Dynamic target monitoring method and system based on artificial intelligence

Abstract

The application discloses a dynamic target monitoring method and system based on artificial intelligence, and relates to the technical field of dynamic target monitoring; the application matches historical detection data for a target frame from a historical detection library based on model consistency and environment consistency, obtains a confidence index and a detection evaluation index through statistical analysis of the historical detection data, dynamically updates a detection threshold of the target frame in combination with a preset detection evaluation benchmark and an optimal confidence interval, and then detects the target frame by using the detection threshold through an intelligent detection model; the application solves the problem that fixed threshold values and single-dimension adaptive threshold values are greatly deviated from confidence distribution under complex dynamic environments, avoids the alternation of target missed detection and false alarm, realizes the synchronous reduction of missed detection rate and false alarm rate, guarantees the stable detection of targets under complex scenes such as light variation, rain, snow, fog and background clutter, improves the perception robustness of target detection under complex scenes, and improves the adaptability of threshold values and detection models since the detection threshold values are model-specific parameters.

Description

Technical Field

[0001] This invention belongs to the field of dynamic target monitoring and involves artificial intelligence technology, specifically a dynamic target monitoring method and system based on artificial intelligence. Background Technology

[0002] Against the backdrop of the convergence of artificial intelligence and computer vision technologies, AI-based dynamic target monitoring methods have been widely applied in the field of target monitoring and have become the mainstream monitoring approach. In existing technologies, image frames are typically extracted from video data used for monitoring dynamic targets as detection frames. Intelligent detection models such as YOLO and RetinaNet are then used to detect targets on these frames. Furthermore, the detection results are filtered by setting detection thresholds to determine the effective targets to be detected.

[0003] In existing technologies, by fixing the confidence threshold or adjusting the detection threshold based solely on the current environmental data, the confidence distribution of the intelligent detection model is prone to significant shifts in complex dynamic environments such as drastic changes in lighting, rain, snow, fog, and cluttered backgrounds. This directly leads to the problem of alternating false alarms and missed detections, making it difficult to guarantee the perceptual robustness of target detection in complex scenarios.

[0004] This invention provides a dynamic target monitoring method and system based on artificial intelligence to solve the above-mentioned technical problems. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a dynamic target monitoring method and system based on artificial intelligence.

[0006] To achieve the above objectives, a first aspect of the present invention provides a dynamic target monitoring method based on artificial intelligence, comprising: The target frame is determined, and the detection threshold of the target frame is updated using the historical detection database; the intelligent detection model uses the detection threshold to detect the target frame and obtain the detection result. Based on the detection results of the target frame, the feature similarity and positional similarity between the detected target in the target frame and the historical trajectory data are calculated, and a comprehensive matching score is obtained after weighted fusion. The overall matching score is compared with the preset tracking matching threshold to determine whether they belong to the same detection target. If they belong to the same detection target, the trajectory ID is bound to complete the target tracking.

[0007] In one possible implementation, the detection threshold of the target frame is updated using a historical detection library, including: Based on model consistency and environment consistency, historical detection data is matched for the target frame from the historical detection database; the historical detection database includes detection datasets for several image frames. Statistical analysis of historical detection data yields confidence indices and detection evaluation indices. The confidence indices include the confidence thresholds and median of the confidence thresholds for each detection dataset, while the detection evaluation indices include the average false negative rate and average false alarm rate for all detection datasets. The detection threshold of the target frame is determined based on the detection evaluation index, the preset detection evaluation benchmark, and the optimal confidence interval; wherein, the optimal confidence interval is determined by the confidence index, the detection evaluation index, and the preset detection evaluation benchmark.

[0008] In one possible implementation, historical detection data is matched for the target frame from a historical detection library based on model consistency and environment consistency, including: Extract environmental data from the target frame, and the intelligent detection model used to detect the target frame; Based on environmental feature vectors and intelligent detection models, the detection dataset is matched from the historical detection database, and the matched detection dataset is used as the historical detection data of the target frame.

[0009] In one possible implementation, the detection threshold of the target frame is determined based on detection evaluation metrics, a preset detection evaluation benchmark, and an optimal confidence interval, including: When all detection evaluation indicators meet the preset detection evaluation benchmark, the median of the confidence index is used as the detection threshold of the target frame; otherwise, proceed to the next step of judgment. Determine whether all the testing and evaluation indicators fail to meet the preset testing and evaluation benchmarks; Yes, the lower limit of the optimal confidence interval is used as the basic detection threshold, and the false alarm compensation value is calculated in combination with the average false alarm rate. The detection threshold is obtained by superimposing the basic detection threshold and the false alarm compensation value. No, if the average false alarm rate does not meet the requirements, the lower limit of the optimal confidence interval shall be used as the detection threshold; if the average false alarm rate does not meet the requirements, the upper limit of the optimal confidence interval shall be used as the detection threshold.

[0010] In one possible implementation, the formula for calculating the false alarm compensation value is: False alarm compensation value = α × (average false alarm rate in the same environment - 3%) / (maximum historical false alarm rate - 3%); where α is 0.1 by default, and the maximum historical false alarm rate is the peak value of the false alarm rate of historical data in the same environment, with a preset upper limit of 10%.

[0011] In one possible implementation, adjusting the tracking matching threshold includes: The occlusion state of the detected target is quantized and calculated based on the detection results of the target frame, and the tracking base threshold is matched based on the occlusion state. The tracking correction coefficient is calculated based on the detection threshold of the target frame; the tracking correction coefficient is then used to correct the tracking base threshold to obtain the tracking matching threshold.

[0012] In one possible implementation, the occlusion state of the detected target is quantized based on the detection results of the target frame, including: The cross-union ratio and the proportion of the visible area of ​​the detected target are calculated based on the detection results of the target frame. The occlusion status of the detected target is calculated by weighted fusion based on the cross-union ratio between targets and the proportion of the visible area of ​​the target; where the occlusion status value ranges from [0,1], where 0 represents no occlusion and 1 represents complete occlusion.

[0013] In one possible implementation, a tracking baseline threshold is matched based on the occlusion state, including: Read the threshold range for tracking and matching thresholds; Establish a linear mapping relationship between occlusion state and threshold range; obtain the tracking matching threshold from the linear mapping relationship based on the occlusion state; wherein the upper limit of the threshold range corresponds to the unoccluded state.

[0014] In one possible implementation, the tracking correction coefficients are calculated based on the detection threshold of the target frame, including: Read the threshold range of the tracking correction coefficient; Establish a linear mapping relationship between the tracking correction coefficient range and the detection threshold range; where the midpoint of the detection threshold range corresponds to a tracking correction coefficient of 0. The tracking correction coefficient for the target frame detection threshold is obtained from the second linear mapping relationship; where the larger the detection threshold, the smaller the tracking correction coefficient.

[0015] A second aspect of the present invention provides an artificial intelligence-based dynamic target monitoring system, including a dynamic monitoring module and a data storage module, wherein the data storage module is used to store data required for the dynamic target monitoring process; Dynamic monitoring module: used to determine the target frame and update the detection threshold of the target frame using the historical detection library; intelligent detection model uses the detection threshold to detect the target frame and obtain the detection result; Based on the detection results of the target frame, the feature similarity and positional similarity between the detected target in the target frame and the historical trajectory data are calculated, and a comprehensive matching score is obtained after weighted fusion; and, The overall matching score is compared with the preset tracking matching threshold to determine whether they belong to the same detection target. If they belong to the same detection target, the trajectory ID is bound to complete the target tracking.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention matches historical detection data for target frames from a historical detection database based on model consistency and environment consistency. Statistical analysis of the historical detection data yields confidence and detection evaluation indices. Combined with a preset detection evaluation benchmark and the optimal confidence interval, the detection threshold for the target frame is dynamically updated. Then, an intelligent detection model uses this detection threshold to detect the target frame. This invention solves the problem of significant shifts in confidence distribution between fixed thresholds and single-dimensional adaptive thresholds in complex dynamic environments, avoiding alternating false alarms and missed detections, achieving a simultaneous reduction in both the false alarm and missed detection rates. It ensures stable target detection in complex scenarios such as dramatic changes in lighting, rain, snow, fog, and cluttered backgrounds, improving the perceptual robustness of target detection in complex scenarios. Furthermore, the detection threshold is a model-specific parameter, enhancing the adaptability of the threshold to the detection model.

[0017] 2. This invention calculates the intersection-union ratio (IU) and visible area ratio of the detected targets based on the detection results of the target frame. After weighted fusion, it quantifies the occlusion state of the detected targets and matches it with the basic tracking threshold. Combined with the detection threshold of the target frame, it calculates the tracking correction coefficient and corrects the basic tracking threshold to obtain the tracking matching threshold. This invention realizes the bidirectional linkage and closed-loop collaboration between detection threshold adjustment and occlusion-aware tracking, solving the problem that the tracking matching strategy cannot adapt to changes in detection results in occluded and dense target scenes. It avoids target ID switching, trajectory interruption, and tracking failure, ensuring trajectory continuity and ID not being lost when the target is occluded or crosses over, effectively filtering detection box noise, improving the target association success rate, and significantly improving the overall reliability and long-term tracking stability of dynamic target tracking. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method steps of the dynamic target monitoring method in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram illustrating the steps of updating the target frame detection threshold using a historical detection library in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the steps for adjusting the tracking matching threshold in Embodiment 2 of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0021] Example:

[0022] In existing dynamic target monitoring technologies, artificial intelligence models generally use fixed confidence thresholds or adjust the thresholds based on only a single dimension of the current environmental data. This approach has obvious drawbacks: 1) The confidence distribution of the model shifts significantly under complex dynamic environments, which can easily lead to alternating missed detections and false alarms. This threshold adjustment approach lacks specificity and continuity, making it difficult to guarantee the robustness of perception in complex scenarios.

[0023] To address the aforementioned technical issues, this invention matches current environmental data with historical detection records, selects historical confidence distributions and detection results under similar environments, and adaptively determines the optimal confidence threshold for the current frame based on this. This fundamentally solves the problem of poor robustness of fixed thresholds and single-dimensional adaptive thresholds in scenarios such as drastic changes in lighting, rain, snow, fog, and cluttered backgrounds, achieving a simultaneous reduction in false alarm rate and false alarm rate, and ensuring stable target detection in complex environments.

[0024] Please see Figure 1 The first aspect of the present invention provides a dynamic target monitoring method based on artificial intelligence, comprising: determining a target frame, updating the detection threshold of the target frame using a historical detection library; and using an intelligent detection model to detect the target frame using the detection threshold to obtain a detection result. Based on the detection results of the target frame, the feature similarity and positional similarity between the detected target in the target frame and the historical trajectory data are calculated, and a comprehensive matching score is obtained after weighted fusion. The overall matching score is compared with the preset tracking matching threshold to determine whether they belong to the same detection target. If they belong to the same detection target, the trajectory ID is bound to complete the target tracking.

[0025] In this invention, the target frame refers to the image frame that the intelligent detection model will process. The target frame is extracted from video data used to monitor moving targets. The target frame is determined preferentially before the intelligent detection model completes its current task, so that the intelligent detection model can immediately use the updated detection threshold to complete the detection of the target frame.

[0026] In one example, a video stream for target detection is first acquired using a video capture device. The video stream is then segmented into frames to obtain several image frames. The intelligent detection model sequentially detects and identifies these image frames to complete target detection. If the intelligent detection model is processing a particular image frame, that image frame is designated as the current frame. While the intelligent detection model is processing the current frame, the image frame following it is identified as the target frame, and the detection threshold for the target frame is updated synchronously. When the intelligent detection model processes the target frame, its corresponding detection threshold can be called in real time, ensuring the continuity of the intelligent detection model's operation and improving the efficiency of dynamic target detection.

[0027] It is worth noting that the intelligent detection model can also process image frames in batches. In this case, target frames can be determined in batches to ensure the continuity of batch processing by the intelligent detection model.

[0028] Intelligent detection models are built upon artificial intelligence models and can improve the accuracy of target detection in complex environments by adjusting the detection threshold. For example, YOLOv8 / v10 or RetinaNet can be used as intelligent detection models for dynamic targets. These models feature fast real-time inference speed, stable standardized confidence output, and strong coupling between detection results and environmental features. They can effectively support adaptive adjustment of confidence thresholds based on environmental matching and historical detection data statistics, ensuring the robustness of dynamic target detection and the accuracy of threshold calculation.

[0029] In a preferred embodiment, updating the detection threshold of the target frame using a historical detection library includes: matching historical detection data for the target frame from the historical detection library based on model consistency and environment consistency; performing statistical analysis on the historical detection data to obtain confidence index and detection evaluation index; and determining the detection threshold of the target frame based on the detection evaluation index, a preset detection evaluation benchmark, and the optimal confidence interval.

[0030] Updating the detection threshold of the target frame using a historical detection database essentially involves determining the optimal confidence threshold for the target frame's corresponding environmental scenario through statistical analysis of the historical detection dataset. This confidence threshold is then combined with the intelligent detection model to complete target recognition in the target frame, ensuring the reliability of detection results in complex environments.

[0031] Before updating the target frame using the historical detection library, a pre-built historical detection library needs to be invoked. The historical detection library consists of detection datasets for several image frames. Each detection dataset includes the environmental feature vector, confidence set, and detection record corresponding to the image frame. The confidence set includes the confidence scores (values ​​ranging from 0 to 1) for all detected targets in the image frame. The detection record includes the number of missed detections, the number of false alarms, the number of valid targets, the missed detection rate, the false alarm rate, and the detection result of the image frame.

[0032] Because different AI detection models have fundamentally different confidence output distributions and detection characteristics, the confidence threshold is a model-specific parameter. Only by using historical detection datasets corresponding to the current intelligent detection model can the statistical validity and distribution consistency of historical detection data be guaranteed, ensuring that the calculated optimal confidence threshold fits the current model and achieves accurate and stable adaptive detection. Therefore, the detection datasets of each image frame in the historical detection library of this invention are all bound to the intelligent detection model that generated the detection dataset. When matching and obtaining historical detection data, model consistency must be ensured first, and then similar environment matching is completed based on environmental features.

[0033] Please see Figure 2 Based on model consistency and environment consistency, historical detection data is matched for the target frame from the historical detection database. This includes: extracting the environmental data of the target frame and the intelligent detection model used to detect the target frame; matching the detection dataset from the historical detection database based on the environmental feature vector and the intelligent detection model, and using the matched detection dataset as the historical detection data of the target frame.

[0034] After determining the target frame, the environmental data corresponding to the target frame is quantized to obtain an environmental feature vector. A similarity calculation algorithm is used to calculate the similarity between the environmental feature vectors of the target frame and image frames in the historical detection database. The detection dataset with a similarity greater than a preset similarity threshold and consistent with the intelligent detection model is selected as the historical detection data.

[0035] The environmental feature vector includes illumination intensity (unit: lux), image blur (0-1 quantization value), haze (unit: ppm), signal-to-noise ratio (dB), and scene complexity (0-1 quantization value), ensuring that the environmental data can be used for subsequent historical detection data matching.

[0036] It should be noted that the similarity calculation algorithm can be either Euclidean distance or cosine similarity. The higher the preset similarity threshold, the higher the consistency between the matched historical detection data and the target frame's environment. Consistency in the intelligent detection model is to avoid affecting the statistical analysis results of historical detection data; for example, the intelligent detection model structure should be consistent.

[0037] In one example, the environmental feature vector of the target frame is compared sequentially with the environmental feature vectors of each detection dataset in the historical detection database. If the environmental similarity is greater than 0.85 and the intelligent detection model is consistent, the detection dataset is included in the historical detection data. Of course, model consistency can be analyzed first, and then environmental consistency can be calculated.

[0038] If the target frame matches sufficient historical detection data in similar environments, the detection threshold of the target frame is adaptively adjusted based on the historical detection data. If insufficient historical detection data is not matched, the detection threshold is not adaptively updated. For example, if the number of detection datasets in the historical detection data corresponding to the target frame exceeds a preset threshold (e.g., 5), then the detection threshold is updated; otherwise, the detection threshold is not updated, and the default confidence threshold is still used. The default confidence threshold can be the native threshold of the intelligent detection model.

[0039] It should be noted that in dynamic target detection, the fundamental causes of confidence fluctuations, missed detections, and false alarms are changes in environmental factors such as illumination, fog, and blurriness, rather than the target itself. Targets exhibit randomness and heterogeneity in quantity, category, location, and occlusion state, lacking stable matching and statistical basis. Furthermore, the confidence threshold is a global scene parameter preceding detection and must be determined before target detection; a closed-loop process of matching, calculation, and detection can only be achieved through environmental features. Therefore, using environmental feature vectors as the basis for matching historical detection data ensures that the selected historical detection data exhibits a consistent confidence distribution pattern and statistical validity, guaranteeing the accurate and reliable calculation of the optimal confidence threshold.

[0040] The confidence metrics include the confidence threshold and the median of the confidence threshold for each detection dataset. The detection evaluation metrics include the average false negative rate and the average false alarm rate for all detection datasets. The false negative rate for each detection dataset = (number of labeled targets - number of targets detected by the model) / number of labeled targets × 100%, and the false alarm rate = number of unlabeled targets detected by the model / number of targets detected by the model × 100%.

[0041] A pre-defined detection evaluation benchmark is used to screen for the optimal confidence threshold. The detection evaluation benchmark includes the false negative rate and the false alarm rate. The detection evaluation benchmark can be set to a false negative rate ≤ 5% and a false alarm rate ≤ 3%, and the confidence interval corresponding to the detection evaluation benchmark is taken as the optimal confidence interval.

[0042] In one example, based on the detection evaluation benchmarks of a false negative rate ≤ 5% and a false alarm rate ≤ 3%, a set of qualified detection datasets is selected from historical detection data. A candidate threshold set is generated based on the confidence thresholds of these datasets. The smallest confidence threshold is selected as the lower limit from the candidate threshold set, and the largest confidence threshold is selected as the upper limit. An optimal confidence interval set is then generated based on the upper and lower limits.

[0043] In a preferred embodiment, the detection threshold of the target frame is determined based on detection evaluation indicators, a preset detection evaluation benchmark, and an optimal confidence interval. This includes: when all detection evaluation indicators meet the preset detection evaluation benchmark, the median of the confidence indicators is used as the detection threshold of the target frame; otherwise, proceed to the next step; determine whether all detection evaluation indicators do not meet the preset detection evaluation benchmark; if yes, the lower limit of the optimal confidence interval is used as the basic detection threshold, and a false alarm compensation value is calculated by combining the average false alarm rate; the basic detection threshold and the false alarm compensation value are then superimposed to obtain the detection threshold; if no, the lower limit of the optimal confidence interval is used as the detection threshold when the average false alarm rate does not meet the requirements, and the upper limit of the optimal confidence interval is used as the detection threshold when the average false alarm rate does not meet the requirements.

[0044] If the average false negative rate of the historical detection data corresponding to the target frame is too high (e.g., >5%), then the lower limit of the optimal confidence interval is selected as the detection threshold for the target frame, and the confidence threshold is lowered to reduce false negatives. If the average false alarm rate of the historical detection data corresponding to the target frame is too high (>3%), then the upper limit of the optimal confidence interval is selected as the detection threshold for the target frame, and the confidence threshold is increased to suppress false alarms. If both the average false negative rate and the average false alarm rate of the historical detection data corresponding to the target frame meet the preset detection evaluation benchmark, then the median of the confidence index is selected as the detection threshold for the target frame.

[0045] After determining the detection threshold for the target frame, its validity is verified. For example, if the detection threshold is within the detection threshold range (e.g., 0.2-0.8), the threshold is considered valid; if it exceeds the range, it is automatically corrected to the corresponding boundary value (0.2 if below 0.2, and 0.8 if above 0.8). It should be noted that the detection threshold range can be set separately for different models.

[0046] In addition to the above situations, there are also cases where both the average false alarm rate and the average missed detection rate do not meet the detection evaluation benchmark. In such cases, it is necessary to prioritize reducing the false alarm rate and then use a dynamic compensation mechanism to suppress false alarms.

[0047] The lower limit of the optimal confidence interval is used as the basic detection threshold to reduce the false negative rate. Simultaneously, a false alarm compensation value is introduced, and the sum of the false alarm compensation value and the basic detection threshold is used as the detection threshold for the target frame.

[0048] False alarm compensation value = α × (average false alarm rate in the same environment - 3%) / (maximum historical false alarm rate - 3%). α defaults to 0.1. The maximum historical false alarm rate is the peak false alarm rate of historical data in the same environment, with a preset upper limit of 10% to ensure coefficient normalization.

[0049] It should be noted that the 3% in the formula is the preset upper limit for the false alarm rate in this invention, which is part of the detection evaluation benchmark. By subtracting this benchmark value, the excess false alarm rate exceeding the standard can be obtained. Compensation values ​​are only calculated for cases where the false alarm rate exceeds the benchmark. When the false alarm rate is within the acceptable range, the compensation value is 0, and threshold correction is not triggered. That is, the false alarm compensation value is not calculated using the above formula, avoiding blindly raising the detection threshold when false alarms are normal, which could lead to missed detections. This ensures that the compensation mechanism strictly corresponds to the preset detection evaluation benchmark, and the compensation intensity is positively correlated with the degree of exceeding the false alarm benchmark, guaranteeing accurate and reasonable threshold adjustment. The same environment uses historical detection data of the target frame as the analysis content.

[0050] In one example, the detection base threshold is 0.3, the false alarm compensation value is 0.05, and the detection threshold is 0.35, which is both lower than the median of the interval to ensure detection and slightly raised to suppress false alarms.

[0051] After obtaining the detection threshold for the target frame, i.e., the optimal confidence threshold, the intelligent detection model performs target recognition on the preprocessed target frame, using the detection threshold to filter out valid targets, thus obtaining the detection result for that target frame. Simultaneously, the detection dataset corresponding to the target frame is updated to the historical detection database. The detection results for each target frame include valid detected targets, as well as their bounding boxes, appearance features, and location coordinates.

[0052] A second aspect of the present invention provides an artificial intelligence-based dynamic target monitoring system, including a dynamic monitoring module and a data storage module. The data storage module is used to store data required for the dynamic target monitoring process. The dynamic monitoring module is used to determine a target frame and update the detection threshold of the target frame using a historical detection library. An intelligent detection model uses the detection threshold to detect the target frame and obtain a detection result. Based on the detection result of the target frame, the feature similarity and position similarity between the detected target in the target frame and the historical trajectory data are calculated, and a comprehensive matching score is obtained after weighted fusion. The comprehensive matching score is compared with a preset tracking matching threshold to determine whether they belong to the same detected target. If they belong to the same detected target, the trajectory ID is bound to complete the target tracking.

[0053] The dynamic monitoring module and the data storage module communicate and / or are electrically connected. The dynamic monitoring module integrates various necessary modules for performing the above methods, and the data storage module stores the data required for the dynamic target monitoring process, such as the intelligent detection model, historical detection database, and various thresholds and their ranges.

[0054] Example:

[0055] In existing target tracking methods, the detection threshold adjustment and occlusion perception tracking are independent of each other and have no information interaction. As a result, in scenarios with occlusion, dense targets, and broken trajectories, the tracking matching strategy cannot adapt to changes in detection results, which easily leads to problems such as target ID switching, trajectory interruption, and tracking failure, making it difficult to achieve long-term stable tracking.

[0056] To address the aforementioned technical issues, this invention quantifies and perceives the occlusion state and degree of the detected target, using occlusion information as a key parameter for adjusting the tracking matching threshold. Simultaneously, it uses an adaptive detection threshold to inversely constrain the dynamic correction of the tracking matching threshold, enabling the detection and tracking processes to form a closed-loop collaboration. This effectively solves the technical challenges of poor target tracking stability and difficulty in maintaining identity consistency in occluded and complex scenarios. Even in extreme cases such as when the detected target is occluded or crosses over, it can still ensure trajectory continuity and no ID loss, significantly improving the overall reliability of dynamic target tracking.

[0057] Please see Figure 3 The tracking matching threshold is adjusted by: quantifying and calculating the occlusion state of the detected target based on the detection results of the target frame, matching the tracking base threshold based on the occlusion state; calculating the tracking correction coefficient based on the detection threshold of the target frame; and correcting the tracking base threshold using the tracking correction coefficient to obtain the tracking matching threshold.

[0058] The detection results are obtained by filtering the target frames using an intelligent detection model and updated detection thresholds. The detection results include valid targets, as well as the detection bounding box, appearance features, and location coordinates of each valid target. Simultaneously, the historical trajectory data of the detected target is read, including trajectory ID, historical location, historical appearance features, and historical occlusion status.

[0059] Based on the detection bounding boxes in the detection results, the intersection-union ratio (CIU) and the percentage of the visible area of ​​the detected targets are calculated, and the occlusion status of the detected targets is output. The occlusion status is a normalized value between 0 and 1, where 0 represents no occlusion and 1 represents complete occlusion.

[0060] Intersection over Union (IoU) refers to the ratio of the intersection area to the union area of ​​the current detected target bounding box and other target bounding boxes, used to quantify the degree of spatial overlap between targets; the visible area ratio refers to the ratio of the actual visible pixel area of ​​the current detected target to the pixel area of ​​the complete detection box, used to quantify the pixel-level visibility of the target.

[0061] When calculating the occlusion state of a detected target, the maximum IoU value between the current detected target and other targets is first extracted, and the visible area ratio of the target is calculated and converted into an occlusion ratio. Next, the maximum IoU value and the occlusion ratio are weighted and fused according to a preset weight (e.g., 0.6:0.4) to obtain the occlusion degree. Finally, the initial occlusion degree is normalized to the range of 0-1 to obtain the occlusion state (0 for no occlusion, 1 for complete occlusion). This method achieves accurate quantification of the occlusion state through the dual constraints of spatial overlap and pixel visibility, providing a reliable basis for subsequent tracking and matching threshold adjustment.

[0062] It should be noted that the maximum intersection-union ratio (IoU) between targets reflects the degree of spatial overlap at the detection box level and is a prerequisite for occlusion; the occlusion percentage reflects the actual coverage of the target at the pixel level and is a result of occlusion. The two dimensions are complementary and can completely characterize the essence of occlusion. The weight allocation follows the objective laws of occlusion formation, using spatial overlap as the primary judgment criterion and the visible area of ​​the target as the auxiliary verification criterion. That is, the weight of the maximum IoU value is greater than the weight of the occlusion percentage, ensuring that the quantization results are clearly distinguished. After fusion, the values ​​are naturally normalized to the 0-1 range, maintaining consistency with the numerical scale of subsequent threshold linkage adjustments, and avoiding the misjudgment problems that are easily caused by a single indicator. This makes the occlusion state quantification accurate, stable, and interpretable, providing a reliable basis for the adaptive correction of the tracking matching threshold.

[0063] In a preferred embodiment, matching a tracking baseline threshold based on the occlusion state includes: reading the threshold range of the tracking matching threshold; establishing a linear mapping relationship between the occlusion state and the threshold range; obtaining the tracking matching threshold from the linear mapping relationship based on the occlusion state; wherein the upper limit of the threshold range corresponds to the occlusion state being 0.

[0064] The stored threshold range for tracking matching is read, and all application tracking matching thresholds fall within this range. A linear mapping relationship 1 between the detected target occlusion state and the threshold range is established. When the occlusion state of the detected target is calculated, the corresponding tracking matching threshold can be directly matched from this linear mapping relationship 1. It is worth noting that the larger the occlusion state, the lower the corresponding tracking matching threshold, and the smaller the occlusion state, the higher the corresponding tracking matching threshold. It should be noted that the threshold range for tracking matching can be [0.3, 0.8].

[0065] In one example, if the detected target is unoccluded (occlusion level is 0), the upper limit of the tracking matching threshold range should be obtained according to the linear mapping relationship to ensure strict matching and avoid erroneous trajectory ID switching. If partially occluded (occlusion level is 0 < occlusion level < 1), the tracking matching threshold decreases linearly as the occlusion level increases to relax the target association conditions. If completely occluded (occlusion level is 1), the lower limit of the tracking matching threshold range should be obtained, retaining only the trajectory position association to prevent trajectory ID loss.

[0066] In a preferred embodiment, calculating the tracking correction coefficient based on the detection threshold of the target frame includes: establishing a second linear mapping relationship between the tracking correction coefficient and the detection threshold range; wherein, when the tracking correction coefficient is 0, it corresponds to the middle value of the detection threshold range; obtaining the tracking correction coefficient of the target frame detection threshold from the second linear mapping relationship; wherein, the larger the detection threshold, the smaller the tracking correction coefficient.

[0067] The threshold range of the stored tracking correction coefficients is read, and all applied tracking correction coefficients fall within this range. A second linear mapping relationship is established between the tracking correction coefficients and the detection threshold range. After obtaining the detection threshold of the target frame, substituting this threshold into the second linear mapping relationship yields the tracking correction coefficients. It is worth noting that the larger the detection threshold, the smaller the tracking correction coefficient, and vice versa. The median value of the detection threshold range corresponds to a tracking correction coefficient of 0. The threshold range of the tracking correction coefficients is set empirically, with a default value of [-0.3, 0.3].

[0068] A low detection threshold is often due to complex environments where the false negative rate is reduced, indicating high noise levels within the detection box. In this case, the tracking and matching threshold should be increased to filter out noise interference. If the detection threshold is moderate and falls under normal environments, the tracking and matching threshold can be directly calculated based on the occlusion status. If the detection threshold is high, it's often due to clear environments where the false alarm rate is reduced, indicating low noise interference within the detection box. In this case, the tracking and matching threshold should be slightly decreased to improve the target association success rate.

[0069] In this invention, a basic tracking threshold is determined based on the occlusion state, a tracking correction coefficient is calculated using the detection threshold, and the tracking correction coefficient is used to correct the basic tracking threshold to obtain a tracking matching threshold. The correction formula is: Tracking matching threshold = (1 + Tracking correction coefficient) × Tracking basic threshold.

[0070] First, the feature similarity and positional similarity between the detected target in the target frame and the historical trajectory data are calculated, and then weighted and fused to obtain a comprehensive matching score. This comprehensive matching score is compared with the tracking matching threshold after bidirectional linkage. If the comprehensive matching score is not less than the tracking matching threshold, it is determined to be the same detected target, and the original trajectory ID is inherited. If the comprehensive matching score is less than the tracking matching threshold, it is determined to be a new detected target or an invalid match, and trajectory creation or filtering is performed. By default, the weights of feature similarity and positional similarity are equal, set to 0.5.

[0071] When calculating feature similarity, the cosine similarity algorithm can be used. For example, the cosine of the angle between the current appearance feature vector of the detected target and the feature vector in the historical trajectory can be calculated, with a value ranging from [-1, 1]. The closer the value is to 1, the higher the feature similarity.

[0072] When calculating positional similarity, the intersection-union ratio (IUGR) can be used. For example, the intersection area and union area of ​​the current target detection box and the historical trajectory target detection boxes can be calculated, with a value range of [0,1]. The closer the value is to 1, the higher the positional overlap and the higher the similarity.

[0073] It should be noted that when tracking detected targets, multiple targets can be dynamically detected simultaneously. The tracking and matching threshold can be dynamically adjusted for each target to achieve accurate monitoring. In scenarios with multiple detected targets, matching is performed sequentially from low to high occlusion level, prioritizing the ID stability of unoccluded / weakly occluded targets.

[0074] For successfully matched targets, update the trajectory position, appearance features, and occlusion status to maintain trajectory continuity; for completely occluded targets, perform trajectory prediction based on historical trajectories, retain the original ID until the occlusion is removed; after the occlusion is removed, quickly re-associate the target through a linkage threshold to restore the real-time trajectory and avoid trajectory breakage.

[0075] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments.

[0076] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any other combination thereof. When implemented using a software program, it can be implemented entirely or partially in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0077] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A dynamic target monitoring method based on artificial intelligence, characterized in that, include: Determine the target frame and update the detection threshold of the target frame using the historical detection library; The intelligent detection model uses a detection threshold to detect the target frame and obtain the detection result; Based on the detection results of the target frame, the feature similarity and positional similarity between the detected target in the target frame and the historical trajectory data are calculated, and a comprehensive matching score is obtained after weighted fusion. The overall matching score is compared with the preset tracking matching threshold to determine whether they belong to the same detection target. If they belong to the same detection target, the trajectory ID is bound to complete the target tracking.

2. The dynamic target monitoring method based on artificial intelligence according to claim 1, characterized in that, Update the detection threshold of the target frame using the historical detection library, including: Based on model consistency and environment consistency, historical detection data is matched for the target frame from the historical detection database; the historical detection database includes detection datasets for several image frames. Statistical analysis of historical detection data yields confidence indices and detection evaluation indices. The confidence indices include the confidence thresholds and median of the confidence thresholds for each detection dataset, while the detection evaluation indices include the average false negative rate and average false alarm rate for all detection datasets. The detection threshold of the target frame is determined based on the detection evaluation index, the preset detection evaluation benchmark, and the optimal confidence interval; wherein, the optimal confidence interval is determined by the confidence index, the detection evaluation index, and the preset detection evaluation benchmark.

3. The dynamic target monitoring method based on artificial intelligence according to claim 2, characterized in that, Based on model consistency and environment consistency, historical detection data is matched for the target frame from the historical detection database, including: Extract environmental data from the target frame and detect the intelligent detection model used for the target frame; Based on the environmental feature vector and the intelligent detection model, a detection dataset is matched from the historical detection database, and the matched detection dataset is used as the historical detection data of the target frame.

4. The dynamic target monitoring method based on artificial intelligence according to claim 2, characterized in that, The detection threshold for the target frame is determined based on detection evaluation indicators, preset detection evaluation benchmarks, and the optimal confidence interval, including: When all the detection evaluation indicators meet the preset detection evaluation benchmark, the median of the confidence index is used as the detection threshold of the target frame; otherwise, proceed to the next step of judgment. Determine whether all of the detection and evaluation indicators fail to meet the preset detection and evaluation benchmarks; Yes, the lower limit of the optimal confidence interval is used as the basic detection threshold, and the false alarm compensation value is calculated in combination with the average false alarm rate. The detection threshold is obtained by superimposing the basic detection threshold and the false alarm compensation value. No, if the average false alarm rate does not meet the requirements, the lower limit of the optimal confidence interval shall be used as the detection threshold; if the average false alarm rate does not meet the requirements, the upper limit of the optimal confidence interval shall be used as the detection threshold.

5. The dynamic target monitoring method based on artificial intelligence according to claim 4, characterized in that, The formula for calculating the false alarm compensation value is: False alarm compensation value = α × (average false alarm rate in the same environment - 3%) / (maximum historical false alarm rate - 3%); where α is 0.1 by default, and the maximum historical false alarm rate is the peak value of the false alarm rate of historical data in the same environment, with a preset upper limit of 10%.

6. The dynamic target monitoring method based on artificial intelligence according to claim 1, characterized in that, Adjust the tracking match threshold, including: The occlusion state of the detected target is quantized and calculated based on the detection results of the target frame, and the tracking base threshold is matched based on the occlusion state. The tracking correction coefficient is calculated based on the detection threshold of the target frame; the tracking correction coefficient is then used to correct the tracking base threshold to obtain the tracking matching threshold.

7. The dynamic target monitoring method based on artificial intelligence according to claim 6, characterized in that, The occlusion state of the detected target is quantized and calculated based on the detection results of the target frame, including: The cross-union ratio and the proportion of the visible area of ​​the detected target are calculated based on the detection results of the target frame. The occlusion status of the detected target is calculated by weighted fusion based on the intersection-union ratio between the targets and the proportion of the visible area of ​​the target; wherein, the occlusion status value ranges from [0,1], where 0 represents no occlusion and 1 represents complete occlusion.

8. The dynamic target monitoring method based on artificial intelligence according to claim 6, characterized in that, Based on the occlusion state, a tracking baseline threshold is matched, including: Read the threshold range for tracking and matching thresholds; Establish a linear mapping relationship between occlusion state and threshold range; obtain the tracking matching threshold from the linear mapping relationship based on the occlusion state; wherein the upper limit of the threshold range corresponds to the unoccluded state.

9. The dynamic target monitoring method based on artificial intelligence according to claim 6, characterized in that, The tracking correction coefficients are calculated based on the detection threshold of the target frame, including: Read the threshold range of the tracking correction coefficient; Establish a linear mapping relationship between the tracking correction coefficient range and the detection threshold range; where the midpoint of the detection threshold range corresponds to a tracking correction coefficient of 0. The tracking correction coefficient for the target frame detection threshold is obtained from the second linear mapping relationship; where the larger the detection threshold, the smaller the tracking correction coefficient.

10. An AI-based dynamic target monitoring system, used to execute the AI-based dynamic target monitoring method according to any one of claims 1 to 9, characterized in that, It includes a dynamic monitoring module and a data storage module. The data storage module is used to store the data required for the dynamic target monitoring process. Dynamic monitoring module: used to determine the target frame and update the detection threshold of the target frame using the historical detection library; The intelligent detection model uses a detection threshold to detect the target frame and obtain the detection result; Based on the detection results of the target frame, the feature similarity and positional similarity between the detected target in the target frame and the historical trajectory data are calculated, and a comprehensive matching score is obtained after weighted fusion. as well as, The overall matching score is compared with the preset tracking matching threshold to determine whether they belong to the same detection target. If they belong to the same detection target, the trajectory ID is bound to complete the target tracking.

Images