Cardio-pulmonary resuscitation video key frame time sequence labeling method based on semi-supervision

Abstract

The invention discloses a cardio-pulmonary resuscitation video key frame time sequence labeling method based on semi-supervision. The cardio-pulmonary resuscitation video key frame time sequence labeling method comprises the following steps: S1, preprocessing a video; s2, labeling the data; s3, a YOLO11 model is trained; s4, carrying out YOLO11 detection; s5, KHT matching is carried out; s6, calculating a final confidence coefficient; compared with the prior art, the method has the advantages that a main video key frame extraction technology, a YOLO11 model, a Kalman filtering algorithm and a Hungary algorithm are combined, video time sequence characteristics are combined, and a KHC filtering algorithm, an KHT matching recognition algorithm and a KHT-YOLO method framework are provided for semi-supervised automatic labeling of time sequence pictures of unknown cardio-pulmonary resuscitation video extraction key frames.

Description

Technical Field

[0001] The present invention combines the main video key frame extraction technology, YOLO11 model, Kalman filter algorithm and Hungarian algorithm, and combines the video timing characteristics to propose KHC filtering algorithm, KHT matching recognition algorithm, and KHT-YOLO method framework for semi-supervised automatic labeling of unknown cardiopulmonary resuscitation video extracted key frame timing pictures, specifically a cardiopulmonary resuscitation video key frame timing labeling method based on semi-supervision. Background Art

[0002] In recent years, medicine and artificial intelligence have been closely integrated. The core of medicine lies in training excellent doctors and using advanced equipment. Cardiac arrest is a life-threatening emergency. The incidence and outcome of out-of-hospital cardiac arrest vary greatly in different regions. Cardiopulmonary resuscitation is crucial in clinical and practical situations. The accuracy of its operation skills is related to the ability to handle emergencies.

[0003] In the field of artificial intelligence, OpenCV has received widespread attention since its birth in 1999 and has promoted the development of computer vision. It can perform a variety of image operations and be used for feature extraction and target detection. It can be explored for application in the medical field, such as training doctors to master accurate cardiopulmonary resuscitation skills.

[0004] The YOLO (You Only Look Once) algorithm has significant advantages in target detection, with the characteristics of fast detection speed, high accuracy and wide applicability. The latest version, YOLOv11, released on September 30, 2024, has gone a step further on the basis of YOLOv8, and optimized the model architecture by reducing the number of parameters, which has increased the detection speed by about 2%, and the mAP on datasets such as COCO by 2%. These improvements not only increase the applicability of the model, but also further improve its real-time performance in resource-constrained environments such as edge devices. As a commonly used optimal estimation algorithm, Kalman filtering iteratively updates the current state by combining the estimated value at the previous moment and the latest observation value, providing the optimal estimation under linear, unbiased, and minimum variance conditions. The Hungarian algorithm is based on the Hall theorem and is often used for bipartite graph matching. It minimizes the total cost by finding augmented paths to complete task allocation.

[0005] Compared with traditional training, hybrid cardiopulmonary resuscitation (CPR) refresher training shows unique advantages. The study by Chien et al. showed that after 6-month and 12-month CPR review courses, the hybrid training model provides new possibilities for improving CPR skills. Tang et al. (2024) explored the feasibility of using a three-dimensional motion capture system to detect and evaluate high-quality CPR, and proposed that the system can accurately record various actions and motion trajectories during CPR, so as to more accurately evaluate the quality of CPR, including key indicators such as compression position, depth and frequency. In addition, Li et al. (2024) proposed a deep learning-based method that uses deep learning algorithms to automatically analyze and evaluate CPR actions, and can accurately identify and judge the correctness of operations such as compression position, depth, frequency and artificial respiration. Through a large amount of training data, the deep learning model can learn CPR action patterns in different situations and accurately evaluate and feedback new CPR actions.

[0006] Since traditional teaching methods often rely on face-to-face instruction and practice, as well as the use of sensors to detect whether compressions are in place and whether the number of compressions is up to standard, including whether artificial respiration is performed during cardiopulmonary resuscitation, these methods are limited in resources and time. Therefore, using image recognition technology to evaluate and improve the practical operation ability of medical students has become an emerging research field.

[0007] The prior art has the following problems:

[0008] Lack of data sets: There is currently no public CPR data set, and it is impossible to use recorded or real-time videos to detect whether medical students or related trainers meet key indicators such as whether the compression position is accurate, the number, depth, frequency of compressions, and whether artificial respiration is performed during CPR training.

[0009] High labor costs: The current labeling method relies almost entirely on manual labor to label a large amount of required data, which makes it impossible to save more time to conduct corresponding experiments.

[0010] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention

[0011] The technical problem to be solved by the present invention is to overcome the above technical defects and provide a semi-supervised cardiopulmonary resuscitation video key frame timing labeling method.

[0012] In order to solve the above problems, the technical solution of the present invention is a semi-supervised cardiopulmonary resuscitation video key frame timing labeling method, which comprises the following steps:

[0013] S1. Video preprocessing: converting video formats and adjusting resolution to prepare for model processing;

[0014] S2, annotated data, manually annotating key actions and feature points in selected frames for training models;

[0015] S3. Train the YOLO11 model and apply basic data enhancement techniques to improve the generalization ability of the model;

[0016] S4, YOLO11 detection, using the YOLO11 model to detect key actions and feature points in video frames;

[0017] S5, KHT matching, for each frame, use KHT matching to predict the key points of the next frame, use the intersection over union (IoU) to verify the match, and iteratively update the prediction results;

[0018] S6. Calculate the final confidence, combine the results of YOLO and KHT with equal weights, and determine the final detection confidence.

[0019] Furthermore, the video preprocessed in S1 is a CPR training assessment video.

[0020] Furthermore, the data enhancement technology in S3 includes blurring, grayscale conversion, brightness adjustment and compression technology.

[0021] Furthermore, the video data preprocessing in S1 is to read and process the video without encoding or decoding the video through OpenCV, and to create a CPR assessment video key frame parser using OpenCV to extract key frame images from the CPR assessment video.

[0022] Furthermore, the data annotation in S2 is integrated by using a video parser to analyze the key frame images extracted from the CPR video, and the data is manually annotated, and there are five labels: Dummy, D-head, D-chest, DH-occlusion, and DC-occlusion. The meanings of the labels are Dummy: CPR training aid; D-head: the head of the training aid without artificial respiration; D-chest: the chest of the training aid without cardiopulmonary compression; DH-occlusion: the head of the training aid for artificial respiration;

[0023] DC-occlusion: A chest training aid for doing cardiopulmonary compressions.

[0024] Furthermore, the method framework of KHT-YOLO in S5 is as follows: extract key frames from the CPR assessment video, manually annotate some frames, and divide the annotated data set into a training set and a validation set, use YOLO11n for recognition, and combine Kalman filtering and Hungarian matching algorithm to perform semi-supervised automatic annotation on unannotated frames;

[0025] When the recognition results are confusing, the IOU edge matching is used to determine whether the positions are similar. If they are similar and the recognition results are consistent, the target with the smallest recognition part in KHT recognition is retained; if they are inconsistent, the recognition situation in the auxiliary perspective is checked and its weight is set to 10 to correct the confusing recognition; if other angles cannot be distinguished, the confidence of KHT (weight 5) and YOLO11n (weight 5) is compared, and the one with the highest confidence is selected as the final recognition result.

[0026] Furthermore, the confidence index introduced by the Kalman recursive filter is a KHC filter, which is used to estimate the next state (such as position and speed) of the dynamic prediction target based on the label observations detected by YOLO and the state equation of the system, and updates the prediction according to the new observation data.

[0027] Further, the KHC filtering algorithm comprises the following steps:

[0028] S1. Prediction stage: Prediction status: Error covariance formula: P t - =FP t-1 F T +Q;

[0029] S2, Update phase: KHC filter gain coefficient: K t =P t - H T (HP t - H T +R) -1 ;

[0030] Best estimate = predicted value + (actual value - predicted value):

[0031] Error covariance update formula: P t =(IK t H)P t - .

[0032] S3, KHT matching algorithm. The KHT matching algorithm combines the Hungarian algorithm with the KHC filter prediction and recognition, introduces confidence, and performs prediction matching according to the weight when confusion occurs in the KHT prediction and recognition, and performs secondary confirmation on the target, thereby improving the accuracy of target recognition.

[0033] Furthermore, the confidence mechanism of YOLO in S6 includes: Objectness Score and ClassConfidence, which use softmax or sigmoid function to calculate the probability of each category and jointly determine the final confidence score (Confidence Score) of a detection box, C = P (object) × P (class | object)

[0034] , where Objectness Score is used to determine whether there is an object in the detection box, and Class Confidence is used to classify which category the object in the detection box belongs to;

[0035] The KHC filter confidence uses partial information of the covariance matrix to estimate the overall uncertainty. Based on this uncertainty, the confidence can be expressed as the inverse of the uncertainty,

[0036] P ii : The first four diagonal elements in the covariance matrix related to position and size, which represent the state error of the tracked target; Uncertainty is the total uncertainty of the state error; Confidence is the confidence of the KHC filter calculated based on the uncertainty.

[0037] The advantages of the present invention compared with the prior art are:

[0038] 1. The present invention proposes a KHC filtering algorithm: utilizing the timing characteristics of video, based on the Kalman filtering algorithm, creating a KHC filter combined with a KHT matching algorithm for more efficient target tracking and prediction.

[0039] 2. The present invention proposes a KHT matching recognition algorithm: combining the YOLO11 model and the KHC filtering algorithm, using a small amount of training sets for training to automatically annotate unknown cardiopulmonary resuscitation video images, thereby improving the efficiency of semi-supervised learning.

[0040] 3. The present invention proposes a KHT-YOLO method framework: YOLO11 is used in combination with KHC filter and KHT matching recognition algorithm to train and match a small amount of artificial data, and a large number of unknown CPR video key frame samples are identified for semi-supervised automatic labeling.

[0041] 4. The present invention proposes a CPR training dataset: by extracting key frames of 120 cardiopulmonary resuscitation videos, a labeled original dataset is created, and it is divided into 240:1916 ratios and manually marked to identify targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a timing annotation diagram of the CPR key frames of the present invention.

[0043] Figure 2 This is a framework diagram of the KHT-YOLO method of the present invention.

[0044] Figure 3 It is the target graph predicted by the KHC filter of the present invention.

[0045] Figure 4 It is the KHT matching algorithm diagram.

[0046] Figure 5 It is a comparison chart of the detection performance of the YLv11 vs YKH vs KHT-Y combination method of the present invention. DETAILED DESCRIPTION

[0047] In order to make the contents of the present invention more clearly understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0048] Based on the semi-supervised cardiopulmonary resuscitation video key frame temporal labeling method framework (KHT-YOLO), a large amount of unknown data can be accurately identified and labeled by labeling a small amount of data. The main contents of the framework are:

[0049] Input: CPR training assessment video

[0050] Output: Keyframe timing annotation

[0051] Step 1: Preprocess the video, convert the video format and adjust the resolution to prepare for model processing.

[0052] Step 2: Label the data and manually label the key actions and feature points in the selected frames for training the model.

[0053] Step 3: Train the YOLO11 model and apply basic data augmentation techniques (such as blurring, grayscale conversion, brightness adjustment, and compression) to improve the generalization ability of the model.

[0054] Step 4: YOLO11 Detection Use the YOLO11 model to detect key actions and feature points in the video frames.

[0055] Step 5: KHT matching, for each frame, use KHT matching to predict the key points of the next frame, use the intersection over union (IoU) to verify the match, and iteratively update the prediction results.

[0056] Step 6: Calculate the final confidence, combine the results of YOLO and KHT with equal weights to determine the final detection confidence.

[0057] Step 7: Output: Generate and output key frame timing annotations based on the final detection results.

[0058] Preparation includes:

[0059] 1. Key frame extraction. OpenCV does not perform encoding and decoding on the video. It can quickly read and process the video. A CPR assessment video key frame parser was created using OpenCV to extract key frame images from the CPR assessment video and perform the first preprocessing of the video data.

[0060] 2. Manually label key frame time series images. Since there is no public CPR dataset, the key frame images extracted from the CPR video are integrated using a video parser, and the data is manually labeled with five labels, as shown in the following table:

[0061]

[0062] 3. Framework method implementation

[0063] 3.1, KHT-YOLO method framework, based on the semi-supervised cardiopulmonary resuscitation video key frame temporal annotation method framework (KHT-YOLO), such as Figure 2 As shown in the figure, this method extracts key frames from the CPR assessment video, manually annotates some frames, and divides the annotated data set into a training set and a validation set, and uses YOLO11n for recognition. Combining Kalman filtering and Hungarian matching algorithm, semi-supervised automatic annotation is performed on unlabeled frames.

[0064] When the recognition results are confusing, the IOU edge matching is used to determine whether the positions are similar. If they are similar and the recognition results are consistent, the target with the smallest recognition part in KHT recognition is retained; if they are inconsistent, the recognition situation in the auxiliary perspective is checked and its weight is set to 10 to correct the confusing recognition. If other angles cannot be distinguished, the confidence of KHT (weight 5) and YOLO11n (weight 5) is compared, and the highest confidence is selected as the final recognition result.

[0065] Since the experiment uses raw video data, which does not contain any key points and feature markers, this data cannot be directly used to detect whether CPR training meets the standards. In order to solve this problem, a semi-supervised CPR video key frame temporal labeling method framework is proposed, named KHT-YOLO. This method is first used to train some labeled CPR training assessment videos, and then target detection and data labeling are performed on unknown videos and videos in complex environments, so as to prepare for the subsequent detection of whether CPR training assessment is standard.

[0066] 3.2. KHC filtering algorithm, based on the Kalman recursive filter, introduces a confidence indicator called KHC filter. It is used to estimate the next state (such as position and speed) of the dynamic prediction target based on the label observations detected by YOLO and the state equation of the system, and updates the prediction based on the new observation data.

[0067] Initialize the KHC filter: set parameters and convert the detected bounding box into a state representation, assigning a unique ID and confidence to each tracker. The initial state of the target is determined by the YOLO11n detection box, including the center position, confidence, width, height, and aspect ratio.

[0068] Prediction: In each frame, the KHC filter predicts the state of the current frame based on the state of the previous frame.

[0069] Update: Update the state of the KHC filter with new boxes detected by YOLO11n.

[0070] In the KHC filtering algorithm, the state estimation and prediction of the next time series image of CPR is divided into a prediction stage and an update stage. The prediction stage is divided into two parts: state prediction and error. The definition of each parameter is closely related to the state estimation and prediction process. The following table describes the parameters in the following stages in detail.

[0071]

[0072]

[0073] 3.2.1. Prediction stage: state prediction:

[0074] Error covariance formula: P t - =FP t-1 F T +Q

[0075] 3.2.2. Update phase: KHC filter gain coefficient: K t =P t - H T(HP t - H T +R) -1

[0076] Best estimate = predicted value + (actual value - predicted value):

[0077] Error covariance update formula: P t =(IK t H)P t -

[0078] 3.3, KHT matching algorithm,KHT matching algorithm is based on KHC filter prediction and recognition combined with Hungarian algorithm, introduces confidence, and performs prediction and matching according to weights when confusion occurs in KHT prediction and recognition, and performs secondary confirmation on the target, thereby improving the accuracy of target recognition.

[0079] 3.3.1,KHT task allocation and target tracking,The core idea of ​​the KHT matching algorithm is that the,Hungarian algorithm determines the optimal allocation scheme through a series of steps,of KHC filters and confidence weights without increasing the cost.

[0080]

[0081]

[0082] The KHT matching algorithm is an optimization based on the Hungarian algorithm, which can best match the detection box with the tracker at the lowest cost and calculate the confidence formula. IOU (intersection over union) is used as the metric.

[0083]

[0084] A: Detection frame area

[0085] B: Prediction box area

[0086]

[0087] 3.3.2 KHT weight selection

[0088] After YOLO11n recognition and the filter and matching algorithm, the recognition confidence of YOLO11n for the label and the confidence after the filter and matching algorithm are extracted.

[0089] YOLO confidence, the confidence mechanism of YOLO (You Only Look Once) mainly includes two parts: Objectness Score and Class Confidence, which uses softmax or sigmoid function to calculate the probability of each category and jointly determine the final confidence (Confidence Score) of a detection box.

[0090] C=P(object)×P(class|object)

[0091] Objectness Score is used to determine whether there is an object in the detection box.

[0092] Class Confidence is used to classify the category to which the object in the detection box belongs.

[0093] The KHC filter confidence uses partial information of the covariance matrix (the relevant part of position and size, i.e. the first 4 elements) to estimate the overall uncertainty. Based on this uncertainty, the confidence can be expressed as the inverse of the uncertainty. Higher uncertainty means lower confidence.

[0094]

[0095] P ii : The first 4 diagonal elements in the covariance matrix related to position and size, which represent the state error of the tracked target.

[0096] Uncertainty: The total uncertainty of the state error.

[0097] Confidence: The confidence of the KHC filter calculated based on the uncertainty.

[0098] 4.2 Results and Analysis

[0099] The experiment used 240 images as training sets and 1916 images as validation sets. The validation set is about 8 times the size of the training set. Finally, the prediction and recognition of YOLO11n and the combination of YOLO11n with Kalman filter algorithm and Hungarian matching algorithm were compared with the results of this method and manual labeling.

[0100] The YOLO11n combined with the original Kalman filter algorithm and the Hungarian matching algorithm (YKH) are compared with five manually labeled labels to show the detection performance of this combined method.

[0101] Note: In this article, "Method" is referred to as "Mtd."; 'YOLO11n' is referred to as 'YLv11'; 'YOLO11n+original Kalman filter algorithm+Hungarian matching algorithm' is referred to as 'YKH'; the 'KHT-YOLO' proposed in this article is referred to as 'KHT-Y'.

[0102] The following table compares the five labels in an environment where the training set: validation set ratio is 240:1916 using YOLO11n, YOLO11n combined with the original Kalman filter algorithm and the Hungarian matching algorithm, and the KHT-YOLO proposed in this article.

[0103]

[0104] Experimental results show that the proposed KHT-YOLO (KHT-Y) outperforms other methods in all evaluation indicators and demonstrates excellent performance. In the "Dummy" task, KHT-Y's precision (Pre) reaches 0.999, the recall rate (Rec) is the same as YOLO11n (YLv11) (1.000), the intersection over union (IoU) is 0.882, and the mAP50 is 0.887, which fully demonstrates the model's ability to balance precision and recall. In contrast, YLv11 and YKH perform slightly worse in terms of intersection over union, which are 0.890 and 0.865, respectively.

[0105] In the “D-head” task, KHT-Y achieves an accuracy of 0.988, significantly higher than YLv11 (0.960) and YKH (0.967). Its F1 score (0.987) and recall (0.985) further verify the robustness of the model in head detection. Its intersection-over-union ratio is 0.799, slightly better than YKH’s 0.793.

[0106] In the "D-chest" category, KHT-Y's accuracy is improved to 0.960, far exceeding YLv11 (0.903) and YKH (0.910). Although the recall rate and F1 score in this category are slightly lower than those in other categories, the overall intersection-over-union ratio (0.699) is basically consistent with other methods, showing the effectiveness of KHT-Y in this task.

[0107] In the complex occlusion "DC-occlusion" and "DH-occlusion" tasks, KHT-Y still performs well. For the "DC-occlusion" scene, KHT-Y's intersection-over-union ratio is 0.691, showing its advantage in dealing with occlusion problems, while in the "DH-occlusion" task, its accuracy reaches 0.963 and the intersection-over-union ratio is 0.748, further demonstrating KHT-Y's robustness in dealing with complex scenes.

[0108] Based on the comprehensive average results, KHT-Y performs best in all tasks, with an average precision (Pre) of 0.973, a recall rate of 0.962, an intersection-over-union ratio of 0.764, and a mAP50 of 0.887, which are significantly higher than YLv11 (average precision 0.951, recall rate 0.964, intersection-over-union ratio 0.769, mAP50 0.861) and YKH (average precision 0.946, recall rate 0.961, intersection-over-union ratio 0.765, mAP50 0.841). This fully demonstrates that KHT-Y has good positioning accuracy and detectability in target detection tasks, and has significant advantages in overall robustness.

[0109] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A semi-supervised cardiopulmonary resuscitation video key frame temporal labeling method, characterized by: The following steps are involved: S1. Video preprocessing: converting video formats and adjusting resolution to prepare for model processing; S2, annotated data, manually annotating key actions and feature points in selected frames for training models; S3. Train the YOLO11 model and apply basic data enhancement techniques to improve the generalization ability of the model; S4, YOLO11 detection, using the YOLO11 model to detect key actions and feature points in video frames; S5, KHT matching, for each frame, use KHT matching to predict the key points of the next frame, use the intersection over union (IoU) to verify the match, and iteratively update the prediction results; S6. Calculate the final confidence, combine the results of YOLO and KHT with equal weights, and determine the final detection confidence.

2. The semi-supervised cardiopulmonary resuscitation video key frame temporal labeling method according to claim 1 is characterized in that: The video preprocessed in S1 is a CPR training assessment video.

3. The semi-supervised cardiopulmonary resuscitation video key frame temporal labeling method according to claim 1, characterized in that: The data enhancement techniques in S3 include blurring, grayscale conversion, brightness adjustment and compression techniques.

4. The semi-supervised cardiopulmonary resuscitation video key frame timing labeling method according to claim 1, characterized in that: The video data preprocessing in S1 is to read and process the video without encoding or decoding the video through OpenCV, and to use OpenCV to create a CPR assessment video key frame parser to extract key frame images from the CPR assessment video.

5. The semi-supervised cardiopulmonary resuscitation video key frame timing labeling method according to claim 1, characterized in that: The data annotation in S2 is performed by integrating the key frame images extracted from the CPR video using a video parser, and manually annotating the data, and there are five labels: Dummy, D-head, D-chest, DH-occlusion, and DC-occlusion. The meanings of the labels are Dummy: CPR training aid; D-head: the head of the training aid without artificial respiration; D-chest: the chest of the training aid without cardiopulmonary compression; DH-occlusion: the head of the training aid for artificial respiration; DC-occlusion: the chest of the training aid for cardiopulmonary compression.

6. The semi-supervised cardiopulmonary resuscitation video key frame timing labeling method according to claim 1, characterized in that: The framework of the KHT-YOLO method in S5 is as follows: extract key frames from the CPR assessment video, manually annotate some frames, and divide the annotated data set into a training set and a validation set, use YOLO11 n for recognition, and combine Kalman filtering and Hungarian matching algorithm to perform semi-supervised automatic annotation on unannotated frames; When the recognition results are confusing, the IOU edge matching is used to determine whether the positions are similar. If they are similar and the recognition results are consistent, the target with the smallest recognition part in KHT recognition is retained; if they are inconsistent, the recognition situation in the auxiliary perspective is checked and its weight is set to 10 to correct the confusing recognition; if other angles cannot be distinguished, the confidence of KHT (weight 5) and YOLO11 n (weight 5) is compared, and the highest confidence is selected as the final recognition result.

7. The semi-supervised cardiopulmonary resuscitation video key frame timing labeling method according to claim 6 is characterized in that: The confidence index introduced by the Kalman recursive filter is the KHC filter, which is used to estimate the next state (such as position and speed) of the dynamic prediction target based on the label observations detected by YOLO and the state equation of the system, and update the prediction according to the new observation data.

8. The semi-supervised cardiopulmonary resuscitation video key frame timing labeling method according to claim 7, characterized in that: The KHC filtering algorithm includes the following steps: S1. Prediction stage: Prediction status: Error covariance formula: P t - =FP t-1 F T +Q; S2, update stage: KHC filter gain coefficient: K t =P t - H T (HP t - H T +R) -1 ; Best estimate = predicted value + (actual value - predicted value): Error covariance update formula: P t =(IK t H)P t - . S3, KHT matching algorithm. The KHT matching algorithm combines the Hungarian algorithm with the KHC filter prediction and recognition, introduces confidence, and performs prediction matching according to the weight when confusion occurs in the KHT prediction and recognition, and performs secondary confirmation on the target, thereby improving the accuracy of target recognition.

9. The semi-supervised cardiopulmonary resuscitation video key frame timing labeling method according to claim 1, characterized in that: The confidence mechanism of YOLO in S6 includes: Objectness Score and Class Confidence, which use softmax or sigmoid function to calculate the probability of each category and jointly determine the final confidence score (Confidence Score) of a detection box, C = P (object) × P (class | object), The Objectness Score is used to determine whether there is an object in the detection box, and the Class Confidence is used to classify which category the object in the detection box belongs to; The KHC filter confidence uses partial information of the covariance matrix to estimate the overall uncertainty. Based on this uncertainty, the confidence can be expressed as the inverse of the uncertainty, P ii : The first four diagonal elements in the covariance matrix related to position and size, which represent the state error of the tracked target; Uncertainty is the total uncertainty of the state error; Confidence is the confidence of the KHC filter calculated based on the uncertainty.

Images