Target tracking method and device based on machine vision, electronic equipment and storage medium

Abstract

The invention discloses a target tracking method and device based on machine vision, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence or other related fields, and the method comprises the steps: obtaining a target video sequence containing at least two video frames, performing feature collection in the current video frame according to a predetermined to-be-tracked target to obtain a feature point set; performing matching analysis according to the feature point set and a previous feature point set corresponding to a previous video frame to obtain a feature point pair set; calculating a scale change value and a displacement value corresponding to the target according to the feature point pair set, and obtaining position information of a target frame in the previous video frame; and updating the position of the target frame in the current video frame according to the scale change value, the displacement value and the position information of the target frame in the previous video frame to obtain a target tracking result. Through the target tracking method and device, the technical problem that the tracking precision and continuity of a target tracking algorithm in a complex changing environment are reduced in the related technology is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology or other related fields. Specifically, it relates to a target tracking method and apparatus, electronic device, and storage medium based on machine vision. Background Technology

[0002] In the field of bank security monitoring, traditional target tracking technologies face numerous challenges, especially in complex environments such as changes in lighting, target scale, occlusion, and deformation, where tracking accuracy and robustness are often insufficient. Existing background modeling-based target tracking algorithms, while simple and direct, often suffer from limited tracking performance in complex scenes, such as densely populated and dynamically changing environments like bank lobbies, particularly when the target is occluded by other objects or people, easily leading to target loss. Keypoint matching-based target tracking algorithms demonstrate good performance when the target's appearance changes significantly; however, once the target is completely or partially occluded, the algorithm cannot effectively utilize the keypoint information of the occluded portion, thus reducing tracking accuracy. Image segmentation-based target tracking algorithms can theoretically accurately separate the target from the background, but in practice, this method, due to its high computational cost and algorithmic complexity, may not be efficient enough for real-time monitoring requirements.

[0003] In recent years, with the development of deep learning technology, deep learning-based target tracking algorithms have gradually become mainstream. These algorithms, through deep neural networks, can learn more complex feature representations, thereby improving tracking accuracy and robustness to some extent. However, these algorithms may still suffer from excessive computational resource consumption when processing large-scale data, especially in large public environments like banks, where the processing of large amounts of real-time video data places higher demands on computing power and power consumption control. Furthermore, deep learning algorithms are highly dependent on training data. When the target object's appearance in the video sequence differs significantly from the training data—for example, when the target size suddenly changes, the lighting suddenly changes, or the target is partially occluded—the tracking performance will be affected, and tracking drift may even occur.

[0004] In summary, current machine vision-based target tracking algorithms suffer from several technical challenges in practical applications of bank security monitoring, particularly when handling tracking tasks in complex environments such as varying lighting conditions, target size, occlusion, and deformation. These challenges include low tracking accuracy, insufficient robustness, low computational efficiency, and limited ability to process large-scale real-time data. This not only affects the effectiveness of bank security monitoring but may also lead to a decline in customer service quality and efficiency.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] The main objective of this application is to provide a machine vision-based target tracking method, apparatus, electronic device, and storage medium, to at least solve the technical problem of decreased tracking accuracy and continuity of target tracking algorithms in complex and changing environments.

[0007] To achieve the above objectives, according to one aspect of this application, a machine vision-based target tracking method is provided. The method includes: acquiring a target video sequence comprising at least two video frames; performing feature acquisition on the current video frame based on a predetermined target to be tracked, obtaining a feature point set comprising N target feature points, where N is a positive integer; performing matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame, obtaining a feature point pair set comprising M feature point pairs, where M is a positive integer less than N; calculating the scale change value and displacement value corresponding to the target based on the feature point pair set, and obtaining the position information of the target bounding box in the previous video frame, wherein the target bounding box is used to identify the target position in the video frame; updating the position of the target bounding box in the current video frame based on the scale change value, the displacement value, and the position information of the target bounding box in the previous video frame, thereby obtaining a target tracking result.

[0008] Further, the step of calculating the scale change value and displacement value corresponding to the target based on the set of feature point pairs includes: calculating the distance ratio between the two corresponding feature points for each feature point pair to obtain a preliminary scale change value; calculating the median of the M preliminary scale change values ​​corresponding to the M feature point pairs to obtain the scale change value; calculating the lateral distance difference and longitudinal distance difference between the two corresponding feature points for each feature point pair, and calculating a preliminary displacement value based on the lateral distance difference and the longitudinal distance difference; and calculating the median of the M preliminary displacement values ​​corresponding to the M feature point pairs to obtain the displacement value.

[0009] Further, the step of updating the position of the target box in the current video frame based on the scale change value, the displacement value, and the position information of the target box in the previous video frame to obtain the target tracking result includes: determining the size of the previous target box and the coordinates of the center point of the previous target box using the position information of the target box in the previous video frame; adjusting the size of the previous target box according to the scale change value to obtain the updated target box size; determining the coordinates of the center point of the updated target box based on the displacement value and the coordinates of the center point of the previous target box; and updating the position of the target box in the current video frame based on the updated target box size and the coordinates of the center point of the updated target box to obtain the target tracking result.

[0010] Furthermore, before feature acquisition in the current video frame, the target tracking method includes: determining whether the target is occluded, obtaining a determination result, and if the determination result indicates that the target is not occluded, performing feature acquisition in the current video frame; or if the determination result indicates that the target is occluded, analyzing the target bounding box position changes of the target in X consecutive historical video frames using a motion prediction algorithm to obtain the predicted position information of the target in the current frame, where X is a specified positive integer; updating the position of the target bounding box in the current video frame according to the predicted position information to obtain the target tracking result.

[0011] Furthermore, after obtaining a set of feature point pairs containing M feature point pairs, the target tracking method further includes: for each feature point pair, calculating the feature similarity of the feature point pair, and determining the feature similarity as the confidence level of the feature point pair; filtering out low-quality feature point pairs corresponding to the confidence level below a preset confidence threshold, and removing all low-quality feature point pairs from the set of feature point pairs.

[0012] Furthermore, after removing all the low-configuration feature point pairs from the feature point pair set, the target tracking method further includes: if the number of elements in the feature point pair set obtained after removal is less than a preset threshold, performing target detection on the current video frame using a global search strategy and identifying a first type of target detection box; or, if the number of elements in the feature point pair set obtained after removal is less than a preset threshold, performing target detection on the current video frame using a local search strategy and identifying a second type of target detection box.

[0013] Furthermore, after updating the position of the target box in the current video frame based on the scale change value, the displacement value, and the position information of the target box in the previous video frame, the target tracking method further includes: if the position of the target box is successfully updated in the current video frame, performing clustering processing on the updated target box and the target detection box, and outputting the resulting comprehensive target box as the target tracking result, wherein the target detection box is either the first type of target detection box or the second type of target detection box.

[0014] To achieve the above objectives, according to another aspect of this application, a target tracking device based on machine vision is also provided. The device includes: an acquisition unit, configured to acquire a target video sequence containing at least two video frames, and to perform feature acquisition in the current video frame according to a predetermined target to be tracked, obtaining a feature point set containing N target feature points, where N is a positive integer; an analysis unit, configured to perform matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame, obtaining a feature point pair set containing M feature point pairs, where M is a positive integer less than N; an acquisition unit, configured to calculate the scale change value and displacement value corresponding to the target based on the feature point pair set, and acquire the position information of the target bounding box in the previous video frame, wherein the target bounding box is used to identify the target position in the video frame; and an update unit, configured to update the position of the target bounding box in the current video frame according to the scale change value, the displacement value, and the position information of the target bounding box in the previous video frame, to obtain a target tracking result.

[0015] Further, the acquisition unit includes: a first calculation module, used to calculate the distance ratio between corresponding two feature points for each feature point pair to obtain a preliminary value of scale change; a second calculation module, used to calculate the median of the M preliminary values ​​of scale change corresponding to the M feature point pairs to obtain the scale change value; a third calculation module, used to calculate the lateral distance difference and longitudinal distance difference between corresponding two feature points for each feature point pair, and calculate a preliminary value of displacement based on the lateral distance difference and the longitudinal distance difference; and a fourth calculation module, used to calculate the median of the M preliminary values ​​of displacement corresponding to the M feature point pairs to obtain the displacement value.

[0016] Further, the updating unit includes: a first determining module, configured to determine the size of the previous target box and the coordinates of the center point of the previous target box based on the position information of the target box in the previous video frame; an adjusting module, configured to adjust the size of the previous target box according to the scale change value to obtain the updated target box size; a second determining module, configured to determine the coordinates of the center point of the updated target box according to the displacement value and the coordinates of the center point of the previous target box; and a first updating module, configured to update the position of the target box in the current video frame according to the updated target box size and the coordinates of the center point of the updated target box to obtain the target tracking result.

[0017] Furthermore, the target tracking device includes: a judgment module, configured to determine whether the target is occluded before feature acquisition in the current video frame, obtain a judgment result, and perform feature acquisition in the current video frame if the judgment result indicates that the target is not occluded; an analysis module, configured to analyze the target bounding box position changes of the target in X consecutive historical video frames using a motion prediction algorithm if the judgment result indicates that the target is occluded, to obtain the predicted position information of the target in the current frame, where X is a specified positive integer; and a second update module, configured to update the position of the target bounding box in the current video frame according to the predicted position information, to obtain the target tracking result.

[0018] Furthermore, the target tracking device further includes: a fifth calculation module, configured to, after obtaining a set of feature point pairs containing M feature point pairs, calculate the feature similarity of each feature point pair and determine the feature similarity as the confidence level of the feature point pair; and a filtering module, configured to filter out low-configuration feature point pairs corresponding to the confidence level below a preset confidence threshold and remove all low-configuration feature point pairs from the set of feature point pairs.

[0019] Furthermore, the target tracking device further includes: a first detection module, configured to perform target detection on the current video frame using a global search strategy and identify a first type of target detection box when the number of elements in the feature point pair set obtained after removing all the low-configuration feature point pairs from the feature point pair set is less than a preset threshold; and a second detection module, configured to perform target detection on the current video frame using a local search strategy and identify a second type of target detection box when the number of elements in the feature point pair set obtained after removing the low-configuration feature point pairs is less than a preset threshold.

[0020] Furthermore, the target tracking device further includes: a clustering module, used to perform clustering processing on the updated target box and the target detection box after updating the position of the target box in the current video frame according to the scale change value, the displacement value and the position information of the target box in the previous video frame, and when the position update of the target box in the current video frame is successful, and output the resulting comprehensive target box as the target tracking result, wherein the target detection box is the first type of target detection box or the second type of target detection box.

[0021] To achieve the above objectives, according to another aspect of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the machine vision-based target tracking method described in any one of the above claims.

[0022] To achieve the above objectives, according to another aspect of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the machine vision-based target tracking method described in any of the above claims.

[0023] To achieve the above objectives, according to another aspect of this application, a computer program product is also provided, including computer instructions, wherein when the computer instructions are executed by a processor, they implement the steps of the machine vision-based target tracking method described in any one of the preceding claims.

[0024] This invention proposes a target tracking method based on machine vision. First, a target video sequence containing at least two video frames is acquired. Then, features are collected in the current video frame based on a pre-determined target to be tracked, resulting in a feature point set containing N target feature points (where N is a positive integer). Next, a matching analysis is performed between the feature point set and the previous feature point set corresponding to the previous video frame, resulting in a feature point pair set containing M feature point pairs (where M is a positive integer less than N). Then, the scale change value and displacement value corresponding to the target are calculated based on the feature point pair set, and the position information of the target bounding box in the previous video frame is obtained. The target bounding box is used to identify the target position in the video frame. Finally, based on the scale change value, displacement value, and the position information of the target bounding box in the previous video frame, the position of the target bounding box in the current video frame is updated to obtain the target tracking result.

[0025] This invention employs a multi-scale adaptive illumination processing approach, combining SURF feature matching and HOG feature point confidence evaluation to achieve accurate and efficient target tracking under complex illumination conditions and varying target scales. This enables intelligent monitoring and customer service optimization in banking scenarios. Specifically, addressing the insufficient robustness of traditional TLD algorithms to changes in illumination and target scale, this invention introduces a multi-scale adaptive illumination adjustment mechanism. This ensures stable operation under different illumination environments while maintaining high sensitivity to changes in target scale, significantly improving the algorithm's adaptability and accuracy.

[0026] Furthermore, to address the challenge of occlusion, this invention incorporates occlusion detection and motion prediction steps. A fusion algorithm combining Kalman filters and feature point matching is used to predict the target's position in the next frame, effectively overcoming tracking interruptions caused by local occlusion and significantly enhancing the algorithm's tracking continuity and reliability in complex scenes. Building upon this, feature point pair confidence assessment further eliminates low-quality point pairs during the matching process, ensuring the accuracy of the tracking results. Through these technical steps, this invention not only significantly improves the accuracy and stability of target tracking but also specifically addresses the needs of the banking environment, enabling real-time monitoring of customer behavior, automatic identification of high-net-worth clients, and precise allocation of internal resources. This solves the shortcomings of traditional monitoring methods in terms of efficiency, accuracy, and intelligent response, and further addresses the technical problem of decreased tracking accuracy and continuity in related technologies under complex and changing environments. Attached Figure Description

[0027] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0028] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a machine vision-based target tracking method is shown.

[0029] Figure 2 This is a flowchart of an optional machine vision-based target tracking method according to an embodiment of the present invention;

[0030] Figure 3 This is an optional tracking flowchart based on SURF feature matching according to an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of an optional machine vision-based target tracking device according to an embodiment of the present invention;

[0032] Figure 5 This is a structural block diagram of an electronic device that performs a machine vision-based target tracking method according to an embodiment of the present invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] To facilitate understanding of the present invention by those skilled in the art, some terms or nouns involved in the various embodiments of the present invention are explained below:

[0036] SURF, Speeded-Up Robust Features, is a feature point detection algorithm for image processing. It is widely used in image matching and object recognition due to its speed and robustness to image changes.

[0037] HOG, Histogram of Oriented Gradients, is an image feature descriptor used for object recognition. It calculates the orientation histogram of pixel intensity gradients in an image, has a strong ability to describe object contours, and exhibits good invariance to changes in the scale and pose of the target.

[0038] The Kalman filter is a recursive Bayesian filter primarily used for state estimation of linear systems. It is particularly effective in predicting and correcting the state when the system state is affected by noise, thereby improving the continuity of tracking and the accuracy of prediction.

[0039] GMS, Geometric Matching Score, is a quality control method for image feature point pairing. It can evaluate the correctness of the pairing based on the geometric relationship between the point pairs, thereby reducing false matches and increasing the reliability of tracking.

[0040] Local search strategy and global search strategy: Local search strategy refers to target detection in the vicinity of the target box in the previous frame, which is suitable for situations where the target position changes little; Global search strategy refers to target detection across the entire video frame, which is suitable for situations where the target may move significantly or temporarily lose tracking.

[0041] Occlusion detection, in target tracking, is the ability to detect whether a target is partially or completely occluded by other objects or environmental elements.

[0042] Motion prediction algorithms are algorithms that predict the possible location of a target in the next frame based on the target bounding box position information of the current frame and one or more historical frames. They are often used in scenarios where the target is occluded so that tracking can be quickly resumed after the occlusion is removed.

[0043] It should be noted that the machine vision-based target tracking method and apparatus in this application can be used in the field of artificial intelligence technology for real-time tracking and behavior analysis of people or objects in bank surveillance videos, and can also be used in any field other than artificial intelligence technology for real-time tracking and behavior analysis of people or objects in bank surveillance videos. This application does not limit the application field of the machine vision-based target tracking method and apparatus.

[0044] It should be noted that all relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) involved in this application are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, processing, transmission, provision, disclosure, use, and handling of such data comply with the laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse access. For example, this system has interfaces with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface, and the relevant information is obtained only after receiving consent from the aforementioned user or organization.

[0045] The information collection (e.g., user voice, video, and text collection) and analysis operations involved in this application have provided users with corresponding operation entry points during execution, allowing users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0046] The following embodiments of the present invention can be applied to various systems / applications / devices that require target detection and tracking, enabling intelligent target tracking solutions based on machine vision. The present invention uses the SURF feature point detection algorithm to acquire and match target features, and then combines it with HOG feature point confidence evaluation and occlusion detection and motion prediction algorithms. This allows for better capture of target changes and motion trajectories in video sequences, especially under complex conditions such as changes in illumination, scale, and occlusion, maintaining the continuity and accuracy of target tracking.

[0047] This invention also dynamically monitors the target's behavior patterns through a real-time analysis and feedback mechanism, and performs interrupted staking prediction processing on the target. That is, when encountering interference such as occlusion, the target position is estimated in advance through motion prediction to ensure the smoothness and efficiency of the tracking process. The target is located quickly and accurately, which greatly improves the application effect in scenarios such as bank monitoring and crowd management.

[0048] The present invention will now be described in detail with reference to various embodiments.

[0049] Example 1

[0050] According to an embodiment of the present invention, an embodiment of a target tracking method based on machine vision is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0051] The target tracking method based on machine vision provided in Embodiment 1 of the present invention can be executed on a mobile terminal, computer terminal or similar computing device. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing a machine vision-based target tracking method is shown. Figure 1 As shown, the computer terminal 10 (or mobile device) may include one or more processors 102 (shown as 102a, 102b, ..., 102n in the figure) 102 (processor 102 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0052] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0053] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the machine vision-based target tracking method in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned machine vision-based target tracking method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0054] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0055] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0056] Under the above operating environment, the present invention provides, as follows: Figure 2The machine vision-based target tracking method shown is implemented in an intelligent bank monitoring system. Combining machine vision and deep learning technologies, it is used for security monitoring and customer behavior analysis in bank branches, ATMs, and vaults. In particular, it addresses the problem of continuous and accurate tracking under conditions of target occlusion, lighting changes, and scale changes. By improving the tracking module, it introduces SURF feature point matching, HOG feature point confidence evaluation, and a fusion algorithm of occlusion detection and Kalman filter. Specifically, it includes an initialization phase, SURF feature matching, GMS denoising, HOG feature similarity calculation, occlusion judgment and motion prediction, detection module invocation, integrated module application, and algorithm iteration. The goal is to achieve intelligent, continuous, and accurate tracking of specific targets in complex environments, thereby improving the bank's security management level and service quality.

[0057] The embodiments of the present invention will now be described in detail with reference to each specific step.

[0058] Figure 2 This is a flowchart of an optional machine vision-based target tracking method according to an embodiment of the present invention, such as... Figure 2 As shown, the method includes the following steps:

[0059] Step S201: Obtain a target video sequence containing at least two video frames, and perform feature acquisition in the current video frame according to the pre-determined target to be tracked to obtain a feature point set containing N target feature points, where N is a positive integer.

[0060] It should be noted that the initialization phase requires target feature extraction and representation construction to provide a stable, distinguishable, and robust target appearance model. This does not rely solely on a single feature but rather achieves efficient target representation through multi-scale, multi-discriminative feature joint acquisition. The target region (i.e., the bounding box) can be delineated in the first frame (or a specified starting frame) of the video sequence. Subsequently, SURF (Speed-Up Robust Feature) feature point sampling is performed on the target region image patch to extract several corner points or blob features with high gradient changes and stable structures, forming an initial feature point set P={p1,p2,…,p...}. N These feature points are scale-invariant and rotation-invariant, which can effectively cope with scale scaling, slight rotation or viewpoint changes that may occur in subsequent frames.

[0061] To enhance the discriminative power of features, embodiments of the present invention can further utilize the target region to generate positive samples (through radiative transformations such as translation, rotation, and scaling), and combine them with the background region to generate negative samples, which are used to train the classifier (such as a variance classifier or a random circle classifier) ​​of the subsequent detection module. This process not only completes feature acquisition but also simultaneously constructs the appearance model of the target and its background discriminative power.

[0062] In a specific implementation scenario, customer A, wearing a dark coat and glasses, is the target to be tracked in front of a bank ATM, and has a clear facial outline. After defining the target bounding box in the first frame, the system can extract approximately 80–120 SURF feature points within that area, covering textured areas such as the shoulders, collar, and glasses frames. These feature points will serve as matching anchor points in subsequent frames. Even if the customer's head is slightly turned or the coat wrinkles due to movement, the feature points can still maintain a stable match, avoiding tracking drift caused by local deformation.

[0063] Optionally, in the machine vision-based target tracking method provided in this embodiment of the invention, before feature acquisition in the current video frame, the target tracking method includes: determining whether the target is occluded, obtaining a determination result, and if the determination result indicates that the target is not occluded, performing feature acquisition in the current video frame; or if the determination result indicates that the target is occluded, analyzing the target bounding box position changes of the target in X consecutive historical video frames using a motion prediction algorithm to obtain the predicted position information of the target in the current frame, where X is a specified positive integer; updating the position of the target bounding box in the current video frame according to the predicted position information to obtain the target tracking result.

[0064] The above steps construct a perception-decision-response closed-loop mechanism, namely: before performing feature acquisition in each frame, it is first determined whether the target is occluded, and the tracking strategy is dynamically switched according to the determination result—high-precision feature matching is used when there is no occlusion, and conservative tracking driven by motion prediction is switched when there is occlusion, thereby maintaining tracking continuity while ensuring accuracy. The complete implementation logic of the optional steps is as follows:

[0065] First, the occlusion determination is based on the feature point matching success rate between the current frame and the previous frame (i.e., the ratio of the number of feature point pairs retained after GMS filtering to the initial number of collected points). If the number of matching point pairs drops sharply by more than a threshold (e.g., below 30%), or the feature point distribution shows "hollowing out" (e.g., feature points in the target area are concentrated at local edges), then it is determined to be occlusion. For example, when a customer is withdrawing money at an ATM, their body is briefly obscured by the teller. At this time, the number of SURF feature point matches drops sharply from 85 to 12, and the remaining points are concentrated in the head area. The system determines this to be "partial occlusion".

[0066] Furthermore, feature acquisition is performed when there is no occlusion. Following the original process (S201–S204), SURF sampling, matching, scale displacement calculation, and target bounding box update are performed to ensure maximum tracking accuracy. When there is occlusion: Motion prediction mode is used: a Kalman filter is invoked, and the state vector can be defined as: ,in The coordinates of the center of the target box Width and height, This corresponds to the velocity component. The position and size changes of historical target boxes from the previous X frames (e.g., X=3) can be used to estimate the "optimal state" of the current frame using a Kalman prediction model, outputting the predicted box. No feature acquisition is performed; the predicted box is directly used as the tracking result for the current frame, avoiding noise or mismatches introduced by sampling occluded areas. Simultaneously, the duration of occlusion is recorded. If it exceeds a set threshold (e.g., 5 frames), the detection module is triggered to perform a global search (see step 5 in the manual), preventing long-term data loss.

[0067] In a specific implementation scenario, during bank vault entrance monitoring, a security personnel using a handheld metal detector briefly obscures the view of the customer being tracked as they pass through the camera's field of view. The system detects a feature point matching success rate below 25% in frame 87 and immediately enters prediction mode. Based on the motion trajectory of the previous three frames (frames 84–86) (moving steadily to the right), it predicts that the target in frame 88 will be located 15 pixels to the right and in front, with unchanged width and height. The system skips feature extraction and directly outputs the predicted bounding box, avoiding misalignment of the bounding box caused by incorrectly extracting background textures in the obscured area. When the customer is fully revealed in frame 90, the system resumes feature matching, seamlessly continuing tracking without frame skipping or loss.

[0068] Step S202: Perform matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame to obtain a feature point pair set containing M feature point pairs, where M is a positive integer less than N.

[0069] It should be noted that the technical goal of cross-frame feature point matching and effective alignment is to establish a spatiotemporal correlation of target appearance between consecutive frames of a video sequence, eliminate mismatches, retain high-confidence correspondences, and provide a reliable basis for subsequent target displacement and scale estimation. Specifically, using the feature point set of the current frame (time t)... The set of feature points in the previous frame (time t-1) Based on this, initial feature point pairing is performed to generate a set of candidate feature point pairs. K is the initial number of feature point pairs, and K is greater than or equal to M. Due to factors such as background interference, lighting fluctuations, and non-rigid deformation in the video, this initial set contains a large number of mismatches (such as mismatches of background points and misassociations of occluded areas).

[0070] To improve the matching quality, embodiments of the present invention may further introduce the GMS (Geometric Matching Score) method. This method analyzes the geometric consistency between feature point pairs (such as the relative position relationship in the spatial neighborhood) and excludes abnormal matching points that do not conform to the rigid body motion assumption. For example, if the relative position of a certain feature point pair with its neighboring points in the current frame deviates from its relative structure in the previous frame (such as sudden changes in angle or distance), it is determined as a false match and excluded, thereby reducing the number of feature point pairs from the initial thousands to an effective M (M < N), significantly improving the matching accuracy.

[0071] In an optional specific implementation scenario, in the bank lobby monitoring, several SURF feature points on the collar of customer B shift in position in two consecutive frames during walking. The initial matching may wrongly correspond this point to the moving light texture points in the background. After GMS analysis, it is found that this "matching" does not conform to the geometric consistency of the surrounding points (such as all other points around the collar move forward steadily, only this point jumps), so it is excluded. The final retained point pairs need to satisfy the local structural invariance and can accurately reflect the overall motion trend of the target.

[0072] Optionally, in the target tracking method based on machine vision provided by embodiments of the present invention, after obtaining the set of feature point pairs containing M feature point pairs, the target tracking method further includes: for each feature point pair, calculating the feature similarity of the feature point pair and determining the feature similarity as the confidence of the feature point pair; screening out the low-confidence feature point pairs corresponding to the confidences lower than the preset confidence threshold and excluding all the low-confidence feature point pairs from the set of feature point pairs.

[0073] In this embodiment, calculating the confidence based on the feature similarity and performing threshold screening can improve the reliability of feature matching. Its essence is to quantify the appearance consistency of visual features into a decision-making confidence index, realize the active filtering of false matches, and thus provide a clean and reliable input for subsequent target box estimation. The specific implementation process is as follows:

[0074] First, calculate the feature similarity. For each feature point pair screened by GMS , respectively extract the corresponding 10×10 local image patches (centered on the feature point) in the previous frame and the current frame. Calculate the Histogram of Oriented Gradients (HOG) feature vectors for these two image patches, that is: divide the image patch into 4×4 cell units, each unit statistically calculates the gradient magnitudes in 8 directions, forms a 62-dimensional local descriptor, and then performs L2 normalization.

[0075] Then, define the confidence. Take the cosine similarity between the two HOG descriptors as the confidence of this feature point pair: , where These are the HOG descriptors for the point pair in the preceding and following frames, respectively. Next, a threshold filtering is performed. The confidence threshold can be set to 0.8, retaining only feature point pairs with a confidence greater than 0.8, while the rest are considered "low-confidence matches" and removed from the set.

[0076] In a specific implementation scenario, during bank lobby surveillance, a customer is moving. Due to changes in lighting, a reflective area appears on the surface of their coat, causing some SURF feature points to fall onto these reflective spots. Although these points spatially match the positions of the previous frame (which GMS failed to remove), their local texture structure has changed drastically due to specular reflection, and the HOG gradient direction distribution differs significantly from the previous frame, with a similarity of only 0.3. The system removes this matching pair to prevent it from participating in displacement and scale calculations. The final retained point pairs all come from structurally stable areas such as the coat texture, collar buttons, and facial contours, ensuring the target bounding box moves forward stably rather than being "dragged" in the wrong direction by reflective points.

[0077] The above mechanism addresses the fundamental flaw in traditional feature matching: geometric consistency does not equal appearance consistency. While GMS can eliminate geometric outliers, it cannot identify appearance distortions caused by lighting, material reflection, or local deformation. HOG confidence assessment provides a supplementary correction to this flaw at the appearance level, ensuring that the matching results are not only geographically accurate but also aesthetically credible.

[0078] Figure 3 This is an optional SURF feature matching-based tracking flowchart according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes: performing feature point matching on the image to be tested, specifically including: SURF feature point detection, BF matching and GMS to eliminate mismatches, then filtering out irrelevant feature point pairs from the feature point set obtained from the initial matching, and then evaluating the confidence of each of the remaining feature point pairs to finally obtain the tracking result.

[0079] Optionally, in the machine vision-based target tracking method provided in this embodiment of the invention, after removing all low-configuration feature point pairs from the feature point pair set, the target tracking method further includes: if the number of elements in the feature point pair set obtained after removal is less than a preset threshold, performing target detection on the current video frame using a global search strategy and identifying a first type of target detection box; or, if the number of elements in the feature point pair set obtained after removal is less than a preset threshold, performing target detection on the current video frame using a local search strategy and identifying a second type of target detection box.

[0080] In this embodiment, the dynamic switching detection search strategy mechanism based on the number of remaining feature points is an important design at the level of tracking failure emergency response. Essentially, when the tracking module becomes unreliable due to severe loss of feature points, the system no longer forcibly relies on low-quality matching, but instead adaptively switches the detection mode to prioritize the recovery of tracking, thus achieving an intelligent transition from tracking-driven to detection-backup.

[0081] Specifically, after completing GMS filtering and HOG confidence screening, if the number of remaining valid feature point pairs is lower than a preset threshold (e.g., less than 10, or less than 15% of the initial number of feature points), it is determined that the tracking module is seriously malfunctioning and cannot reliably estimate the target motion.

[0082] If a global search strategy is adopted, the detection window slides across the entire image, and a variance classifier, a random circle classifier, and a nearest neighbor classifier are used to perform triple discrimination on all candidate regions, outputting the first type of target detection box (i.e., the full-image re-detection result). This strategy is suitable for scenarios where the target may have completely left the original area, been occluded for a long time, or undergone drastic displacement. It has a large computational cost but a wide coverage.

[0083] If a local search strategy is adopted, the detection window is scanned only in the extended region of the target bounding box in the previous frame (such as a rectangular area centered on the original bounding box and expanded by twice its width and height). The classifier is only performed within this local area, and the second type of target detection box is output (i.e., the local re-detection result). This strategy is suitable for scenarios where the target is still located near the original region and the feature is lost only due to brief occlusion or changes in lighting. It has high computational efficiency and fast response.

[0084] In one possible implementation scenario, at a bank counter, a customer briefly looks down at their phone, obscuring their face and upper body, causing the number of SURF feature point matches to plummet from 80 to 6. The system determines the following: If the customer's movement trajectory was previously stable (no sudden changes in speed or direction) and the occlusion time is less than 3 frames, the system uses a local search, expanding the detection range by 50 pixels above, below, left, and right of the original target bounding box, quickly finding the target near its original position, outputting a second-type target detection box, and resuming tracking. If the customer was previously moving rapidly towards the door, and only 4 feature points remain in the current frame, with the position deviating from the predicted direction of the previous frame by more than 30 pixels, the system determines that the target has left the original area, switches to a global search, rescans the entire frame, and finally finds the target in the doorway area, outputting a first-type target detection box, completing the target rebinding. Without this mechanism, the system might mistakenly judge tracking failure due to "insufficient feature points," yet stubbornly use the incomplete feature points to calculate displacement, causing the target bounding box to continuously drift; or it might completely abandon tracking, requiring manual intervention to restart.

[0085] Step S203: Calculate the scale change value and displacement value corresponding to the target based on the set of feature points, and obtain the position information of the target box in the previous video frame. The target box is used to identify the target position in the video frame.

[0086] It should be noted that deriving the rigid motion parameters (displacement and scale changes) of the target in space based on high-confidence feature point pairs, and combining this with the target box state of the previous frame to achieve continuous estimation of the target position, is a key step in transforming "feature matching" into "target localization". The goal is to extract global motion information from a small number of stable feature point pairs after removing mismatches, so as to achieve robust estimation of the overall position and size changes of the target and avoid target box jitter or deviation caused by local feature drift.

[0087] The feature point set filtered by GMS For example, scale variation can be calculated by comparing the Euclidean distances of each pair of feature points in consecutive frames, i.e.: By using the median instead of the mean, interference from outliers (such as mismatches or local deformations) can be effectively suppressed, ensuring stable scale estimation. For example, if a target moves closer to the camera, the spacing between its facial feature points increases, and this ratio will be greater than 1, allowing the system to enlarge the target bounding box accordingly.

[0088] Furthermore, displacement This can be achieved by differentiating the horizontal and vertical displacements of each pair of feature points and then taking the median, i.e.: ; The median mechanism also ensures the robustness of displacement estimation to local noise and occluded areas. For example, when a customer walks in front of an ATM, their foot feature points may be occluded, but their upper body feature points still move forward steadily. The system only relies on the median displacement of the unoccluded points to avoid the entire frame being "dragged" to the wrong position.

[0089] Furthermore, combining the target bounding box from the previous frame... Calculate the target bounding box of the current frame using the following formula. Complete the pose update of the rigid body model based on feature motion:

[0090]

[0091]

[0092]

[0093]

[0094] In one possible implementation scenario, a customer walks from a bank counter to an ATM, with their bounding box gradually increasing in size (closer to the camera) and moving forward in consecutive frames. The geometric relationship of SURF feature point pairs shows that the average distance between points increases by a factor of 1.2. The median horizontal displacement is +15 pixels, and the median vertical displacement is +3 pixels. Based on this, the system enlarges the original target bounding box by 20% and offsets it by 15×3 pixels to accurately track the target. Even if the target's sleeve deforms locally due to movement, it will not affect the overall stability of the bounding box. This step solves the problem of target bounding box inaccuracy caused by "single-point drift" in traditional tracking algorithms. Through a statistical robustness mechanism (median), the tracking results have the ability to resist local interference, noise, and partial occlusion, which is the core engineering means to achieve stable tracking.

[0095] Optionally, in the target tracking method based on machine vision provided in this embodiment of the invention, the step of calculating the scale change value and displacement value corresponding to the target based on the set of feature point pairs includes: calculating the distance ratio between the two corresponding feature points for each feature point pair to obtain a preliminary scale change value; calculating the median of the M preliminary scale change values ​​corresponding to the M feature point pairs to obtain a scale change value; calculating the lateral distance difference and longitudinal distance difference between the two corresponding feature points for each feature point pair, and calculating a preliminary displacement value based on the lateral distance difference and longitudinal distance difference; and calculating the median of the M preliminary displacement values ​​corresponding to the M feature point pairs to obtain a displacement value.

[0096] In this embodiment, the core idea of ​​the precise and robust implementation of the target motion parameters (scale change and displacement) calculation process is: instead of relying on single points or average values, the interference of abnormal matching point pairs is eliminated through the median mechanism of group statistics, so as to achieve a stable estimate of the overall rigid motion of the target.

[0097] Specifically, for each pair of feature points in the set, the Euclidean distance between them in the previous frame and the current frame is calculated, and the distance ratio is used as the preliminary value of the scale change for that pair. The preliminary values ​​of scale change are calculated for all valid feature point pairs, and the median is taken as the final scale change value. For each feature point pair, the displacement components in the horizontal and vertical directions are calculated. Similarly, for M point pairs, the displacement component sets are obtained, and the median of each set is taken as the final displacement value.

[0098] The role of the median mechanism includes: strong resistance to outliers. If a feature point falls in a local deformation area of ​​the target (such as a sleeve fold), a reflective area, or an occluded edge, its distance ratio or displacement may deviate significantly from the true value, but the median is not affected by extreme values ​​and can still reflect the consensus trend of most point pairs; no need to assume a distribution pattern. Unlike the mean, the median does not depend on the assumption of a normal distribution and is suitable for real-world scenarios where feature points are sparsely distributed and noise is non-Gaussian.

[0099] In one optional implementation scenario, at a bank counter, a customer leans forward while withdrawing money. Their upper body feature points (such as collar and glasses) remain stable, but their arm deforms locally due to the movement, causing an abnormal distance ratio (e.g., abruptly changing from 1.0 to 2.5) for three feature point pairs due to arm folding. If the mean is used, the scale change might be amplified to 1.3, resulting in the target bounding box being incorrectly magnified by 30%. However, using the median, out of 100 point pairs, 97 point pairs have distance ratios distributed between 0.95 and 1.05, while the median remains at 1.01. The target bounding box scale remains almost unchanged, translating only based on the remaining 97 stable point pairs. The target bounding box accurately follows the customer's head movement without expanding due to arm deformation. This mechanism ensures that the update of the target bounding box is driven entirely by the majority of consistent motion patterns, rather than being skewed by a few noisy points, which is the key mathematical guarantee for achieving stable tracking in this invention.

[0100] Step S204: Based on the scale change value, displacement value, and the position information of the target box in the previous video frame, update the position of the target box in the current video frame to obtain the target tracking result.

[0101] This embodiment integrates the motion parameters (scale changes and displacements) calculated in the previous steps with the historical target state to achieve accurate, smooth, and physically reasonable updates of the target bounding box. The technical goal is to maintain the consistency of the target's appearance while avoiding target bounding box jumps, stretching, or drifts caused by instantaneous noise, feature jitter, or local mismatches, thus ensuring the spatiotemporal continuity and visual rationality of the tracking results.

[0102] It should be noted that the calculation formula for the target bounding box in the current frame is not simply a matter of superimposing the displacement onto the original box center, but rather takes into account the effect of scale on the target center position. When the target is enlarged ( When zooming in, if only the object is translated without adjusting the center, the bounding box will "shift backward"—because the target occupies more space after zooming in, and its geometric center should be slightly adjusted in the opposite direction of the movement to maintain visual "center alignment." In the formula... Designed for this purpose, it ensures that the target box expands based on the original visual center after scaling, avoiding the "stretching" effect.

[0103] In one alternative implementation scenario, a customer gets up and leaves the bank counter; their bounding box is in the previous frame. That is, the center is located at (145, 115), with a width of 50 and a height of 70. The current frame calculates... (Slightly closer to the camera), displacement If only the target is translated: the new center = (145+10, 115+5) = (155, 120), and the width and height become 55×77. However, the upper left corner of the target box will be located at (155-55 / 2, 120-77 / 2) = (127.5, 81.5), and the target will appear to have shifted to the lower right.

[0104] After compensation using this formula:

[0105]

[0106]

[0107]

[0108] The new bounding box has its top-left corner at (127, 81) and its center at (154.5, 119.5), perfectly matching the motion trend. Visually, the target smoothly enlarges and moves forward without any "tailing" or "jumping" sensation. The above steps allow the "motion parameters" output by the tracking module to be transformed into "standard target boxes that can be directly used by the detection and synthesis modules," thereby ensuring the consistency of the interfaces between the tracking, detection, and learning modules in the TLD architecture.

[0109] Optionally, in the machine vision-based target tracking method provided in this embodiment of the invention, the step of updating the position of the target box in the current video frame based on the scale change value, displacement value, and the position information of the target box in the previous video frame to obtain the target tracking result includes: determining the size of the previous target box and the coordinates of the center point of the previous target box through the position information of the target box in the previous video frame; adjusting the size of the previous target box according to the scale change value to obtain the updated target box size; determining the coordinates of the center point of the updated target box based on the displacement value and the coordinates of the center point of the previous target box; and updating the position of the target box in the current video frame based on the updated target box size and the coordinates of the center point of the updated target box to obtain the target tracking result.

[0110] In this embodiment, the technical process of gradually updating the target bounding box based on scale changes and displacement is the core computational link for the continuous modeling of target motion in this invention. Its essence is to accurately map the abstract motion parameters (scale changes, displacement) into target bounding box coordinates and dimensions that conform to visual geometry, thereby ensuring that the tracking results are consistent, reasonable and interpretable in space.

[0111] Specifically, the basic parameters of the target bounding box in the previous frame are first obtained. The updated target bounding box from the previous frame is read, the implicit calculation center point is calculated, the target bounding box size is updated according to the scale change value, and the width and height of the target bounding box are scaled proportionally. This operation can achieve proportional scaling, which conforms to the rigid body motion assumption and avoids target distortion caused by non-uniform deformation. Further, a new center point is determined based on the displacement value and the center point. The key is that the displacement is not directly applied to the upper left corner, but acts on the center point and compensates for the influence of scale change on the center. In the traditional approach, if only translation is performed without adjusting the center, the target bounding box will shift to the lower right. In this embodiment, the center point is finely adjusted in the opposite direction through the compensation term to ensure that the target bounding box still expands based on the visual center after magnification. Then, the new center point is calculated, and the coordinates of the upper left corner are deduced from the new center point and the new size.

[0112] Optionally, in the machine vision-based target tracking method provided in the embodiments of the present invention, after updating the position of the target box in the current video frame according to the scale change value, displacement value and the position information of the target box in the previous video frame, the target tracking method further includes: if the position of the target box in the current video frame is successfully updated, performing clustering processing on the updated target box and the target detection box, and outputting the resulting comprehensive target box as the target tracking result, wherein the target detection box is a first type of target detection box or a second type of target detection box.

[0113] This embodiment clusters and fuses the target bounding box and the detection bounding box after successful tracking, outputting a comprehensive target bounding box. Essentially, it constructs a dynamic fusion strategy that prioritizes tracking and supplements it with detection. By fusing two independent but complementary information sources—the highly efficient but easily interfered-prone tracking module output (B1) and the highly robust but computationally expensive detection module output (B2)—it significantly improves the stability and accuracy of tracking while ensuring real-time performance. The complete implementation logic is as follows:

[0114] B1 (tracking module output) is the target bounding box obtained based on SURF+GMS+HOG confidence screening and median motion estimation. It has high frame rate, low latency and good continuity, but is sensitive to occlusion and drastic deformation. B2 (detection module output) is the candidate target bounding box detected in the local or whole image by variance classifier, random circle classifier and nearest neighbor classifier. It has strong discriminative power and can cope with occlusion and lighting changes, but has high computational cost and low frame rate.

[0115] In this embodiment of the invention, B1 and B2 are considered as a set of candidate bounding boxes, and a clustering algorithm based on position and size overlap (such as DBSCAN or weighted mean clustering) is used to merge them. The similarity between the two boxes is defined as: If IoU > a threshold (e.g., 0.6), then the two are considered to be the same target and weighted fusion is performed. ,in, That is, the tracking box accounts for 70% of the weight, and the detection box accounts for 30%, ensuring that the system prioritizes and trusts efficient tracking results, and only fine-tunes when the detection box is significantly better.

[0116] Final output As the tracking result of the current frame, it retains the continuity of the tracking module and incorporates the accuracy of the detection module, achieving a 1+1>2 effect.

[0117] In one optional implementation scenario, during bank lobby surveillance, customer A is withdrawing cash at a self-service terminal when a passing employee suddenly obscures their view for 1-2 frames. In frame t: due to feature point loss, the tracking module's output bounding box B1 drifts slightly (5 pixels to the right, slightly enlarged); the detection module still identifies the target in the local search (because the random Circle feature can still identify the customer's clothing texture), and the output bounding box B2 is accurately positioned; the fusion module calculates IoU=0.68, determines it to be the same target, and uses weighted fusion to correct the target box position back to the correct area, slightly reducing the scale. This corrects the tracking drift without interrupting continuity, without triggering a global search or Kalman prediction, and the system runs smoothly. Without this fusion mechanism, the system would either continuously output the drifting B1 (leading to a misjudgment of "abnormal customer movement") or switch to the detection result (causing the target box to "jump"), affecting the accuracy of behavior analysis.

[0118] Through steps S201 to S204, a target video sequence containing at least two video frames can be obtained first. Based on the predetermined target to be tracked, feature acquisition is performed in the current video frame to obtain a feature point set containing N target feature points, where N is a positive integer. Then, matching analysis is performed between the feature point set and the previous feature point set corresponding to the previous video frame to obtain a feature point pair set containing M feature point pairs, where M is a positive integer less than N. Then, the scale change value and displacement value corresponding to the target are calculated based on the feature point pair set, and the position information of the target box in the previous video frame is obtained. The target box is used to identify the target position in the video frame. Finally, based on the scale change value, displacement value, and the position information of the target box in the previous video frame, the position of the target box in the current video frame is updated to obtain the target tracking result.

[0119] In this embodiment of the invention, a multi-scale adaptive illumination processing method is employed. By combining SURF feature matching and HOG feature point confidence evaluation, the goal of accurately and efficiently tracking targets under complex illumination conditions and varying target scales is achieved. This enables intelligent monitoring and customer service optimization in banking scenarios. Specifically, this embodiment addresses the insufficient robustness of the traditional TLD algorithm to changes in illumination and target scale by introducing a multi-scale adaptive illumination adjustment mechanism. This ensures stable operation of the algorithm under different illumination environments while maintaining high sensitivity to changes in target scale, significantly improving the algorithm's adaptability and accuracy.

[0120] Furthermore, to address the challenge of occlusion, this invention adds an occlusion detection and motion prediction step. A fusion algorithm combining Kalman filters and feature point matching is used to predict the target's position in the next frame, effectively overcoming tracking interruptions caused by local occlusion and significantly enhancing the algorithm's tracking continuity and reliability in complex scenes. Based on this, feature point pair confidence evaluation further eliminates low-quality point pairs during the matching process, ensuring the accuracy of the tracking results. Through these technical steps, this invention not only significantly improves the accuracy and stability of target tracking but also specifically addresses the needs of the banking environment, enabling real-time monitoring of customer behavior, automatic identification of high-net-worth clients, and precise allocation of internal resources. This solves the shortcomings of traditional monitoring methods in terms of efficiency, accuracy, and intelligent response, and further addresses the technical problem of decreased tracking accuracy and continuity in target tracking algorithms under complex and changing environments.

[0121] The present invention will now be described in conjunction with another specific embodiment.

[0122] This invention discloses an optional overall architecture of a machine vision-based TLD target tracking algorithm. The original TLD algorithm mainly comprises four parts: a tracking module, a detection module, a learning module, and a synthesis module. The algorithm proposed in this embodiment is based on the architecture of the original TLD algorithm. The tracking module uses a SURF feature matching-based tracking module. The detection module is composed of the variance classifier and random circle feature classifier of the original TLD algorithm detection module, and the nearest neighbor classifier of the TLD algorithm detection module. The learning module still uses the PN learning strategy of the TLD algorithm learning module. The specific workflow of this algorithm is as follows:

[0123] (1) Algorithm initialization. First, the target bounding box of the initial frame is manually marked, and the SURF feature points of the current frame are detected. At the same time, the initial target image block is subjected to radiative transformation to generate positive samples, and the background image block is used as negative samples. The initial detection module is trained using these positive and negative samples.

[0124] (2) Determine the video frame. Determine whether the current frame number is greater than 2. If the current frame number is greater than 2, proceed to step (3); otherwise, proceed to step (4).

[0125] (3) Determine whether the target is occluded. If the target is determined to be occluded, the Kalman filter and feature point matching fusion algorithm is used to predict the motion of the target's position in the next frame, and the optimal estimate of the Kalman filter in the previous frame is used as the observation value, and proceed to step (8); if the target is determined not to be occluded, proceed to step (4).

[0126] (4) Call the tracking module. The tracking module uses the SURF feature matching method to match images of two adjacent frames, then uses the GMS method to eliminate mismatches, and calculates the feature point pairs. The HOG feature similarity of 10 image blocks is used to evaluate the confidence of feature point pairs. Feature point pairs with confidence below a set threshold are filtered out. Finally, the tracking target box of the current frame is calculated as the result B1 of the tracking module. If the tracking module is considered to have successfully tracked, the tracking module is considered to have failed. Proceed to step (5).

[0127] (5) Call the detection module. First, perform adaptive histogram equalization with limited contrast. If the tracking module tracks successfully, use a local search method to filter the bounding boxes within the region of interest; otherwise, use a global search method to filter the bounding boxes within the entire image. The bounding boxes that pass the variance classifier, random circle classifier, and nearest neighbor classifier are taken as the result B2 of the detection module, indicating that the detection module has tracked successfully; otherwise, it is considered that the detection module has failed to track. Proceed to step (6).

[0128] (6) Call the synthesis module. If both the tracking module and the detection module track successfully, then B1 and B2 are clustered, and the tracking result is the synthesis result of B1 and B2, where B1 has a higher weight, and proceed to step (7); if the tracking module tracks successfully but the detection module fails to track, then B1 is output as the result and proceed to step (7); if the tracking module fails to track but the detection module tracks successfully, then the optimal target box in B2 is output as the result and proceed to step (7); if both the tracking module and the detection module fail to track, then the current frame is considered to have failed to track, and proceed to step (8).

[0129] (7) Determine whether the tracking result of the algorithm meets the learning conditions of the learning module. If it does, update the sample library of the learning module with the tracking result of the algorithm and train the classifier in the detection module; otherwise, do not train the classifier in the detection module. Proceed to step (8).

[0130] (8) Determine whether the current frame is the last frame. If it is, the tracking is considered to have ended; otherwise, proceed to step (2).

[0131] This invention proposes a machine vision-based TLD target tracking algorithm based on the above implementation methods. Addressing the shortcomings of the original TLD algorithm in robustness to changes in illumination and target scale, a multi-scale adaptive illumination TLD algorithm is proposed. To address the TLD algorithm's insufficient resistance to occlusion in moving backgrounds, occlusion detection and motion prediction are added, resulting in an improved occlusion-resistant TLD algorithm for moving backgrounds. This invention retains the monitoring capabilities of the original TLD algorithm while exhibiting adaptability to occlusion, deformation, illumination changes, and scale variations, thus meeting the practical needs of real-world tracking. This invention enables better real-time capture of customer and employee targets, leading to improved banking service levels.

[0132] The invention will now be described in conjunction with another alternative embodiment.

[0133] Example 2

[0134] This invention also provides a machine vision-based target tracking device. It should be noted that the machine vision-based target tracking device of this invention includes multiple implementation units, which can be used to execute the machine vision-based target tracking method provided in the first embodiment above. Each implementation unit corresponds to each implementation step in the first embodiment above.

[0135] Figure 4 This is a schematic diagram of an optional machine vision-based target tracking device according to an embodiment of the present invention, such as... Figure 4 As shown, the device may include: a data acquisition unit 41, an analysis unit 42, an acquisition unit 43, and an update unit 44.

[0136] The acquisition unit 41 is used to acquire a target video sequence containing at least two video frames, and to perform feature acquisition in the current video frame according to the pre-determined target to be tracked, to obtain a feature point set containing N target feature points, where N is a positive integer.

[0137] Analysis unit 42 is used to perform matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame to obtain a feature point pair set containing M feature point pairs, where M is a positive integer less than N.

[0138] The acquisition unit 43 is used to calculate the scale change value and displacement value corresponding to the target based on the set of feature points, and to acquire the position information of the target box in the previous video frame, wherein the target box is used to identify the target position in the video frame.

[0139] The update unit 44 is used to update the position of the target box in the current video frame based on the scale change value, displacement value and the position information of the target box in the previous video frame, so as to obtain the target tracking result.

[0140] The aforementioned machine vision-based target tracking device first acquires a target video sequence containing at least two video frames through the acquisition unit 41, and performs feature acquisition in the current video frame according to the pre-determined target to be tracked, obtaining a feature point set containing N target feature points, where N is a positive integer. Then, the analysis unit 42 performs matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame, obtaining a feature point pair set containing M feature point pairs, where M is a positive integer less than N. Then, the acquisition unit 43 calculates the scale change value and displacement value corresponding to the target based on the feature point pair set, and obtains the position information of the target box in the previous video frame, where the target box is used to identify the target position in the video frame. Finally, the update unit 44 updates the position of the target box in the current video frame according to the scale change value, displacement value, and the position information of the target box in the previous video frame, thus obtaining the target tracking result.

[0141] In this embodiment of the invention, a multi-scale adaptive illumination processing approach is adopted. By combining SURF feature matching and HOG feature point confidence evaluation, the goal of accurately and efficiently tracking targets under complex illumination conditions and changes in target scale is achieved, thereby realizing the technical effect of intelligent monitoring and customer service optimization in banking scenarios. Specifically, this embodiment of the invention addresses the insufficient robustness of the traditional TLD algorithm to changes in illumination and target scale by introducing a multi-scale adaptive illumination adjustment mechanism. This ensures that the algorithm can still operate stably under different illumination environments, while maintaining a high sensitivity to changes in target scale, significantly improving the algorithm's adaptability and accuracy.

[0142] Furthermore, to address the challenge of occlusion, this invention adds an occlusion detection and motion prediction step. A fusion algorithm combining Kalman filters and feature point matching is used to predict the target's position in the next frame, effectively overcoming tracking interruptions caused by local occlusion and significantly enhancing the algorithm's tracking continuity and reliability in complex scenes. Based on this, feature point pair confidence evaluation further eliminates low-quality point pairs during the matching process, ensuring the accuracy of the tracking results. Through these technical steps, this invention not only significantly improves the accuracy and stability of target tracking but also specifically addresses the needs of the banking environment, enabling real-time monitoring of customer behavior, automatic identification of high-net-worth clients, and precise allocation of internal resources. This solves the shortcomings of traditional monitoring methods in terms of efficiency, accuracy, and intelligent response, and further addresses the technical problem of decreased tracking accuracy and continuity in target tracking algorithms under complex and changing environments.

[0143] Furthermore, the acquisition unit includes: a first calculation module, used to calculate the distance ratio between corresponding two feature points for each feature point pair to obtain a preliminary value of scale change; a second calculation module, used to calculate the median of the M preliminary values ​​of scale change corresponding to the M feature point pairs to obtain a scale change value; a third calculation module, used to calculate the lateral distance difference and longitudinal distance difference between corresponding two feature points for each feature point pair, and calculate a preliminary value of displacement based on the lateral distance difference and longitudinal distance difference; and a fourth calculation module, used to calculate the median of the M preliminary values ​​of displacement corresponding to the M feature point pairs to obtain a displacement value.

[0144] Furthermore, the updating unit includes: a first determining module, used to determine the size of the previous target box and the coordinates of the center point of the previous target box based on the position information of the target box in the previous video frame; an adjusting module, used to adjust the size of the previous target box according to the scale change value to obtain the updated target box size; a second determining module, used to determine the coordinates of the center point of the updated target box according to the displacement value and the coordinates of the center point of the previous target box; and a first updating module, used to update the position of the target box in the current video frame according to the updated target box size and the coordinates of the center point of the updated target box to obtain the target tracking result.

[0145] Furthermore, the target tracking device includes: a judgment module, used to determine whether the target is occluded before feature acquisition in the current video frame, obtain a judgment result, and perform feature acquisition in the current video frame if the judgment result indicates that the target is not occluded; an analysis module, used to analyze the changes in the target bounding box position of the target in X consecutive historical video frames using a motion prediction algorithm if the judgment result indicates that the target is occluded, to obtain the predicted position information of the target in the current frame, where X is a specified positive integer; and a second update module, used to update the position of the target bounding box in the current video frame according to the predicted position information, to obtain the target tracking result.

[0146] Furthermore, the target tracking device also includes: a fifth calculation module, used to calculate the feature similarity of each feature point pair after obtaining a feature point pair set containing M feature point pairs, and to determine the feature similarity as the confidence level of the feature point pair; and a filtering module, used to filter out low-configuration feature point pairs corresponding to confidence levels below a preset confidence threshold, and to remove all low-configuration feature point pairs from the feature point pair set.

[0147] Furthermore, the target tracking device also includes: a first detection module, used to perform target detection on the current video frame using a global search strategy and identify a first type of target detection box when the number of elements in the feature point pair set obtained after the removal of all low-configuration feature point pairs is less than a preset threshold; and a second detection module, used to perform target detection on the current video frame using a local search strategy and identify a second type of target detection box when the number of elements in the feature point pair set obtained after the removal of all low-configuration feature point pairs is less than a preset threshold.

[0148] Furthermore, the target tracking device also includes a clustering module, which, after updating the position of the target box in the current video frame based on the scale change value, displacement value and the position information of the target box in the previous video frame, performs clustering processing on the updated target box and the target detection box if the position update of the target box in the current video frame is successful, and outputs the resulting comprehensive target box as the target tracking result, wherein the target detection box is a first type of target detection box or a second type of target detection box.

[0149] It should be noted that the aforementioned acquisition unit 41, analysis unit 42, acquisition unit 43, and update unit 44 correspond to steps S201 to S204 in Embodiment 1. The instances and application scenarios implemented by the aforementioned units and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the aforementioned modules or units may be hardware or software components stored in a memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The aforementioned modules or units may also be part of a device and can run in the computer terminal 10 provided in Embodiment 1.

[0150] The invention will now be described in conjunction with another alternative embodiment.

[0151] Example 3

[0152] The present invention can also provide an electronic device. Figure 5 This is a structural block diagram of an electronic device that performs a machine vision-based target tracking method according to an embodiment of the present invention, such as... Figure 5 As shown, the electronic device may include: one or more ( Figure 5 (Only one is shown) processor 502, memory 504, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0153] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the machine vision-based target tracking method and device in this application embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned machine vision-based target tracking method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0154] The processor can access information and applications stored in the memory via a transmission device to perform the following steps: acquire a target video sequence containing at least two video frames, and perform feature acquisition in the current video frame based on a pre-determined target to be tracked, obtaining a feature point set containing N target feature points, where N is a positive integer; perform matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame, obtaining a feature point pair set containing M feature point pairs, where M is a positive integer less than N; calculate the scale change value and displacement value corresponding to the target based on the feature point pair set, and obtain the position information of the target bounding box in the previous video frame, where the target bounding box is used to identify the target position in the video frame; update the position of the target bounding box in the current video frame based on the scale change value, displacement value, and the position information of the target bounding box in the previous video frame, to obtain the target tracking result.

[0155] The processor can also call the information and application program stored in the memory through the transmission device to perform the following steps: calculate the distance ratio between the two corresponding feature points for each feature point pair to obtain the preliminary value of scale change; calculate the median of the M preliminary values ​​of scale change corresponding to the M feature point pairs to obtain the scale change value; for each feature point pair, calculate the difference in lateral distance and the difference in longitudinal distance between the two corresponding feature points, and calculate the preliminary value of displacement based on the difference in lateral distance and the difference in longitudinal distance; calculate the median of the M preliminary values ​​of displacement corresponding to the M feature point pairs to obtain the displacement value.

[0156] The processor can also call the information and application program stored in the memory through the transmission device to perform the following steps: determine the size of the previous target box and the coordinates of the center point of the previous target box based on the position information of the target box in the previous video frame; adjust the size of the previous target box according to the scale change value to obtain the updated target box size; determine the coordinates of the center point of the updated target box according to the displacement value and the coordinates of the center point of the previous target box; update the position of the target box in the current video frame according to the updated target box size and the updated target box center point coordinates to obtain the target tracking result.

[0157] The processor can also call the information and application program stored in the memory through the transmission device to perform the following steps: determine whether the target is occluded, obtain the determination result, and if the determination result indicates that the target is not occluded, perform feature acquisition in the current video frame; or if the determination result indicates that the target is occluded, analyze the changes in the target bounding box position of the target in X consecutive historical video frames through a motion prediction algorithm to obtain the predicted position information of the target in the current frame, where X is a specified positive integer; update the position of the target bounding box in the current video frame according to the predicted position information to obtain the target tracking result.

[0158] The processor can also call the information and application program stored in the memory through the transmission device to perform the following steps: for each feature point pair, calculate the feature similarity of the feature point pair and determine the feature similarity as the confidence of the feature point pair; filter out low-quality feature point pairs corresponding to confidence values ​​below a preset confidence threshold and remove all low-quality feature point pairs from the feature point pair set.

[0159] The processor can also call the information and application program stored in the memory through the transmission device to perform the following steps: if the number of elements in the feature point pair set obtained after removal is less than a preset threshold, a global search strategy is used to perform target detection on the current video frame and identify the first type of target detection box; or, if the number of elements in the feature point pair set obtained after removal is less than a preset threshold, a local search strategy is used to perform target detection on the current video frame and identify the second type of target detection box.

[0160] The processor can also call the information and application stored in the memory through the transmission device to perform the following steps: if the position of the target box is successfully updated in the current video frame, perform clustering processing on the updated target box and the target detection box, and output the resulting comprehensive target box as the target tracking result, wherein the target detection box is a first-class target detection box or a second-class target detection box.

[0161] This invention provides a machine vision-based target tracking scheme. Employing a multi-scale adaptive illumination processing approach, combining SURF feature matching and HOG feature point confidence evaluation, it achieves accurate and efficient target tracking under complex lighting conditions and varying target scales, thereby realizing the technical effects of intelligent monitoring and customer service optimization in banking scenarios. Specifically, this invention addresses the insufficient robustness of the traditional TLD algorithm to changes in illumination and target scale by introducing a multi-scale adaptive illumination adjustment mechanism. This ensures stable operation of the algorithm under different lighting environments while maintaining high sensitivity to changes in target scale, significantly improving the algorithm's adaptability and accuracy.

[0162] Furthermore, to address the challenge of occlusion, this invention adds an occlusion detection and motion prediction step. A fusion algorithm combining Kalman filters and feature point matching is used to predict the target's position in the next frame, effectively overcoming tracking interruptions caused by local occlusion and significantly enhancing the algorithm's tracking continuity and reliability in complex scenes. Based on this, feature point pair confidence evaluation further eliminates low-quality point pairs during the matching process, ensuring the accuracy of the tracking results. Through these technical steps, this invention not only significantly improves the accuracy and stability of target tracking but also specifically addresses the needs of the banking environment, enabling real-time monitoring of customer behavior, automatic identification of high-net-worth clients, and precise allocation of internal resources. This solves the shortcomings of traditional monitoring methods in terms of efficiency, accuracy, and intelligent response, and further addresses the technical problem of decreased tracking accuracy and continuity in target tracking algorithms under complex and changing environments.

[0163] Those skilled in the art will understand that Figure 5 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones, tablets, handheld computers, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 5 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 5 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 5 The different configurations shown.

[0164] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0165] The invention will now be described in conjunction with another alternative embodiment.

[0166] Example 4

[0167] This invention also provides a computer-readable storage medium. Optionally, in this invention, the computer-readable storage medium can be used to store the program code executed by the machine vision-based target tracking method provided in Embodiment 1.

[0168] Optionally, in this embodiment of the invention, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0169] This invention also provides a computer program product, which, when executed on a data processing device, is suitable for executing the steps of a machine vision-based target tracking method: acquiring a target video sequence containing at least two video frames, and performing feature acquisition in the current video frame according to a pre-determined target to be tracked, obtaining a feature point set containing N target feature points, where N is a positive integer; performing matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame, obtaining a feature point pair set containing M feature point pairs, where M is a positive integer less than N; calculating the scale change value and displacement value corresponding to the target based on the feature point pair set, and obtaining the position information of the target bounding box in the previous video frame, where the target bounding box is used to identify the target position in the video frame; updating the position of the target bounding box in the current video frame according to the scale change value, displacement value, and the position information of the target bounding box in the previous video frame, to obtain the target tracking result.

[0170] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0171] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0172] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0173] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0174] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0175] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0176] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A target tracking method based on machine vision, characterized in that, include: Obtain a target video sequence containing at least two video frames, and perform feature acquisition in the current video frame according to the pre-determined target to be tracked to obtain a feature point set containing N target feature points, where N is a positive integer; Based on the matching analysis of the feature point set and the previous feature point set corresponding to the previous video frame, a feature point pair set containing M feature point pairs is obtained, where M is a positive integer less than N. The scale change value and displacement value corresponding to the target are calculated based on the set of feature points, and the position information of the target box in the previous video frame is obtained, wherein the target box is used to identify the target position in the video frame; Based on the scale change value, the displacement value, and the position information of the target box in the previous video frame, the position of the target box is updated in the current video frame to obtain the target tracking result.

2. The target tracking method according to claim 1, characterized in that, The steps of calculating the scale change value and displacement value corresponding to the target based on the set of feature points include: For each pair of feature points, the distance ratio between the two corresponding feature points is calculated to obtain a preliminary value of the scale change; Calculate the median of the M preliminary scale change values ​​corresponding to the M feature point pairs to obtain the scale change value; For each pair of feature points, calculate the difference in lateral distance and the difference in longitudinal distance between the two corresponding feature points, and calculate the preliminary value of displacement based on the difference in lateral distance and the difference in longitudinal distance. The displacement value is obtained by calculating the median of the M preliminary displacement values ​​corresponding to the M feature point pairs.

3. The target tracking method according to claim 1, characterized in that, The step of updating the position of the target bounding box in the current video frame based on the scale change value, the displacement value, and the position information of the target bounding box in the previous video frame to obtain the target tracking result includes: The size of the previous target box and the coordinates of its center point are determined by the position information of the target box in the previous video frame. Adjust the size of the previous target box based on the scale change value to obtain the updated target box size; The coordinates of the updated target box center point are determined based on the displacement value and the coordinates of the center point of the previous target box. The target box position is updated in the current video frame based on the updated target box size and the updated target box center point coordinates to obtain the target tracking result.

4. The target tracking method according to claim 1, characterized in that, Before feature acquisition in the current video frame, the target tracking method includes: Determine whether the target is occluded, obtain a determination result, and if the determination result indicates that the target is not occluded, perform feature acquisition in the current video frame; or... If the judgment result indicates that the target is occluded, the target bounding box position change of the target in X consecutive historical video frames is analyzed by a motion prediction algorithm to obtain the predicted position information of the target in the current frame, where X is a specified positive integer; The target bounding box position is updated in the current video frame based on the predicted position information to obtain the target tracking result.

5. The target tracking method according to claim 1, characterized in that, After obtaining a set of feature point pairs containing M feature point pairs, the target tracking method further includes: For each of the feature point pairs, the feature similarity of the feature point pairs is calculated, and the feature similarity is determined as the confidence level of the feature point pairs; Filter out low-quality feature point pairs corresponding to the confidence level that is lower than the preset confidence threshold, and remove all low-quality feature point pairs from the feature point pair set.

6. The target tracking method according to claim 5, characterized in that, After removing all the low-spec feature point pairs from the feature point pair set, the target tracking method further includes: If the number of elements in the feature point pair set obtained after removal is less than a preset threshold, a global search strategy is used to detect targets in the current video frame and identify the first type of target detection box; or, If the number of elements in the feature point pair set obtained after elimination is less than a preset threshold, a local search strategy is used to perform target detection on the current video frame and identify the second type of target detection box.

7. The target tracking method according to claim 6, characterized in that, After updating the position of the target bounding box in the current video frame based on the scale change value, the displacement value, and the position information of the target bounding box in the previous video frame, the target tracking method further includes: If the position of the target box is successfully updated in the current video frame, the updated target box and the target detection box are clustered, and the resulting composite target box is output as the target tracking result. The target detection box is either the first type of target detection box or the second type of target detection box.

8. A target tracking device based on machine vision, characterized in that, include: The acquisition unit is used to acquire a target video sequence containing at least two video frames, and to perform feature acquisition in the current video frame according to the pre-determined target to be tracked, to obtain a feature point set containing N target feature points, where N is a positive integer; The analysis unit is used to perform matching analysis based on the feature point set and the previous feature point set corresponding to the previous video frame to obtain a feature point pair set containing M feature point pairs, where M is a positive integer less than N; The acquisition unit is used to calculate the scale change value and displacement value corresponding to the target based on the set of feature points, and to acquire the position information of the target box in the previous video frame, wherein the target box is used to identify the target position in the video frame; The update unit is used to update the position of the target box in the current video frame based on the scale change value, the displacement value, and the position information of the target box in the previous video frame, so as to obtain the target tracking result.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the machine vision-based target tracking method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the machine vision-based target tracking method according to any one of claims 1 to 7.

11. A computer program product, characterized in that, Includes computer instructions, wherein when executed by a processor, the computer instructions implement the steps of the machine vision-based target tracking method according to any one of claims 1 to 7.

Images