A personnel re-tracking method based on image detection technology
By employing a person re-tracking method based on image detection technology, and utilizing human detection and skeletal point verification to construct a multi-stage closed-loop mechanism, this method solves the problems of high target false detection rate, low re-identification accuracy after tracking loss, and weak anti-interference ability in existing technologies. It achieves high-precision person target detection and continuous tracking, while reducing hardware costs and response time.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIACHEN YUNKONG NEW ENERGY (SHANGHAI) CO LTD
- Filing Date
- 2026-03-11
- Publication Date
- 2026-05-29
AI Technical Summary
Existing personnel tracking technologies suffer from problems such as high false detection rate of targets, low accuracy of re-identification after tracking loss, and weak anti-interference ability.
A person re-tracking method based on image detection technology is adopted. By acquiring monitoring images in real time, human detection and skeletal point verification are performed, a distance relationship table is constructed, and a multi-stage closed-loop tracking mechanism is realized. High-confidence human targets are screened out, and human skeletal point features are used for tracking and re-identification.
It achieves high-precision human target detection and continuous tracking, reduces interference from non-human targets, effectively resists environmental interference, reduces hardware computing power requirements, reduces deployment costs, and ensures tracking continuity and re-identification response time.
Smart Images
Figure CN122116480A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of personnel re-tracking, and more specifically to a personnel re-tracking method based on image detection technology. Background Technology
[0002] With the continuous and in-depth research of electronic and sensor technologies, a large number of autonomous following robots have been widely used in various fields. Since these robots need to operate fully autonomously without human control, from target recognition and path following to obstacle avoidance and interaction, they rely entirely on target tracking technology for core support. If there is a lack of high-confidence target screening and human feature verification mechanisms, non-human objects may be misjudged as tracking objects, leading to work interruption or service failure. If the target is lost during the tracking process due to factors such as occlusion or rapid movement, and it cannot be accurately re-identified through stable features, it will directly affect the robot's autonomous service capabilities. Therefore, target tracking technology is a key core technology to ensure the stable and efficient operation of autonomous following robots in industrial manufacturing and home service scenarios.
[0003] In the field of personnel target detection and re-tracking technology, with the increasing demands for tracking accuracy and continuity in applications such as security monitoring and intelligent robots, related technical solutions have been continuously iterated. However, the two mainstream core technologies still have significant shortcomings and are difficult to meet the application needs of complex real-world scenarios. One type of solution is to introduce a ReID (Person Re-identification) model. This solution integrates a third-party pre-trained model that covers multi-dimensional visual information such as target color features, shape contour features, and local texture features to detect and identify targets in the monitoring screen. It uses the model's deep extraction capability of target visual features to achieve tracking and matching. However, this technical approach places extremely high demands on hardware performance. Not only does it require devices equipped with high-performance processors and large-capacity video memory to support the model's real-time inference operations, significantly increasing the deployment cost and hardware threshold, but more importantly, its core color feature channel is highly susceptible to interference from various environmental factors in real-world scenarios. Such as drastic changes in light intensity, shadow occlusion, reflection and superposition of complex background colors, and the convergence of the target's clothing color with the surrounding environment, all of these can lead to color feature extraction distortion, resulting in target matching failure and causing... Interruptions in the tracking link cannot meet the requirements of scenarios with high tracking continuity. Another mainstream solution is to use the Kalman prediction algorithm. This method establishes a mathematical prediction model based on the target's historical motion state (such as position coordinates, motion speed, acceleration, etc.) and estimates the target's trajectory at the next moment through recursive calculations, thereby completing tracking confirmation. Although this solution has certain advantages in hardware adaptability and does not rely on high computing power, its tracking logic relies excessively on the target's historical motion data. When the target is lost due to prolonged occlusion, rapid change of direction, sudden stillness, or leaving the monitoring field of view, the method of prediction based solely on historical motion trajectory data is completely unable to cope with possible changes in attitude and direction of motion when the target reappears. This results in extremely low accuracy of re-tracking and often misidentifies other moving objects in the scene as the original tracked target, causing erroneous tracking problems. This not only affects the accuracy of the tracking results but may also lead to serious deviations in subsequent event analysis and security warnings based on tracking data. As a result, existing personnel tracking technologies suffer from high target false detection rates, low re-identification accuracy after tracking loss, and weak anti-interference capabilities. Summary of the Invention
[0004] To address the technical problems of high false detection rate, low re-identification accuracy after tracking loss, and weak anti-interference capability in existing personnel tracking technologies, the present invention aims to provide a personnel re-tracking method based on image detection technology. The specific technical solution adopted is as follows:
[0005] This invention proposes a method for re-tracking people based on image detection technology, the method comprising:
[0006] Real-time acquisition of surveillance images of the monitored area;
[0007] Perform human detection on the current frame image and filter out the tracking targets in the current frame image;
[0008] Feature point detection is performed on the tracked target in the current frame image to obtain the skeleton points of the tracked target; based on the number of skeleton points, it is determined whether the tracked target is a human target. If it is not a human target, the next frame image is processed; otherwise, a distance relationship table containing the distances between the skeleton points of the human target is generated for the current frame image, and the next frame image is processed and the distance relationship table is updated in real time.
[0009] Based on the number of images containing the human target, it is determined whether the human target is lost. If it is not lost, the human target is continuously tracked. Otherwise, the same processing is performed on the new monitoring images of the reacquired monitoring area, and the human target in the new monitoring images is tracked based on the difference between the new distance relationship table and the historical distance relationship table.
[0010] Furthermore, the step of filtering out the tracking targets in the current frame image includes:
[0011] The current frame image is input into the preset human target detection model, and the detected targets in the current frame image and the confidence parameters of the detected targets are output.
[0012] Based on the confidence parameter, the target to be tracked is selected from the detected targets in the current frame image.
[0013] Furthermore, the step of filtering out the tracking target from the detected targets in the current frame image based on the confidence parameter includes:
[0014] In the current frame image, the detected targets whose confidence parameters are greater than or equal to the preset confidence threshold are taken as the tracking targets.
[0015] Furthermore, obtaining the skeletal points of the tracked target includes:
[0016] The tracking target in the current frame image is cropped to obtain a cropped region containing the tracking target;
[0017] The cropped area is input into a preset human skeleton point detection model, and the skeleton points of the tracking target are output.
[0018] Furthermore, determining whether the tracking target is a human target includes:
[0019] If the number of skeletal points of the tracked target in the current frame image is equal to the number of elements contained in the complete set of skeletal points, then the tracked target in the current frame image is a human target; otherwise, it is not a human target.
[0020] Furthermore, generating a distance relationship table for the current frame image, which includes the distances between skeletal points of the human target, includes:
[0021] Select multiple preset key bone points from all the bone points of the human target in the current frame image;
[0022] The three-dimensional spatial coordinates of the key skeletal points of the human target in the current frame image are acquired using a depth camera;
[0023] Take any two key skeleton points as a group, and use the Euclidean distance between the three-dimensional spatial coordinates of the two key skeleton points in each key skeleton point group as the distance coefficient of each key skeleton point group.
[0024] Record each key skeleton point group and the distance coefficient of each key skeleton point group in the distance relationship table.
[0025] Furthermore, determining whether the human target is lost includes:
[0026] If the human target is not detected in a consecutive preset number of frames, it is determined that the human target is lost; otherwise, it is determined that the human target is not lost.
[0027] Furthermore, the tracking of human targets in the new surveillance images includes:
[0028] Based on the difference in distance coefficients between the new distance relationship table and the historical distance relationship table for the same key skeleton point group, the feature similarity of human targets in the new surveillance image is obtained;
[0029] Based on the feature similarity, it is determined whether the human target in the new monitoring image is the same human target that was lost before. If it is the same human target that was lost before, the human target in the new monitoring image is continuously tracked. Otherwise, the human target in the new monitoring image is tracked as a new target.
[0030] Furthermore, obtaining the feature similarity of human targets in the new surveillance image includes:
[0031] The absolute value of the difference between the distance coefficients of the same key skeleton point group between the new distance relationship table and the historical distance relationship table is negatively correlated and normalized to obtain the similarity coefficient of the same key skeleton point group between the new distance relationship table and the historical distance relationship table.
[0032] The average of the similarity coefficients of all identical key skeleton point groups between the new distance relationship table and the historical distance relationship table is used as the feature similarity of the human target in the new surveillance image.
[0033] Furthermore, determining whether the human target in the new surveillance image is the same human target that was lost before, based on the feature similarity, includes:
[0034] If the feature similarity is greater than or equal to a preset similarity threshold, then the human target in the new surveillance image is the human target that was lost before; otherwise, the human target in the new surveillance image is not the human target that was lost before.
[0035] The present invention has the following beneficial effects:
[0036] This invention addresses the problems of high false detection rates, low re-identification accuracy after tracking loss, and weak anti-interference capabilities in existing personnel tracking technologies. It constructs a multi-stage closed-loop tracking mechanism—"confidence screening - skeletal point verification - feature comparison"—to achieve accurate detection and continuous tracking of personnel targets. First, human detection is performed on the current frame image to filter out tracking targets, directly filtering out low-confidence interference items to reduce invalid tracking at the source. Then, human skeletal point features are used to accurately identify human targets, reducing interference from non-human targets and allowing system resources to focus on valid targets. Furthermore, the distance between key human skeletal points is used as a feature identifier to re-identify lost targets, effectively resisting environmental interference such as changes in lighting and posture. This approach requires no high-performance hardware, reducing deployment costs. The complete closed-loop process ensures tracking continuity, and the re-identification response time is low. Attached Figure Description
[0037] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 A flowchart illustrating a person re-tracking method based on image detection technology, provided as an embodiment of the present invention;
[0039] Figure 2 This is a schematic diagram of the logical structure of a person re-tracking method based on image detection technology in a specific scenario, as provided in an embodiment of the present invention. Detailed Implementation
[0040] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a person re-tracking method based on image detection technology proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0042] Example 1
[0043] The following description, in conjunction with the accompanying drawings, details a specific scheme for a personnel re-tracking method based on image detection technology provided by the present invention.
[0044] Please see Figure 1 The diagram illustrates a flowchart of a person re-tracking method based on image detection technology according to an embodiment of the present invention, the method comprising:
[0045] Step S1: Acquire real-time monitoring images of the monitored area.
[0046] In this embodiment of the invention, the vehicle-mounted image acquisition device (e.g., a high-definition camera) is first used to acquire real-time monitoring images of the monitored area. The monitoring images are then preprocessed, such as distortion correction and noise reduction, to improve the image quality of the camera images. This enhances the clarity and recognizability of objects in the monitoring images, facilitating the accurate identification and tracking of personnel targets in the future. The image distortion correction and noise reduction preprocessing methods are well-known to those skilled in the art and will not be described in detail here.
[0047] Step S2: Perform human detection on the current frame image and filter out the tracking targets in the current frame image.
[0048] In the monitored area, there are not only human targets but also non-human targets. Therefore, this embodiment of the invention needs to start the analysis from the first frame of the monitoring image, take the first frame of the monitoring image as the current frame image, and perform human detection on the current frame image to quickly and initially screen out the tracking targets in the current frame image. The tracking targets include the human targets that need to be identified in this tracking task. Subsequently, the tracking targets can be further screened more precisely to accurately identify human targets in the monitored area environment.
[0049] Preferably, in one embodiment of the present invention, the method for obtaining the tracking target in the current frame image specifically includes:
[0050] First, the current frame image is input into a preset human target detection model. The human target detection model outputs the detected targets in the current frame image and the confidence parameters of the detected targets. The confidence parameter ranges from 0 to 1. The higher the confidence parameter of a detected target, the more likely the detected target is to be a human target. Therefore, based on the confidence parameter, tracking targets can be further filtered from the detected targets in the current frame image. In one embodiment of the present invention, the preset human target detection model can be, for example, a trained YOLO model or an SSD model. This model can output the bounding rectangles of all suspected human targets (i.e., detected targets) in the image and their corresponding confidence scores. The confidence score represents the degree to which the detected suspected human targets match real human bodies. This is not limited or elaborated here.
[0051] Preferably, in one embodiment of the present invention, the method for acquiring the tracking target specifically includes:
[0052] The above analysis shows that the higher the confidence parameter of a detected target, the more likely the target is to be a person or human body. Therefore, in the current frame image, targets with a confidence parameter greater than or equal to a preset confidence threshold can be used as tracking targets. The preset confidence threshold typically ranges from close to 1. In order to more directly and effectively eliminate non-human interference, in one embodiment of the present invention, the preset reliability threshold is set to 0.9. The specific value of the preset reliability threshold can also be set by the implementer according to the specific implementation scenario, and is not limited here.
[0053] Step S3: Perform feature point detection on the tracked target in the current frame image to obtain the skeleton points of the tracked target; determine whether the tracked target is a human target based on the number of skeleton points. If it is not a human target, process the next frame image; otherwise, generate a distance relationship table containing the distances between the skeleton points of the human target for the current frame image, process the next frame image, and update the distance relationship table in real time.
[0054] In order to more accurately identify human targets in images and reduce interference from non-human targets during the identification process, it is also necessary to detect the unique features of the human body. In this embodiment of the invention, feature point detection is first performed on the tracked target in the current frame image to obtain the skeletal points of the tracked target. Since the number and relative position of the human body's skeletal points are key and unique features for identifying the human body, the skeletal points of the tracked target in the current frame image can be further verified based on the skeletal point features to accurately identify the human target.
[0055] Preferably, in one embodiment of the present invention, the method for obtaining the skeletal points of the tracked target in the current frame image specifically includes:
[0056] First, the tracking target in the current frame image is cropped to obtain a cropped region containing the tracking target. The cropped region mainly contains the tracking target, thereby avoiding interference from other objects in the feature point detection of the tracking target in the cropped region. Image cropping is a well-known technique in the art, and can be implemented using existing edge detection algorithms, thresholding or semantic segmentation algorithms, etc.
[0057] The cropped area is then input into a preset human skeleton point detection model, which outputs the skeleton points of the target. Since the human skeleton points are less affected by environmental interference such as changes in lighting and posture, the human target can be accurately analyzed based on the skeleton point features, and the accuracy of re-identifying lost targets can be improved. The human skeleton point detection model can automatically identify and locate the positions of human joints such as the head, neck, shoulders, and elbows from the image. Common human skeleton point detection models can detect 17, 25, or even more points. In one embodiment of the present invention, the preset human skeleton point detection model can be selected from existing OpenPose models or HRNet models, etc., which are not limited or described in detail here.
[0058] For human targets in the current frame image, complete human skeleton points can be detected after skeletal point detection. However, for non-human targets, complete human skeleton points cannot be detected after skeletal point detection. Therefore, based on the number of skeleton points, it can be further determined whether the tracked target is a human target, thereby reducing interference from non-human targets and enabling system resources to focus on serving valid targets. Subsequently, different processing can be performed on the monitoring image based on the judgment results.
[0059] Preferably, in one embodiment of the present invention, the method for determining whether the tracking target is a human target specifically includes:
[0060] If the number of skeletal points of the tracked target in the current frame image is equal to the number of elements contained in the complete skeletal point set, then the tracked target in the current frame image is a human target; otherwise, it is not a human target. The complete skeletal point set contains all the key joints on the human body. In one embodiment of the present invention, the elements of the complete skeletal point set include head points, neck points, left / right shoulder points, left / right elbow points, left / right wrist points, left / right hip points, left / right knee points, and left / right ankle points. In other embodiments of the present invention, the elements in the complete skeletal point set may also be added or modified by the implementer according to the specific implementation scenario, which is not limited here.
[0061] After judgment, if the tracking target in the current frame image is not a human target, it means that there is no human target in the current frame image. In this case, the same processing needs to be performed on the next frame image, that is, the same method is used to continue to detect human targets in the next frame image, so as to ensure the continuity of human detection and tracking. Otherwise, it means that the tracking target in the current frame image is a human target. In this case, a distance relationship table containing the distance between the skeleton points of the human target is generated for the current frame image, and the next frame image is processed. Since people in the monitoring environment are usually in dynamic motion, the relative positions of their skeleton points will change. Therefore, the distance relationship table is also updated in real time to ensure real-time tracking of people. At the same time, the distance relationship table can also be used for re-identification in the event of subsequent target loss.
[0062] Preferably, in one embodiment of the present invention, the method for obtaining the distance relationship table specifically includes:
[0063] First, select multiple preset key bone points from all the skeletal points of the human target in the current frame image. In one embodiment of the present invention, the preset key bone points are set as left / right shoulder bone points, left / right hip bone points and left / right knee bone points. The type and number of preset key bone points can also be set by the implementer according to the specific implementation scenario, and are not limited here.
[0064] Then, the depth camera is used to acquire the three-dimensional spatial coordinates of key skeletal points of the human target in the current frame image, and any two key skeletal points are grouped together. The Euclidean distance between the three-dimensional spatial coordinates of the two key skeletal points in each group is used as the distance coefficient of each group. The distance coefficient is the distance feature between two skeletal points. For the same person, the distance between skeletal points is generally not much different during dynamic movement. Therefore, the distance coefficients between skeletal points of moving human targets can be compared and analyzed to achieve the re-identification of lost targets.
[0065] Then, the distance coefficients of each key skeletal point group and each key skeletal point group are recorded in the distance relationship table.
[0066] Step S4: Based on the number of images containing human targets, determine whether the human targets are lost. If they are not lost, continue to track the human targets. Otherwise, perform the same processing on the new monitoring images of the newly acquired monitoring area, and track the human targets in the new monitoring images based on the difference between the new distance relationship table and the historical distance relationship table.
[0067] During the continuous tracking of human targets in the surveillance images, the tracked personnel may be lost due to obstruction by the external environment or exceeding the monitoring range. Therefore, this embodiment of the invention also needs to determine whether the human target has been lost based on the number of images containing the human target. Subsequently, the tracked target can be continuously tracked or re-identified based on the judgment result, thereby ensuring the continuity of personnel tracking.
[0068] Preferably, in one embodiment of the present invention, the method for determining whether a human target has been lost specifically includes:
[0069] If no human target is detected in a consecutive preset number of frames, it is determined that the human target is lost; otherwise, it is determined that the human target is not lost. The preset number is usually in the range of 3 to 5. In one embodiment of the present invention, the preset number is set to 3, that is, when no human target is detected in 3 consecutive frames, it is determined that the human target is lost. The specific value of the preset number can also be set by the implementer according to the specific implementation scenario, and is not limited here.
[0070] After judgment, if the human target is not lost, it can be continuously tracked. Otherwise, if any target is lost, it needs to be re-identified and tracked. Therefore, this embodiment of the invention uses the image acquisition device again to perform the same processing on the new monitoring image of the newly acquired monitoring area, that is, to perform human detection, feature point detection, and human target judgment processing on the new monitoring image, and generate a new distance relationship table corresponding to the new monitoring image, i.e., the new distance relationship table. The smaller the difference between the new distance relationship table and the distance relationship table before the target was lost, the more likely that the human target detected in the new monitoring image is the human target before it was lost. Therefore, the human target in the new monitoring image can be tracked according to the difference between the new distance relationship table and the historical distance relationship table, thereby restoring the re-identification and tracking of the lost target. At the same time, it can effectively resist environmental interference such as changes in lighting and posture, and does not require high computing power hardware support, reducing deployment costs. The complete closed-loop process ensures tracking continuity and has a low re-identification response time. The historical distance relationship table refers to the distance relationship table after the last update before the human target was lost.
[0071] Preferably, in one embodiment of the present invention, the method for tracking a human target in a new surveillance image specifically includes:
[0072] For the same human target, the relative positions of its human skeleton points do not change much before and after the loss, that is, the distance features between each skeleton point do not change much before and after the loss. Therefore, the feature similarity of human targets in the new monitoring image can be obtained by the difference in the distance coefficients of the same key skeleton point groups between the new distance relationship table and the historical distance relationship table. Subsequently, based on the feature similarity, it can be accurately determined whether the human target detected in the new monitoring image is the same human target before the loss.
[0073] Preferably, in one embodiment of the present invention, the method for obtaining the feature similarity of human targets in the new surveillance image specifically includes:
[0074] The absolute values of the differences in distance coefficients between the new distance relationship table and the historical distance relationship table for the same key skeleton point group are negatively correlated and normalized to limit the calculation results to within a certain range. Within the range, the similarity coefficient of the same key skeleton point group between the new distance relationship table and the historical distance relationship table is obtained.
[0075] In embodiments of the present invention, a negative exponential function with the natural constant e as its base can be used or The function form is used to perform normalization of negative correlations, which is not limited here. Wherein... The normalization function is represented by the normalization function. In one embodiment of the present invention, the normalization process can be, for example, maximum and minimum value normalization. In other embodiments of the present invention, other normalization methods can be selected according to the specific range of values, or activation functions and hyperbolic tangent functions can be used to implement the normalization process. These will not be elaborated or limited further.
[0076] The average similarity coefficient of all identical key skeleton point groups between the new distance relationship table and the historical distance relationship table is used as the feature similarity of human targets in the new surveillance image.
[0077] The greater the feature similarity, the closer the relative positions of the key skeletal points of the human target in the new surveillance image are to the relative positions of the key skeletal points of the human target before it was lost. This indicates that the human target in the new surveillance image is more likely to be the lost human target. Therefore, based on feature similarity, it can be determined whether the human target in the new surveillance image is the same human target as before it was lost. If it is the same human target as before it was lost, the human target in the new surveillance image will continue to be tracked to restore the re-identification and tracking of the lost target. Otherwise, the human target in the new surveillance image will be tracked as a new target.
[0078] Preferably, in one embodiment of the present invention, the method for determining whether a human target in a new surveillance image is the same human target that was lost before is specifically included:
[0079] If the feature similarity is greater than or equal to a preset similarity threshold, then the human target in the new surveillance image is the previously missing human target; otherwise, the human target in the new surveillance image is not the previously missing human target. The preset similarity threshold ranges from close to 1, typically being [value missing]. To avoid misidentification of lost targets, in one embodiment of the present invention, the preset confidence threshold is set to 0.95. The specific value of the preset confidence threshold can also be set by the implementer according to the specific implementation scenario, and is not limited here.
[0080] Example 2
[0081] Please see Figure 2 The diagram illustrates the logical structure of a person re-tracking method based on image detection technology provided by an embodiment of the present invention in a specific scenario.
[0082] This embodiment takes a vehicle-based personnel tracking system tracking personnel in an external environment as an example. The steps for re-tracking external personnel targets are as follows:
[0083] Step 1: The human body model is pre-trained and loaded into the tracking system hardware. The human body model is used to detect suspected targets and skeletal points. Commonly used human body models include YOLO, SSD, OpenPose, and HRNet models, thus preparing for subsequent target detection.
[0084] Step 2: Install hardware devices on the following vehicle to meet the needs of environmental monitoring. The hardware devices include high-definition cameras, storage media, etc.
[0085] Step 3: Start the vehicle and power it on to ensure that the entire equipment can work normally.
[0086] Step 4: Perform distortion correction, filtering, and noise reduction on the camera footage to improve image quality and facilitate accurate identification and tracking of personnel targets.
[0087] Step 5: Detect the target and output its position and confidence level in the image. The higher the confidence level, the more likely the detected target is a person or human body. Subsequently, based on the confidence level, suspected persons can be screened from the detected targets in the image. Thus, by using confidence level screening, the accuracy of personnel target detection can be initially guaranteed.
[0088] Step Six: Determine if the confidence level is greater than or equal to 90%. If yes, it means that the detected target may be a person, and proceed to Step Seven for further judgment. If not, return to Step Four to continue detecting possible person targets. This iterative step is used to ensure continuous detection and tracking of people in the monitored area.
[0089] Step 7: Identify the target to be tracked and detect the skeletal points. By utilizing the characteristics of human skeletal points, a secondary verification of the target can be achieved, further improving the accuracy of identifying human targets in the monitored area and reducing interference from non-human targets.
[0090] Step 8: If a skeletal point is detected, it means that the target initially screened in Step 6 can be basically identified as a human target, and then proceed to Step 9 for further processing; if not, the target initially screened in Step 6 can be identified as a non-human target, and then proceed to Step 4 to continue detecting human targets. This cyclical step is used to ensure continuous detection and tracking of people in the monitoring area. At the same time, the human skeletal point features can also effectively resist environmental interference such as changes in lighting and posture.
[0091] Step 9: Record the position coordinates of skeletal points 5, 6, 11, 12, 13, 14 (i.e., the preset key skeletal points) and generate a distance relationship table. The distance relationship table can be used to re-identify lost targets, thereby ensuring continuous tracking of the same person.
[0092] Step 10: Continue target detection to determine if the target is currently missing. If yes, proceed to Step 11; otherwise, proceed to Step 14. This iterative process ensures continuous detection and tracking of missing personnel targets within the monitored area.
[0093] Step 11: Re-collect surrounding target information, detect the targets, and determine the confidence level.
[0094] Step 12: Calculate the distance between the target skeleton points and compare it with the recorded distance relationship table. By comparing the distance features between skeleton points in the distance relationship table, it is possible to accurately identify whether the newly detected personnel target is a lost target, thus ensuring the continuous tracking of the lost target.
[0095] Step 13: Compare whether the similarity is greater than or equal to 90%. If yes, it means that the new target is a lost target, and proceed to Step 14; otherwise, proceed to Step 11 to continue detecting lost targets. This cyclical step is used to ensure the continuous detection and tracking of lost personnel targets in the monitored area.
[0096] Step Fourteen: Continuous Tracking Mode.
[0097] In summary, this embodiment of the invention first performs human detection on the current frame image, filtering out the tracking targets in the current frame image and directly filtering out low-confidence interference items to reduce invalid tracking from the source. Then, it uses human skeleton point features to accurately identify human targets, reducing interference from non-human targets and allowing system resources to focus on effective targets. Then, it uses the distance of key human skeleton points as feature identifiers to further improve the accuracy of human target identification in the monitoring area and reduce the false detection rate of human targets. It also uses the constructed distance relationship table to re-identify lost targets, thereby ensuring continuous tracking of lost targets. Furthermore, using human skeleton points as features can effectively resist environmental interference such as changes in lighting and posture, and does not require high computing power hardware support, reducing deployment costs. The complete closed-loop process ensures tracking continuity, and the re-identification response time is low, realizing sustainable real-time tracking of human targets in the monitoring area.
[0098] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0099] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
Claims
1. A method for re-tracking people based on image detection technology, characterized in that, The method includes: Real-time acquisition of surveillance images of the monitored area; Perform human detection on the current frame image and filter out the tracking targets in the current frame image; Feature point detection is performed on the tracked target in the current frame image to obtain the skeleton points of the tracked target; based on the number of skeleton points, it is determined whether the tracked target is a human target. If it is not a human target, the next frame image is processed; otherwise, a distance relationship table containing the distances between the skeleton points of the human target is generated for the current frame image, and the next frame image is processed and the distance relationship table is updated in real time. Based on the number of images containing the human target, it is determined whether the human target is lost. If it is not lost, the human target is continuously tracked. Otherwise, the same processing is performed on the new monitoring images of the reacquired monitoring area, and the human target in the new monitoring images is tracked based on the difference between the new distance relationship table and the historical distance relationship table.
2. The method for re-tracking people based on image detection technology according to claim 1, characterized in that, The filtering of tracking targets in the current frame image includes: The current frame image is input into the preset human target detection model, and the detected targets in the current frame image and the confidence parameters of the detected targets are output. Based on the confidence parameter, the target to be tracked is selected from the detected targets in the current frame image.
3. The method for re-tracking people based on image detection technology according to claim 2, characterized in that, The step of filtering out the tracking target from the detected targets in the current frame image based on the confidence parameter includes: In the current frame image, the detected targets whose confidence parameters are greater than or equal to the preset confidence threshold are taken as the tracking targets.
4. The method for re-tracking people based on image detection technology according to claim 1, characterized in that, The obtained skeleton points of the tracked target include: The tracking target in the current frame image is cropped to obtain a cropped region containing the tracking target; The cropped area is input into a preset human skeleton point detection model, and the skeleton points of the tracking target are output.
5. The method for re-tracking people based on image detection technology according to claim 1, characterized in that, The determination of whether the tracking target is a human target includes: If the number of skeletal points of the tracked target in the current frame image is equal to the number of elements contained in the complete set of skeletal points, then the tracked target in the current frame image is a human target; otherwise, it is not a human target.
6. The method for re-tracking people based on image detection technology according to claim 1, characterized in that, The step of generating a distance relationship table for the current frame image, which includes the distances between skeletal points of the human target, includes: Select multiple preset key bone points from all the bone points of the human target in the current frame image; The three-dimensional spatial coordinates of the key skeletal points of the human target in the current frame image are acquired using a depth camera; Take any two key skeleton points as a group, and use the Euclidean distance between the three-dimensional spatial coordinates of the two key skeleton points in each key skeleton point group as the distance coefficient of each key skeleton point group. Record each key skeleton point group and the distance coefficient of each key skeleton point group in the distance relationship table.
7. The method for re-tracking people based on image detection technology according to claim 1, characterized in that, The determination of whether the human target is lost includes: If the human target is not detected in a consecutive preset number of frames, it is determined that the human target is lost; otherwise, it is determined that the human target is not lost.
8. A method for re-tracking people based on image detection technology according to claim 6, characterized in that, The tracking of human targets in new surveillance images includes: Based on the difference in distance coefficients between the new distance relationship table and the historical distance relationship table for the same key skeleton point group, the feature similarity of human targets in the new surveillance image is obtained; Based on the feature similarity, it is determined whether the human target in the new monitoring image is the same human target that was lost before. If it is the same human target that was lost before, the human target in the new monitoring image is continuously tracked. Otherwise, the human target in the new monitoring image is tracked as a new target.
9. A method for re-tracking people based on image detection technology according to claim 8, characterized in that, The feature similarity of human targets in the new surveillance image obtained includes: The absolute value of the difference between the distance coefficients of the same key skeleton point group between the new distance relationship table and the historical distance relationship table is negatively correlated and normalized to obtain the similarity coefficient of the same key skeleton point group between the new distance relationship table and the historical distance relationship table. The average of the similarity coefficients of all identical key skeleton point groups between the new distance relationship table and the historical distance relationship table is used as the feature similarity of the human target in the new surveillance image.
10. A method for re-tracking people based on image detection technology according to claim 8, characterized in that, The step of determining whether a human target in a new surveillance image is the same human target that was lost before, based on the feature similarity, includes: If the feature similarity is greater than or equal to a preset similarity threshold, then the human target in the new surveillance image is the human target that was lost before; otherwise, the human target in the new surveillance image is not the human target that was lost before.