A Multi-Target Tracking Method Based on Gaussian Strong Tracking and Adaptive Kalman Filtering

Abstract

This invention relates to the field of computer vision recognition technology, and more particularly to a multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering. The method includes: acquiring and preprocessing video images to obtain target bounding boxes, category labels, and detection confidence scores; associating trajectories based on the SORT tracking framework; updating the state of the matched trajectories using Gaussian strong tracking adaptive Kalman filtering; continuously adjusting the update strategy according to the detection confidence scores; simultaneously calculating a strong tracking factor based on the observation residuals and jointly modeling it with the confidence scores to adaptively adjust the observation noise covariance; and finally outputting the corrected trajectory information. The multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering provided by this invention can maintain trajectory stability under complex conditions such as occlusion, overlap, detection fluctuations, and abrupt changes in motion state, achieving a better balance between filter stability and agility, and improving the robustness and continuity of multi-target tracking.

Description

Technical Field

[0001] This invention relates to the field of computer vision recognition technology, and in particular to a multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering. Background Technology

[0002] Industrial sites typically contain high-risk equipment such as forklifts, robotic arms, lifting equipment, and conveyor belts, as well as high-risk areas such as electrified areas, confined space entrances, and hazardous chemical storage areas. To ensure worker safety, existing enterprises commonly manage personnel activities using methods such as camera video surveillance, manual inspections, electronic fences, access control, or location tags. Among these, video image-based personnel safety monitoring methods offer advantages such as convenient deployment, non-contact operation, wide coverage, and strong traceability, and are gradually becoming an important technological means for intelligent industrial safety management.

[0003] Existing vision-based industrial personnel safety monitoring methods typically include personnel detection, target tracking, hazardous area identification, and alarm output. The system first detects workers from the monitoring video, then establishes personnel trajectories based on the positional relationships and appearance features between adjacent frames, and finally determines whether personnel have entered prohibited areas, approached operating equipment, lingered in hazardous areas for extended periods, or experienced abnormal gatherings.

[0004] However, industrial video environments are more complex than ordinary scenarios. On the one hand, workers are easily obstructed by equipment, shelves, vehicles, pipelines, pillars, or other people; on the other hand, there are frequent intersections and overlaps between workers and mobile devices such as forklifts, robotic arms, and material carts, which can easily lead to problems such as detection frame jitter, false detections, missed detections, and decreased confidence. Traditional multi-target tracking methods often use Kalman filters for predicting and updating worker positions. However, conventional Kalman filters are usually based on the assumption of steady motion. When workers suddenly turn, stop abruptly, reappear after being obstructed by equipment, cross paths with other workers, or the detection confidence changes drastically, the tracker's internal state can easily mismatch with the current observation, resulting in trajectory drift, identity switching, tracking interruption, or incorrect identification of dangerous areas.

[0005] In industrial safety monitoring tasks, tracking stability directly affects alarm reliability. If a person's trajectory breaks in obstructed or overlapping scenarios, the system may miss intrusions into dangerous areas; if the identity changes or the trajectory drifts, the system may misjudge that a person is approaching equipment or falsely trigger an alarm. Existing methods that rely solely on adjusting observation noise based on detection confidence struggle to handle prediction model mismatch; simply introducing a strong tracking mechanism may lead to over-updating of trajectory states under low-quality detection; and using a fixed threshold to switch between different update strategies can easily cause discontinuous updates at boundary states. Summary of the Invention

[0006] (a) Technical problems to be solved

[0007] This invention provides a multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering to overcome the problems of existing technologies in complex industrial scenarios, such as tracking trajectory drift, identity switching, interruption, and unreliable judgment of dangerous areas caused by personnel occlusion, cross-over, detection errors, and sudden changes in motion state.

[0008] (II) Technical Solution

[0009] To achieve the above objectives, this invention provides a multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering, comprising the following steps:

[0010] Step S1: Acquire video images of the target industrial site and preprocess the video images;

[0011] Step S2: Use the target detection model to detect the preprocessed video images to obtain a detection set of target industrial site video images;

[0012] The detection set includes: target bounding box, category label, and detection confidence score;

[0013] Step S3: Based on the SORT tracking framework, the detection set obtained in step S2 is associated with the existing target trajectory information to distinguish between matched trajectories, unmatched trajectories, and newly appearing targets.

[0014] Step S4: For the matched trajectory, a Gaussian strong tracking adaptive Kalman filter mechanism is used to update the state in order to correct the state vector of the target trajectory.

[0015] The Gaussian strong tracking adaptive Kalman filter mechanism includes:

[0016] According to the continuously adjusted filtering update strategy based on the detection confidence, a strong tracking factor characterizing the degree of abrupt change in the target motion state is calculated based on the observation residuals, and the strong tracking factor and the detection confidence are jointly modeled to adaptively adjust the observation noise covariance.

[0017] Step S5: Output the corrected target trajectory information.

[0018] Preferably, in step S4, continuously adjusting the filter update strategy according to the detection confidence includes:

[0019] Step S41, based on the current detection confidence level Constructing continuously varying minimum NSA weights using the Sigmoid function , ,in This is a parameter for adjusting the curve slope. For the function center bias parameter, For the Sigmoid function;

[0020] Determine the strong tracking drive weight based on the minimum NSA weight. ,in This represents the strong tracking drive strength corresponding to the detection confidence level.

[0021] Preferably, in step S4, calculating the strong tracking factor characterizing the degree of abrupt change in the target's motion state includes:

[0022] Step S42: Construct the observation residual based on the difference between the predicted state and the detection result. ;

[0023]

[0024] Calculate residual covariance Through the formula:

[0025]

[0026] in, The observed value at the current moment, To predict the observed values, A forgetting factor used to balance historical residuals with current residual information;

[0027] Step S43, calculate the strong tracking factor Through the formula:

[0028]

[0029] in Represents the trace of a matrix. and It is obtained recursively from the observation residuals and their covariance;

[0030] In step S4, the strong tracking factor and the detection confidence are jointly modeled to adaptively adjust the observation noise covariance, including:

[0031] Step S44, standardize the strong tracking factor and detection confidence using the formula:

[0032]

[0033]

[0034]

[0035] in, This represents the normalized strong tracking quantity. and These represent the standardized state change information and the detection reliability information, respectively. , as well as These are the central parameter and scale parameter of the corresponding variable, respectively;

[0036] The and They are respectively Mean and standard deviation in historical statistics and Detection confidence level Mean and standard deviation in historical statistics;

[0037] Step S45, constructing the two-dimensional Gaussian function and the adaptive observation noise adjustment factor includes:

[0038]

[0039]

[0040] Step S46 yields the adaptively adjusted observation noise covariance matrix:

[0041]

[0042] in, This represents the final adaptive observation noise covariance matrix. Based on the observation noise covariance matrix.

[0043] (III) Beneficial Effects

[0044] This invention provides a multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering. It can effectively solve the problems of easy trajectory drift, easy identity switching and easy tracking interruption under complex interference conditions such as occlusion, overlap, detection fluctuation and sudden change of motion state in existing methods by jointly introducing a continuous adjustment strategy driven by detection confidence, calculation of strong tracking factor based on observation residual and adaptive observation noise adjustment method of the two in the Gaussian strong tracking adaptive Kalman filtering mechanism.

[0045] This method can continuously adjust the filter update strategy according to the reliability of the current detection results, avoiding state jumps caused by fixed threshold switching. At the same time, by sensing the intensity of sudden changes in the target's motion state in real time, and on this basis, jointly modeling the state change information and detection reliability information, the observation noise covariance is dynamically optimized, so that the filter can maintain smooth tracking when the target moves smoothly and maintain stable updates when the target's state changes suddenly or the observation quality deteriorates, thereby achieving a better balance between the stability and agility of the filter. Attached Figure Description

[0046] Figure 1 This invention presents a flowchart of a multi-target tracking method based on Gaussian strong tracking and adaptive Kalman filtering.

[0047] Figure 2 A radar chart comparing Kalman filter algorithms is shown. Detailed Implementation

[0048] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The technical solutions in the embodiments of the present invention will be clearly and completely described. 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 are within the scope of protection of the present invention.

[0049] In the description of this invention, it is necessary to understand that the orientations or positional relationships indicated by terms such as "upper," "lower," "left," "right," "inner," "outer," "top," and "bottom" are based on the orientations or positional relationships shown in the accompanying drawings. They are intended only to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the components referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0050] like Figure 1-2 As shown, this invention provides a multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering, comprising the following steps:

[0051] Step S1: Acquire video images of the target industrial site and preprocess the video images;

[0052] Step S1 is used to eliminate the interference of complex industrial environments on subsequent target detection and state estimation through a standardized and enhanced image preprocessing process, providing high-quality input data.

[0053] The preprocessing in step S1 includes:

[0054] Step S11: Decode and synchronize the acquired video images in time, extract frames from the decoded video to obtain a single frame image;

[0055] This video decoding function is used to convert real-time video streams or offline video files of different encapsulation formats into standard digital image sequences, and to synchronize the timestamps of video streams collected by multi-angle camera equipment in the industrial field, ensuring the time consistency of personnel trajectories in multi-view tracking scenarios.

[0056] Step S12: Unify the resolution and normalize the size of the single frame image, adjust the brightness and enhance the contrast of the normalized image, and suppress motion blur and remove local noise.

[0057] This resolution unification and size normalization function is used to unify the differentiated images captured by different cameras to a preset size, thereby reducing the computational complexity of subsequent target detection models and SORT tracking frameworks.

[0058] This brightness adjustment and contrast enhancement function is used to improve image visibility in backlit, highly reflective, or low-light environments, and to enhance the confidence stability of the personnel detection frame.

[0059] The motion blur suppression and local noise removal functions are designed to filter out image blur caused by high-speed mobile devices such as forklifts and robotic arms in industrial settings, as well as pixel interference caused by electrical noise and compression artifacts, thereby reducing detection jitter and trajectory drift caused by image quality issues.

[0060] Step S2: Use the target detection model to detect the preprocessed video images to obtain a detection set of target industrial site video images;

[0061] The detection set includes: target bounding boxes, category labels, and detection confidence scores;

[0062] In step S2, the video frames of the video image are detected based on the YOLO target detection model. In actual work, considering the characteristics of changing personnel postures, different equipment shapes and complex backgrounds in industrial scenarios, YOLOv8, YOLOv10 and YOLO11 series algorithms are used to improve the recognition accuracy of work clothes, safety helmets and specific equipment features.

[0063] The target bounding box is represented as ,in , With the center coordinates, The aspect ratio or width of the bounding box. This represents the height of the bounding box.

[0064] Step S3: Based on the SORT tracking framework, the detection set obtained in step S2 is associated with the existing target trajectory information to distinguish between matched trajectories, unmatched trajectories, and newly appearing targets.

[0065] The SORT tracing framework includes DeepSORT, StrongSORT, and BoT-SORT, among others.

[0066] Step S3 is used to solve the problems of trajectory interruption and identity change caused by personnel crossing, equipment interweaving, and short-term occlusion in industrial scenarios. Through multi-dimensional cost measurement and hierarchical matching strategy, an accurate correspondence between the detection result of the current frame and the historical trajectory is established.

[0067] In step S3, the detection set is associated with the existing target trajectory information by a combination of cascaded matching and Hungarian algorithm. The matching cost is the weighted sum of Mahalanobis distance, intersection-union ratio and appearance feature cosine distance between the predicted state and the detection result.

[0068] This cascaded matching prioritizes matching the most frequently occurring trajectories, rather than a simple one-to-one global matching. It addresses the common short-term occlusion issues in industrial settings by avoiding the initialization of new trajectories due to a single missed detection, thereby reducing identity switching and maintaining the continuity of trajectories during occlusion.

[0069] The Mahalanobis distance measures the distance between the detection box and the Kalman filter prediction box in the motion state space. It is used to eliminate matching pairs with obviously inconsistent motion patterns. The intersection-union ratio is used to calculate the proportion of the overlapping area between the detection box and the prediction box. It is robust to minor offsets of the bounding box caused by camera perspective or detection jitter in industrial scenarios. The appearance feature cosine distance is used to extract the depth appearance features of people in the detection box and compare them with the feature library of historical trajectories to avoid misidentification caused by appearance confusion after occlusion is removed.

[0070] It is important to note that before performing the association operation in step S3, it is necessary to first determine whether there is an established target trajectory. If there is no historical trajectory, the process will jump directly to the new target initialization process. If there is, a priori prediction will be performed on each trajectory in the previous time step to obtain the prediction status at the current time step. The system will then associate the prediction results with the detection results of the current frame. Based on the association results, the targets will be divided into matched trajectories, unmatched trajectories, and newly appearing targets, and a differentiated processing strategy will be adopted for different categories.

[0071] For successfully matched detection boxes and historical trajectories, they are sent to the Gaussian strong tracking adaptive Kalman filter process in step S4 for state update, ensuring that the trajectory can still maintain stable correction under complex interference.

[0072] For existing trajectories for which no corresponding detection box is found in the current frame, they are briefly retained based on DeepSORT;

[0073] For detection boxes that fail to match any historical trajectories in the current frame, they are determined to be newly appearing targets. Newly appearing targets need to undergo stability verification over several consecutive frames before a new trajectory number is initialized.

[0074] Step S4: For the matched trajectory, a Gaussian strong tracking adaptive Kalman filter mechanism is used to update the state in order to correct the state vector of the target trajectory.

[0075] This Gaussian strong tracking adaptive Kalman filter mechanism includes:

[0076] Based on the detection confidence, the filtering update strategy is continuously adjusted. At the same time, the strong tracking factor, which characterizes the degree of abrupt change in the target motion state, is calculated based on the observation residual. The strong tracking factor and the detection confidence are then jointly modeled to adaptively adjust the observation noise covariance.

[0077] In step S4, the continuous adjustment of the filter update strategy based on the detection confidence includes:

[0078] Step S41, based on the current detection confidence level Constructing continuously varying minimum NSA weights using the Sigmoid function , ,in This is a parameter for adjusting the curve slope. For the function center bias parameter, For the Sigmoid function;

[0079] The strong tracking driving weight is determined based on the minimum NSA weight. ,in This represents the strong tracking drive strength corresponding to the detection confidence level;

[0080] By using the nonlinear mapping of the Sigmoid function, a smooth transition of the update strategy is achieved. Unlike traditional methods that set a fixed confidence threshold, this method allows for the retention of certain update weights even at low confidence levels, avoiding trajectory jumps caused by threshold switching and ensuring continuous utilization of detection under smoke and strong light interference.

[0081] In step S4, the calculation of the strong tracking factor, which characterizes the degree of abrupt change in the target's motion state, includes:

[0082] Step S42: Construct the observation residual based on the difference between the predicted state and the detection result. ;

[0083]

[0084] Calculate residual covariance Through the formula:

[0085]

[0086] in, The observed value at the current moment, To predict the observed values, A forgetting factor used to balance historical residuals with current residual information;

[0087] Step S43, calculate the strong tracking factor Through the formula:

[0088]

[0089] in Represents the trace of a matrix. and It is obtained recursively from the observation residuals and their covariance;

[0090] By monitoring changes in the observation residuals and their covariance, the degree of mismatch between the target's motion state and the filter prediction model is quantified. When a person suddenly stops, turns, or is obstructed and then reappears, the observation residuals are analyzed. It will increase significantly, leading to a strong tracking factor. This method utilizes the tracking capability of the strong tracking factor adaptive enhancement filter to quickly correct trajectory deviations and prevent tracking loss caused by prediction model lag, thereby solving the problem of mismatch in conventional Kalman filtering when motion states change abruptly.

[0091] In step S4, the strong tracking factor and the detection confidence are jointly modeled to adaptively adjust the observation noise covariance, including:

[0092] Step S44, standardize the strong tracking factor and detection confidence using the formula:

[0093]

[0094]

[0095]

[0096] in, This represents the normalized strong tracking quantity. and These represent the standardized state change information and the detection reliability information, respectively. , as well as These are the central parameter and scale parameter of the corresponding variable, respectively. and They are respectively Mean and standard deviation in historical statistics and Detection confidence level Mean and standard deviation in historical statistics;

[0097] Step S45, constructing the two-dimensional Gaussian function and the adaptive observation noise adjustment factor includes:

[0098]

[0099]

[0100] Step S46 yields the adaptively adjusted observation noise covariance matrix:

[0101]

[0102] in, This represents the final adaptive observation noise covariance matrix. Based on the observation noise covariance matrix;

[0103] This two-dimensional Gaussian adaptive observation noise mechanism achieves dual adaptive adjustment through joint modeling. When the detection confidence is high and the state change is strong, the noise is appropriately reduced for a rapid response. When the detection confidence is low or the state is stable, the noise weight is increased to suppress interference and improve trajectory robustness in complex industrial environments.

[0104] Step S5: Output the corrected target trajectory information.

[0105] like Figure 2 The radar chart comparing Kalman filter algorithms shown includes: High-Order Tracking Accuracy (HOTA), Multi-Target Tracking Accuracy (MOTA), Multi-Target Tracking Precision (MOTP), Identity Preservation F1 Value (IDF1), Detection Accuracy (DetA), Association Accuracy (AssA), and Positioning Accuracy (LocA).

[0106] It is understood that the various embodiments mentioned above in this invention can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this invention will not elaborate further.

[0107] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0108] This invention provides a multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering. By jointly introducing a detection confidence-driven continuous adjustment strategy, a strong tracking factor calculation based on observation residuals, and an adaptive observation noise adjustment method that coordinates the two into the Gaussian strong tracking adaptive Kalman filtering mechanism, it effectively solves the problems of easy trajectory drift, easy identity switching, and easy tracking interruption of existing methods under complex interference conditions such as occlusion, overlap, detection fluctuations, and sudden changes in motion state.

[0109] This method can continuously adjust the filter update strategy according to the reliability of the current detection results, avoiding state jumps caused by fixed threshold switching. At the same time, by sensing the intensity of sudden changes in the target's motion state in real time, and on this basis, jointly modeling the state change information and detection reliability information, the observation noise covariance is dynamically optimized, so that the filter can maintain smooth tracking when the target moves smoothly and maintain stable updates when the target's state changes suddenly or the observation quality deteriorates, thereby achieving a better balance between the stability and agility of the filter.

[0110] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering, characterized in that, Includes the following steps: Step S1: Acquire video images of the target industrial site and preprocess the video images; Step S2: Use the target detection model to detect the preprocessed video images to obtain a detection set of target industrial site video images; The detection set includes: target bounding box, category label, and detection confidence score; Step S3: Based on the SORT tracking framework, the detection set obtained in step S2 is associated with the existing target trajectory information to distinguish between matched trajectories, unmatched trajectories, and newly appearing targets. Step S4: For the matched trajectory, a Gaussian strong tracking adaptive Kalman filter mechanism is used to update the state, so as to correct the state vector of the target trajectory. The Gaussian strong tracking adaptive Kalman filter mechanism includes: According to the continuously adjusted filtering update strategy based on the detection confidence, a strong tracking factor characterizing the degree of abrupt change in the target motion state is calculated based on the observation residuals, and the strong tracking factor and the detection confidence are jointly modeled to adaptively adjust the observation noise covariance. In step S4, the strong tracking factor and the detection confidence are jointly modeled to adaptively adjust the observation noise covariance, including: The strong tracking factor and detection confidence are standardized using the following formula: in, As a strong tracking factor, This represents the normalized strong tracking value. and These represent the standardized state change information and the detection reliability information, respectively. , as well as These are the central parameter and scale parameter of the corresponding variable, respectively; The and They are respectively Mean and standard deviation in historical statistics and Detection confidence level Mean and standard deviation in historical statistics; Constructing a two-dimensional Gaussian function and an adaptive observation noise adjustment factor includes: The adaptively adjusted observation noise covariance matrix is ​​obtained as follows: in, To test confidence level Construct continuously varying minimum NSA weights using the Sigmoid function. This represents the final adaptive observation noise covariance matrix. Based on the observation noise covariance matrix; Step S5: Output the corrected target trajectory information.

2. The multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering according to claim 1, characterized in that, In step S4, the continuous adjustment of the filter update strategy based on the detection confidence includes: Step S41, based on the current detection confidence level Constructing continuously varying minimum NSA weights using the Sigmoid function , ,in This is a parameter for adjusting the curve slope. For the function center bias parameter, For the Sigmoid function; The strong tracking driving weight is determined based on the minimum NSA weight. ,in This represents the strong tracking drive strength corresponding to the detection confidence level.

3. The multi-target tracking method based on Gaussian strong tracking adaptive Kalman filtering according to claim 2, characterized in that, In step S4, calculating the strong tracking factor characterizing the degree of abrupt change in the target's motion state includes: Step S42: Construct the observation residual based on the difference between the predicted state and the detection result. ; Calculate residual covariance Through the formula: in, The observed value at the current moment, To predict the observed values, A forgetting factor used to balance historical residuals with current residual information; Step S43, calculate the strong tracking factor Through the formula: in Represents the trace of a matrix. and It is obtained recursively from the observation residuals and their covariance.

Images