of ref document: EP

Kind code of ref document: A1