Non-inductive identity recognition and real-time tracking method based on personnel in video

Abstract

The invention discloses a non-inductive identity recognition and real-time tracking method based on personnel in a video. The method comprises the following steps: S1, acquiring video stream data and extracting a key frame image; s2, key point coordinates are extracted, and a time sequence skeleton sequence is constructed; s3, decomposing gait, action and attitude features to generate standardized features; s4, inputting the standardized features into an improved multi-scale image convolutional neural network, extracting identity representation features, and generating a target identity code; s5, matching identity codes based on similarity measurement, calculating a consistency score, and establishing an identity tracking trajectory; s6, tracking is carried out in combination with identity codes and consistency scores, identity mapping is dynamically updated, and non-inductive recognition is achieved; and S7, updating a track in real time, adjusting a tracking state, and ensuring identity stability. According to the method, the improved graph convolutional network and the consistency measurement technology are combined, the identity recognition precision and tracking stability are improved, and cross-scene non-inductive identity recognition is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and biometric recognition, and particularly to a method for contactless identity recognition and real-time tracking of personnel in a video. Background Art

[0002] With the development of computer vision and artificial intelligence technologies, video-based identity recognition and tracking technologies have been widely applied in many fields such as smart cities, intelligent monitoring, public security, and unmanned retail. Traditional identity recognition technologies mainly rely on active biometrics such as face recognition, iris recognition, and fingerprint recognition. However, these methods have certain limitations in practical applications. For example, the recognition accuracy of face recognition will drop significantly under conditions such as light changes, angle offsets, and occlusion interference; iris recognition and fingerprint recognition require the active cooperation of the target individual and cannot perform identity recognition at a long distance or without contact. Therefore, how to perform identity recognition in a contactless, long-distance, and anti-interference manner has become an urgent problem to be solved in this field.

[0003] In recent years, gait recognition, as a contactless identity recognition method, has received wide attention due to its advantages such as long distance, non-contact, and anti-occlusion. Gait is a unique biometric behavior feature of an individual, and it can maintain a high recognition accuracy even in long-distance and low-resolution environments. However, existing gait recognition methods mainly rely on static gait sequences for identity discrimination and are difficult to adapt to complex dynamic scenarios, such as camera angle changes, pedestrian speed changes, and irregular movements. In addition, most gait recognition models only focus on gait features and ignore other human behavior features, such as action patterns and posture changes, resulting in poor recognition robustness in a changing environment. For example, existing gait recognition methods are difficult to handle gait deformations caused by environmental factors or individual state changes, such as pedestrians carrying backpacks, hand movements changing, and clothing changes, and these factors will all affect the accuracy of gait recognition.

[0004] For the identity tracking task, multi-object tracking (MOT) technology has been widely applied in fields such as pedestrian tracking and security monitoring. Traditional MOT methods mainly rely on the detection boxes and motion trajectories of targets for matching, such as SORT, DeepSORT, etc. These methods usually use Kalman filtering for motion prediction during target tracking and combine appearance features for target association. However, such methods are difficult to solve problems such as identity loss or identity drift caused by target occlusion, overlap, and posture changes. For example, in a crowded scene, multiple targets may cross-move in a short period of time, making it difficult for traditional MOT methods to maintain the continuity of identity. In addition, MOT methods usually do not combine the biometric behavior features of target individuals, such as gait, posture, and action patterns, so there are certain limitations in long-term identity maintenance and cross-scene adaptability.

[0005] In the video-based identity recognition and tracking tasks, the application of deep learning technology has enhanced the model's feature extraction and recognition capabilities. Models such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have achieved significant breakthroughs in tasks such as object detection, behavior recognition, and temporal modeling. However, the existing deep learning methods still have the following problems in the process of identity recognition and tracking: (1) Traditional gait recognition models usually use single gait contour or skeleton features for identity discrimination, and fail to fully utilize multi-modal information such as action features and pose features, resulting in a decline in recognition accuracy in complex environments; (2) In the process of human skeleton modeling by existing graph convolutional networks (GCNs), only local spatial relationships are concerned, while the aggregation of global information is ignored, making it difficult to adapt to changes in different camera angles and walking styles; (3) Existing object tracking methods fail to combine gait and temporal information for identity consistency constraints during the identity maintenance process, resulting in identity drift problems during long-term tracking.

[0006] Therefore, how to provide a method for non-intrusive identity recognition and real-time tracking of people in videos is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose a method for non-intrusive identity recognition and real-time tracking of people in videos. The present invention makes full use of multi-dimensional feature analysis methods such as gait, action, and pose, combines an improved multi-scale graph convolutional neural network for identity representation, and adopts an identity consistency score calculation and dynamic identity mapping strategy for real-time tracking. The present invention details algorithms for adaptive multi-dimensional feature decomposition, gait trajectory optimization, cross-frame identity matching, and dynamic identity tracking, and has the advantages of non-intrusive identity recognition, high robustness, high stability, and strong cross-scene adaptability.

[0008] The method for non-intrusive identity recognition and real-time tracking of people in videos according to an embodiment of the present invention includes the following steps: S1. Obtain video stream data and perform preprocessing to extract key frame images containing human targets; S2. Extract the coordinates of human key points from the key frame images and construct a temporal skeleton sequence; S3. Perform adaptive multi-dimensional feature decomposition on the temporal skeleton sequence, extract gait features, action features, and pose features respectively, and perform normalization processing to generate standardized features; [[ID=ID=19]] S4. Input the standardized features into an improved multi-scale graph convolutional neural network, combine the residual connection and non-local aggregation mechanism, extract identity representation features, and generate target identity codes; S5. Use similarity measurement technology to perform cross-frame matching on the target identity codes, calculate the identity consistency score, and establish an identity tracking trajectory; S6. Combine the target identity encoding and identity consistency score in the video stream for identity tracking, dynamically update the target identity mapping, and achieve seamless identity recognition of the target individual; S7. Real-time update the movement trajectory of the target individual in the video, adjust the tracking status in combination with the identity consistency score, and ensure the continuous stability of the identity of the target individual throughout the video sequence.

[0009] Optionally, the S2 specifically includes: S21. Perform human body detection on the key frame image, extract the region containing the human body target, and perform scale normalization processing; S22. Perform skeleton key point detection on the normalized human body target region to obtain the skeleton key point coordinate set , where represents the two-dimensional coordinates of the i-th skeleton key point, respectively represent the horizontal and vertical coordinates of the skeleton key point, and N is the number of skeleton key points; S23. According to the skeleton key point coordinate set, calculate the Euclidean distance matrix between adjacent skeleton key points: ; Among them, is the Euclidean distance between the i-th and j-th skeleton key points, which is used to represent the skeleton structure relationship; S24. Construct a temporal skeleton sequence, set the time window length to Ι, and combine the skeleton key point coordinates of consecutive Ι frames to form a temporal feature matrix S, where each row corresponds to the coordinates of all skeleton key points of one frame; S25. Perform time interpolation on the temporal feature matrix S, and use linear interpolation to complete the coordinates of the missing frame skeleton key points to ensure data integrity; S26. Perform normalization processing on the interpolated temporal feature matrix S to normalize each skeleton key point coordinate to the range of [-1, 1] to obtain a normalized temporal skeleton sequence.

[0010] Optionally, the S3 specifically includes: S31. Perform dimensionality splitting on the normalized temporal skeleton sequence, and respectively extract the skeleton spatial feature matrix and the skeleton motion feature matrix , where represents the spatial coordinate information of the skeleton key point, represents the velocity change information of the skeleton key point between adjacent frames; S32. Perform local feature decomposition on the skeleton spatial feature matrix to calculate the relative relationship matrix between skeleton key points: ; Among them, is the relative relationship value between the i-th and j-th skeleton key points, is the Euclidean distance between the i-th and j-th skeleton key points, is the Euclidean distance between the i-th and k-th skeleton key points, and N is the number of skeleton key points; S33. Perform adaptive time series decomposition on the skeleton motion feature matrix to construct a multi-scale time feature representation: ; Among them, T is the multi-scale time feature representation, m is the number of time scale features, is the time feature representation of the m-th scale, is the decomposition weight parameter, is the velocity change information between adjacent frames of the skeleton key points at time t; S34. Construct a gait feature matrix B using gait information, and extract gait features through a non-linear mapping where f is the non-linear mapping function; S35. Construct a pose feature matrix P using pose information, and extract pose features through a non-linear mapping where g is the non-linear mapping function and R is the relative relationship matrix; S36. Construct an action feature matrix O using action information, and extract action features through a non-linear mapping where h is the non-linear mapping function; S37. Perform feature fusion on the gait feature matrix B, the pose feature matrix P, and the action feature matrix O to generate a normalized feature representation , where is the fusion weight parameter, and perform normalization processing.

[0011] Optionally, the specific steps of S4 include: S41. Perform graph modeling on the normalized feature representation F to construct an adjacency matrix and a node feature matrix X, where represents the association weight between the i-th and j-th skeleton key points, and respectively represent the feature vectors of the i-th and j-th skeleton key points, is the scale factor, is the natural exponential function with the real number e as the base; S42. Use an improved multi-scale graph convolutional neural network to extract features from the input features, introduce an improved adaptive scale weight allocation mechanism to dynamically adjust the adjacency matrices of different scales, and calculate the multi-scale feature representation of the key points: ; Among them, is the adaptive scale weight, K is the number of scales, is the degree matrix of the corresponding adjacency matrix and is defined as , is the feature representation after the -th layer of graph convolution, is the feature representation after the -th layer of graph convolution, is the trainable weight matrix, is the non-linear activation function, and normalizes the scale weight to ensure the effective fusion of information at different scales; S43. After multi-scale graph convolution calculation, introduce a cross-scale feature aggregation mechanism to optimize the information flow of features at different scales through residual connection, and calculate the global fusion feature: ; Among them, is the global fusion feature, is the number of scales, is the adaptive scale weight, is the feature representation after the -th layer of graph convolution for the k-th scale, is the node feature matrix of the initial input, and uses cross-scale residual connection to enhance feature stability; S44. Use the non-local aggregation mechanism to optimize the global fusion feature, calculate the long-distance dependence relationship between skeleton key points globally, and construct the non-local relationship matrix G: ; ; Among them, is the natural exponential function with the real number e as the base, K is the number of scales, and are different mapping functions used to calculate the similarity of key point features, and are the feature representations of the i-th and j-th skeleton key points after multi-scale graph convolution respectively, represents the non-local association weight between key point i and key point j, is the global fusion feature, is the feature representation after non-local enhancement, combines local and global information, and optimizes the identity representation feature; S45. The optimized feature representation Perform a linear transformation, map it to the identity coding space, generate the target identity coding, and perform normalization processing on the target identity coding so that it is distributed on the unit sphere to ensure the stability and distinctiveness of the identity representation features.

[0012] Optionally, the S5 specifically includes: S51. Perform cross-frame matching on the target identity coding to construct an identity similarity matrix based on feature association: ; Wherein, is the natural exponential function with the real number e as the base, K is the number of scales, and are the target identity codings at time step i and time step j respectively, is the identity matching probability between time step i and time step j, represents the norm operation; S52. Based on the identity similarity matrix, calculate the identity consistency score by combining the non-local relationship matrix: ; Wherein, is the identity consistency score at time step i, k is the number of scales, is the non-local association weight between key point i and key point k, is the identity matching probability between time step i and time step k, is the gait trajectory constraint parameter to ensure that the identity matching not only depends on the appearance features but also combines the gait time series information; S53. Use the identity consistency score to optimize the target identity matching and construct an identity tracking trajectory: ; ; Wherein, is the identity tracking trajectory, is the identity tracking state at time step i, K is the number of scales, is the target identity coding at time step i, is the target identity coding at time step k, is the identity matching probability between time step i and time step k, is the identity consistency score at time step i, representing the identity stability weighting factor to improve the continuity of identity tracking, is the identity historical matching weight parameter to ensure that the tracking trajectory remains stable between frames and reduce the identity drift problem.

[0013] Optionally, the S6 specifically includes: S61. Calculate the identity matching score matrix between the current frame and the historical frames based on the target identity encoding and the identity consistency score: ; where, represents the identity matching score between time step i and time step j, is the identity matching probability between time step i and time step j, represents the non-local association weight between key point i and key point j, is the identity consistency score of time step i, represents the Euclidean distance of the target identity encoding, serving as a constraint on identity change, is the different feature weighting coefficient; S62. Calculate the identity mapping probability matrix U between the current frame and the historical frames based on the identity matching score matrix: ; where, is the identity mapping probability that the current time step i is associated with the historical time step j, K is the number of scales, represents the identity matching score between time step i and time step j, represents the identity matching score between time step i and time step k, is the natural exponential function with the real number e as the base, and the formula adopts the softmax normalization operation to make the identity mapping change smoothly between frames; Perform regularization constraints on the identity mapping matrix to ensure that the identity mapping does not drift, and calculate the regularized identity mapping matrix: ; where, is the regularized identity mapping matrix, U is the identity mapping probability matrix, is the regularization parameter, and I is the identity matrix; S63. Combine the identity mapping matrix and the identity tracking trajectory to update and calculate the identity tracking state of time step : ; where, is the updated and calculated identity tracking state of time step i, is the regularized identity mapping matrix that the previous time step i is associated with the historical time step k, is the historical identity tracking state of time step k, Calculate the identity deviation as the identity correction term, is the identity correction coefficient, enabling the identity tracking to be dynamically adjusted to ensure cross-frame identity consistency and achieve seamless identity recognition of target individuals.

[0014] The beneficial effects of the present invention are as follows: Firstly, by deeply analyzing multi-dimensional features such as gait, movement, and posture, the present invention can extract stable and individual feature information from the video stream, realizing non-intrusive identification of personnel identities in complex environments. Different from traditional biometric identification technologies based on face, fingerprint, etc., the present invention does not rely on direct contact or device cooperation, avoiding the limitations of traditional identity identification technologies. Especially in dynamic monitoring and unsupervised environments, it can provide higher identification accuracy and stability.

[0015] Secondly, the present invention adopts an improved multi-scale graph convolutional neural network, combining residual connection and non-local aggregation mechanism, enabling gait features to be more precisely extracted in the multi-scale spatial domain. This innovative improvement greatly enhances the model's ability to extract identity representation features in complex scenarios and can still maintain high-precision identity identification from different perspectives and in different environments. In addition, by adaptively decomposing the standardized features, the present invention can effectively eliminate noise and adapt to different motion states, further enhancing the system's performance in practical applications.

[0016] Finally, through the calculation of identity consistency scores and dynamic identity mapping strategies, the present invention successfully realizes real-time tracking and non-intrusive identification of target individuals. Through cross-frame identity matching and target trajectory tracking, the identity of the target individual can be continuously and stably tracked in dynamic scenarios. This method not only improves the real-time performance and accuracy of identity identification but also ensures identity consistency in complex backgrounds through dynamic updates of the target identity, providing a strong guarantee for long-term monitoring in practical applications.

[0017] In summary, the present invention has high robustness, precision, and real-time performance, and can be widely applied in the fields of public security, intelligent security, intelligent transportation, etc. in practical applications, having significant technical advantages and broad application prospects. Brief Description of the Drawings

[0018] The drawings are used to provide further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 It is a flowchart of the non-intrusive identity identification and real-time tracking method for personnel in the video proposed by the present invention. Detailed Embodiments

[0019] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0020] Reference Figure 1, A method for contactless identity recognition and real-time tracking of personnel in a video, including the following steps: S1. Obtain video stream data and perform preprocessing to extract key frame images containing human targets; S2. Extract the coordinates of human key points from the key frame images and construct a temporal skeleton sequence; S3. Perform adaptive multi-dimensional feature decomposition on the temporal skeleton sequence, extract gait features, action features, and pose features respectively, and perform normalization processing to generate standardized features; S4. Input the standardized features into an improved multi-scale graph convolutional neural network, combine the residual connection and non-local aggregation mechanism, extract identity representation features, and generate target identity codes; S5. Use similarity measurement technology to perform cross-frame matching on the target identity codes, calculate the identity consistency score, and establish an identity tracking trajectory; S6. Combine the target identity code and the identity consistency score in the video stream to perform identity tracking, dynamically update the target identity mapping, and achieve contactless identity recognition of the target individual; S7. Real-time update the movement trajectory of the target individual in the video, adjust the tracking state in combination with the identity consistency score, and ensure the continuous stability of the identity of the target individual in the entire video sequence.

[0021] In this embodiment, the S2 specifically includes: S21. Perform human detection on the key frame images, extract the regions containing human targets, and perform scale normalization processing; S22. Perform skeleton key point detection on the normalized human target regions to obtain a set of skeleton key point coordinates , where represents the two-dimensional coordinates of the i-th skeleton key point, respectively represent the horizontal and vertical coordinates of the skeleton key point, and N is the number of skeleton key points; S23. According to the set of skeleton key point coordinates, calculate the Euclidean distance matrix between adjacent skeleton key points: ; Among them, is the Euclidean distance between the i-th and j-th skeleton key points, which is used to represent the skeleton structure relationship; S24. Construct a temporal skeleton sequence, set the time window length to Ι, and combine the skeleton key point coordinates of consecutive Ι frames to form a temporal feature matrix S, where each row corresponds to the coordinates of all skeleton key points in one frame; S25. Perform time interpolation on the temporal feature matrix S, and use the linear interpolation method to complete the coordinates of the skeleton key points of the missing frames to ensure data integrity; S26. Normalize the interpolated temporal feature matrix S so that the coordinates of each skeleton key point are normalized to the range of [-1, 1], obtaining a normalized temporal skeleton sequence.

[0022] In this embodiment, the specific steps of S3 are as follows: S31. Split the dimensions of the normalized temporal skeleton sequence, and separately extract the skeleton spatial feature matrix and the skeleton motion feature matrix , where represents the spatial coordinate information of the skeleton key points, represents the velocity change information of the skeleton key points between adjacent frames; S32. Perform local feature decomposition on the skeleton spatial feature matrix to calculate the relative relationship matrix between the skeleton key points: ; where, is the relative relationship value between the i-th and j-th skeleton key points, is the Euclidean distance between the i-th and j-th skeleton key points, is the Euclidean distance between the i-th and k-th skeleton key points, and N is the number of skeleton key points; S33. Perform adaptive time series decomposition on the skeleton motion feature matrix to construct a multi-scale time feature representation: ; where, T is the multi-scale time feature representation, m is the number of time scale features, is the m-th scale time feature representation, is the decomposition weight parameter, is the velocity change information of the skeleton key points between adjacent frames at time t; S34. Construct a gait feature matrix B using gait information, and extract gait features through a non-linear mapping , where f is the non-linear mapping function; S35. Construct a pose feature matrix P using pose information, and extract pose features through a non-linear mapping , where g is the non-linear mapping function and R is the relative relationship matrix; S36. Construct an action feature matrix O using action information, and extract action features through a non-linear mapping , where h is the non-linear mapping function; S37. Perform feature fusion on the gait feature matrix B, the pose feature matrix P, and the action feature matrix O to generate a normalized feature representation , where As the fusion weight parameter, and perform normalization processing.

[0023] In this embodiment, the S4 specifically includes: S41. Perform graph modeling on the standardized feature representation F to construct an adjacency matrix and a node feature matrix X, where represents the association weight between the skeleton key points i and j, and represent the feature vectors of the i-th and j-th skeleton key points respectively, is the scale factor, is the natural exponential function with the real number e as the base; S42. Use the improved multi-scale graph convolutional neural network to extract features from the input features, introduce the improved adaptive scale weight allocation mechanism, dynamically adjust the adjacency matrices of different scales, and calculate the multi-scale feature representation of the key points: ; Among them, is the adaptive scale weight, K is the number of scales, is the degree matrix corresponding to the adjacency matrix defined as , is the feature representation after the -th layer of graph convolution, is the feature representation after the -th layer of graph convolution, is the trainable weight matrix, is the non-linear activation function, and perform normalization processing on the scale weight to ensure the effective fusion of information of different scales; S43. After the multi-scale graph convolution calculation, introduce the cross-scale feature aggregation mechanism, optimize the information flow of features of different scales through residual connection, and calculate the global fusion feature: ; Among them, is the global fusion feature, is the number of scales, is the adaptive scale weight, is the feature representation after the k-th scale passes through layers of graph convolution, is the initial input node feature matrix, and use cross-scale residual connection to enhance feature stability; S44. Use the non-local aggregation mechanism to optimize the global fusion feature, calculate the long-distance dependence relationship between the skeleton key points globally, and construct the non-local relationship matrix G: ; ; Among them, is the natural exponential function with the real number e as the base, K is the number of scales, and are different mapping functions, used to calculate the similarity of key point features, and are the feature representations of the i-th and j-th skeleton key points after multi-scale graph convolution respectively, represents the non-local association weight between key point i and key point j, is the global fusion feature, As the feature representation after non-local enhancement, combining local and global information to optimize the identity representation feature; S45. Linearly transform the optimized feature representation map it to the identity coding space, generate the target identity coding, and perform normalization processing on the target identity coding to make it distributed on the unit sphere to ensure the stability and distinctiveness of the identity representation feature.

[0024] In this embodiment, the specific steps of S5 are as follows: S51. Perform cross-frame matching on the target identity coding to construct an identity similarity matrix based on feature association: ; Among them, is the natural exponential function with the real number e as the base, K is the number of scales, and are the target identity codings at time step i and time step j respectively, is the identity matching probability between time step i and time step j, represents the norm operation; S52. Based on the identity similarity matrix, calculate the identity consistency score by combining the non-local relationship matrix: ; Among them, is the identity consistency score at time step i, k is the number of scales, is the non-local association weight between key point i and key point k, is the identity matching probability between time step i and time step k, is the gait trajectory constraint parameter, ensuring that the identity matching not only depends on the appearance features but also combines the gait time series information; S53. Use the identity consistency score to optimize the target identity matching and construct an identity tracking trajectory: ; ; Among them, is the identity tracking trajectory, is the identity tracking state at time step i, K is the number of scales, is the target identity encoding at time step i, is the target identity encoding at time step k, is the identity matching probability between time step i and time step k, is the identity consistency score at time step i, representing the identity stability weighting factor to improve the continuity of identity tracking, is the identity historical matching weight parameter to ensure the stability of the tracking trajectory between frames and reduce the identity drift problem.

[0025] In this embodiment, the S6 specifically includes: S61. Based on the target identity encoding and the identity consistency score, calculate the identity matching score matrix between the current frame and the historical frames: ; Among them, represents the identity matching score between time step i and time step j, is the identity matching probability between time step i and time step j, represents the non-local association weight between key point i and key point j, is the identity consistency score at time step i, represents the Euclidean distance of the target identity encoding, serving as a constraint on the identity change, is the different feature weighting coefficient; S62. Based on the identity matching score matrix, calculate the identity mapping probability matrix U between the current frame and the historical frames: ; Among them, is the identity mapping probability that the current time step i is associated with the historical time step j, K is the number of scales, represents the identity matching score between time step i and time step j, represents the identity matching score between time step i and time step k, is the natural exponential function with the real number e as the base, and the formula adopts the softmax normalization operation to make the identity mapping change smoothly between frames; Perform regularization constraints on the identity mapping matrix to ensure that the identity mapping does not drift, and calculate the regularized identity mapping matrix: ; Among them, is the regularized identity mapping matrix, U is the identity mapping probability matrix, is the regularization parameter, and I is the identity matrix; S63. Update the calculation time step by combining the identity mapping matrix and the identity tracking trajectory of the identity tracking status: ; wherein, is the identity tracking status after the update calculation for time step i, is the regularized identity mapping matrix associated with the previous time step i to the historical time step k, is the historical identity tracking status of time step k, Calculate the identity deviation as the identity correction term, is the identity correction coefficient, enabling the identity tracking to be dynamically adjusted to ensure cross-frame identity consistency and achieve seamless identity recognition of target individuals.

[0026] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to a large-scale public security monitoring scenario. This scenario is established in a certain urban transportation hub area, and the main objective is to perform seamless identity recognition and target tracking through video stream data to enhance the security and monitoring efficiency of this area. There is a large volume of people flow in this area, especially during peak hours, and there are often situations where multiple people walk side by side. Traditional identity confirmation methods based on facial recognition or fingerprint recognition are prone to misrecognition or missed recognition, and these methods usually rely on the active cooperation of users or the support of specific devices, bringing inconvenience and limitations to practical applications.

[0027] During the implementation of the present invention, first, multiple high-definition cameras are deployed to capture the dynamic video stream data of this area in real time. The video stream data covers the whole process of pedestrians entering and leaving the monitoring area. Then, through the preprocessing of the video stream, key frame images containing human targets are extracted, and the key point coordinates of the humans in these images are extracted to construct a temporal skeleton sequence. This sequence is decomposed in multiple dimensions through the motion features, gait features, and posture features of the human body, and a standardized feature set is generated through normalization processing. This feature set reflects the behavioral patterns, gait, and posture features of individuals, can effectively distinguish different individuals, and provides key support for subsequent identity recognition and target tracking.

[0028] Next, the standardized features are input into an improved multi-scale graph convolutional neural network, and by combining the residual connection and the non-local aggregation mechanism, the identity characterization features are extracted. Through this improvement, the present invention can more accurately identify individuals in complex environments, especially in situations with large perspective changes and complex backgrounds, improving the accuracy and stability of identity extraction. After extracting the identity characterization features of the target individual, the system further generates the identity code of the target.

[0029] Another key innovation of the present invention is that the system performs real-time tracking and identity mapping of the target through target identity coding and identity consistency scores. As the target individual moves, the system dynamically updates the identity mapping of the target according to the target's identity coding and consistency scores, thereby realizing the seamless identity recognition and real-time tracking of the target individual in the video stream. This method does not rely on traditional manual interaction or external devices and can maintain a high recognition accuracy and stability in complex dynamic scenarios.

[0030] Through field tests conducted in transportation hub areas, the identity recognition accuracy and tracking stability of the system have been effectively verified. Specifically, during the test in this area, the system can accurately identify and track up to 2000 pedestrians during peak hours (such as morning and evening rush hours). The following is a data test comparison table showing the performance under different dates and traffic volumes: Table 1 Data Test Comparison Table ; Table 1 shows that the identity recognition accuracy in different time periods is higher than 95%. Especially during peak hours with dense personnel flow, the real-time performance and accuracy of the system remain stable, with the false recognition rate controlled within 3% and the missed recognition rate controlled within 2%. For example, in the test on March 1, 2025, the traffic volume in the monitored area reached 1500 person-times. After testing, the identity matching accuracy of the system reached 96%, the false recognition rate was 3%, and the missed recognition rate was 2%. In the test on this date, the average processing time of the system was 2.8 seconds, fully meeting the real-time requirements. In the peak-hour test on March 3, 2025, with a traffic volume of 2000 person-times, the identity recognition accuracy of the system still remained above 95%, and the real-time performance was also very stable, with the processing time remaining within 3 seconds.

[0031] Through this series of tests and verifications, the system has not only successfully solved the problems of traditional identity recognition systems in high-traffic and complex scenarios but also demonstrated high reliability and adaptability. This shows that the present invention provides an effective method for seamless identity recognition and real-time tracking, with good application prospects. Especially in high-traffic places such as transportation and security, it will greatly improve the monitoring efficiency and security. These data show that under different traffic volumes, the method adopted in the present invention always maintains a high level in terms of identity recognition accuracy and tracking stability, fully demonstrating its application advantages in complex environments.

[0032] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, should be covered within the protection scope of the present invention.

Claims

1. A method for non-intrusive identity recognition and real-time tracking of personnel in a video, characterized in that, It includes the following steps: S1. Obtain video stream data and perform preprocessing to extract key frame images containing human targets; S2. Extract human key point coordinates from the key frame images and construct a temporal skeleton sequence; S3. Perform adaptive multi-dimensional feature decomposition on the temporal skeleton sequence, extract gait features, action features, and pose features respectively, and perform normalization processing to generate standardized features; S4. Input the standardized features into an improved multi-scale graph convolutional neural network, combine the residual connection and non-local aggregation mechanism, extract identity representation features, and generate target identity codes; S5. Use similarity measurement technology to perform cross-frame matching on the target identity codes, calculate the identity consistency score, and establish an identity tracking trajectory; S6. Combine the target identity code and the identity consistency score in the video stream to perform identity tracking, dynamically update the target identity mapping, and realize the non-intrusive identity recognition of the target individual; S7. Real-time update the movement trajectory of the target individual in the video, adjust the tracking state in combination with the identity consistency score, and ensure the continuous stability of the identity of the target individual in the entire video sequence.

2. The method for contactless identity recognition and real-time tracking based on the people in the video according to claim 1, characterized in that The specific content of S2 includes: S21. Perform human detection on the key frame images, extract the regions containing human targets, and perform scale normalization processing; S22. Perform skeleton key point detection on the normalized human target area to obtain a set of skeleton key point coordinates , where represents the two-dimensional coordinates of the i-th skeleton key point, respectively represent the horizontal and vertical coordinates of the skeleton key point, and N is the number of skeleton key points; S23. According to the set of skeleton key point coordinates, calculate the Euclidean distance matrix between adjacent skeleton key points: ； Among them, is the Euclidean distance between the $i$-th and $j$-th skeletal key points, which is used to represent the skeletal structure relationship; S24. Construct a temporal skeleton sequence, set the time window length to Ι, and combine the skeleton key point coordinates of consecutive Ι frames to form a temporal feature matrix S, where each row corresponds to all the skeleton key point coordinates of one frame; S25. Perform time interpolation on the temporal feature matrix S, and use the linear interpolation method to complete the skeleton key point coordinates of the missing frames to ensure data integrity; S26. Perform normalization processing on the interpolated temporal feature matrix S to normalize each skeleton key point coordinate to the range of [-1, 1] to obtain a standardized temporal skeleton sequence.

3. The method for contactless identity recognition and real-time tracking based on the people in the video according to claim 1, wherein, The specific content of S3 includes: S31. Split the standardized timing skeleton sequence in terms of dimensions, and separately extract the skeleton spatial feature matrix and the skeleton motion feature matrix , where represents the spatial coordinate information of the skeleton key points, represents the velocity change information of the skeleton key points between adjacent frames; S32. Perform local eigen - decomposition on the skeleton spatial feature matrix to calculate the relative relationship matrix between skeleton key points: ； Among them, is the relative relationship value between the $i$-th and $j$-th skeleton key points, is the Euclidean distance between the $i$-th and $j$-th skeleton key points, is the Euclidean distance between the $i$-th and $k$-th skeleton key points, and $N$ is the number of skeleton key points; S33. Perform adaptive time series decomposition on the skeleton motion feature matrix to construct a multi-scale time feature representation: ； Among them, T is the multi-scale time feature representation, m is the number of time scale features, is the time feature representation of the m-th scale, is the decomposition weight parameter, is the velocity change information between adjacent frames of the skeleton key points at time t; S34. Construct a gait feature matrix B using gait information and extract gait features through a non-linear mapping where f is a non-linear mapping function. S35. Construct an attitude feature matrix P using the attitude information, and extract attitude features through a non-linear mapping where g is a non-linear mapping function and R is a relative relationship matrix; S36. Construct an action feature matrix O using the action information and extract action features through a non-linear mapping where h is a non-linear mapping function. S37. Perform feature fusion on the gait feature matrix B, the posture feature matrix P, and the action feature matrix O to generate a standardized feature representation , where is a fusion weight parameter and is normalized 4. The method for contactless identity recognition and real-time tracking of personnel in a video according to claim 1, wherein The specific content of S4 includes: S41. Perform graph modeling on the standardized feature representation F to construct an adjacency matrix and a node feature matrix X, where represents the association weight between the skeleton key points i and j, and represent the feature vectors of the i-th and j-th skeleton key points respectively, is a scale factor, is the natural exponential function with the real number e as the base; S42. Use the improved multi-scale graph convolutional neural network to extract features from the input features, introduce an improved adaptive scale weight allocation mechanism, dynamically adjust the adjacency matrices of different scales, and calculate the multi-scale feature representation of the key points: ； Among them, is the adaptive scale weight, K is the number of scales, is the corresponding adjacency matrix of the degree matrix, defined as , is the feature representation after graph convolution in the th layer, is the feature representation after graph convolution in the th layer, is the trainable weight matrix, is the non-linear activation function, and normalizes the scale weight to ensure the effective fusion of information at different scales; S43. After the multi-scale graph convolution calculation, introduce a cross-scale feature aggregation mechanism, optimize the information flow of different scale features through residual connection, and calculate the global fusion feature: ； Among them, is the global fusion feature, is the number of scales, is the adaptive scale weight, is the feature representation after the k-th level scale passes through layers of graph convolution, is the node feature matrix of the initial input, and the cross-scale residual connection is used to enhance the feature stability; S44. Use the non-local aggregation mechanism to optimize the global fusion feature, calculate the long-distance dependence relationship between the skeleton key points globally, and construct a non-local relationship matrix G: ； ； wherein, is the natural exponential function with the real number e as the base, K is the number of scales, and are different mapping functions for calculating the similarity of key point features, and are the feature representations of the i-th and j-th skeleton key points after multi-scale graph convolution respectively, represents the non-local association weight between key point i and key point j, is the global fusion feature, as the feature representation after non-local enhancement, combines local and global information to optimize the identity representation feature; S45. Perform a linear transformation on the optimized feature representation to map it to the identity coding space, generate the target identity coding, and perform normalization on the target identity coding so that it is distributed on the unit sphere to ensure the stability and distinctiveness of the identity representation features.

5. The method for contactless identity recognition and real-time tracking of personnel in a video according to claim 1, wherein, The specific content of S5 includes: S51. Perform cross-frame matching on the target identity code to construct an identity similarity matrix based on feature association: ； Among them, is the natural exponential function with the real number e as the base, K is the number of scales, and are the target identity encodings at time step i and time step j respectively, is the identity matching probability between time step i and time step j, represents the norm operation; S52. Based on the identity similarity matrix, calculate the identity consistency score in combination with the non-local relationship matrix: ； wherein, is the identity consistency score at time step i, k is the number of scales, is the non-local association weight between key point i and key point k, is the identity matching probability between time step i and time step k, is the gait trajectory constraint parameter, ensuring that the identity matching depends not only on appearance features but also combines gait time series information; S53. Use the identity consistency score to optimize the target identity matching and construct an identity tracking trajectory: ； ； Among them, is the identity tracking trajectory, is the identity tracking state at time step i, K is the number of scales, is the target identity encoding at time step i, is the target identity encoding at time step k, is the identity matching probability between time step i and time step k, is the identity consistency score at time step i, representing the identity stability weighting factor to improve the continuity of identity tracking, is the identity historical matching weight parameter to ensure the stability of the tracking trajectory between frames and reduce the identity drift problem.

6. The method for contactless identity recognition and real-time tracking based on people in a video according to claim 1, characterized in that, The specific content of S6 includes: S61. Based on the target identity code and the identity consistency score, calculate the identity matching score matrix between the current frame and the historical frames: ； Among them, represents the identity matching score between time step i and time step j, is the identity matching probability between time step i and time step j, represents the non-local correlation weight between key point i and key point j, is the identity consistency score of time step i, represents the Euclidean distance of the target identity encoding, serving as a constraint on identity change, is the different feature weighting coefficient; S62. Calculate the identity mapping probability matrix U between the current frame and the historical frame based on the identity matching score matrix: ； Among them, is the identity mapping probability associated with the historical time step j at the current time step i, K is the number of scales, represents the identity matching score between time step i and time step j, represents the identity matching score between time step i and time step k, is the natural exponential function with the real number e as the base. The formula adopts the softmax normalization operation to make the identity mapping distribution change smoothly between frames; Perform regularization constraints on the identity mapping matrix to ensure that the identity mapping does not drift, and calculate the regularized identity mapping matrix: ； Among them, is a regularization identity mapping matrix, and U is an identity mapping probability matrix. is a regularization parameter, and I is an identity matrix. S63. Update the calculated time step by combining the identity mapping matrix and the identity tracking trajectory of the identity tracking status: ； Among them, is the identity tracking status after the update calculation at time step i, is the regularization identity mapping matrix associated with the historical time step k at the previous time step i, is the historical identity tracking status at time step k, calculate the identity deviation as the identity correction term, is the identity correction coefficient, enabling the identity tracking to be dynamically adjusted to ensure cross-frame identity consistency and achieve seamless identity recognition of target individuals.

Images