Method for positioning personnel track in bid evaluation base based on multi-camera and face recognition

Abstract

This invention relates to the field of intelligent monitoring technology, specifically a method for locating personnel trajectories in a bidding evaluation base based on multiple cameras and facial recognition. The method includes: acquiring real-time video streams from various areas of the bidding evaluation base; binding identity tags to areas where facial recognition is successful; converting facial pixel coordinates to 3D world coordinates using camera intrinsic parameters and pose to obtain instantaneous position points; associating consecutive timestamp position points of the same identity to form a single-camera original sequence; based on the spatial layout and overlapping fields of view of multiple cameras, fusing sequences from different cameras of the same identity, and deduplicating and optimizing to generate a global position point sequence; forming a smooth sequence through time-aligned interpolation; calculating displacement vectors and combining them with map corrections to generate personnel movement trajectories. This method achieves high-precision positioning of personnel trajectories in a bidding evaluation base through accurate conversion from 2D pixels to 3D coordinates and optimization of position point fusion from overlapping fields of view of multiple cameras.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology, and in particular to a method for locating personnel trajectories in a bidding evaluation base based on multiple cameras and facial recognition. Background Technology

[0002] Accurate tracking of personnel movement within an enclosed space is crucial for evaluation site management. Existing technologies largely rely on single-camera two-dimensional planar tracking or simple RFID tag positioning. The former marks locations in video frames using facial or body features, but only outputs pixel coordinates, failing to reflect the actual spatial distribution of personnel within the site. The latter requires worn devices, which suffers from issues such as missed wearers and signal interference, and is difficult to cover the movements of untagged personnel. Current technologies generally fail to address the correspondence between two-dimensional pixel coordinates and physical space, resulting in a disconnect between location information and the actual scene, hindering refined behavioral analysis.

[0003] The shortcomings of existing technical solutions are: lack of a conversion mechanism from pixel coordinates of the face region to three-dimensional world coordinates, only using two-dimensional plane position to approximate the position of the person, ignoring spatial dimensions such as height and depth, thus limiting the positioning accuracy; when multiple cameras are deployed, the same person is repeatedly identified under adjacent cameras due to overlapping fields of view, resulting in redundant position points and accumulated coordinate deviations. Without establishing spatial association rules for deduplication and fusion, the global trajectory is prone to breakage or drift.

[0004] It is necessary to accurately map the pixel coordinates of faces into three-dimensional world coordinates through camera intrinsic parameters and pose calibration, and to establish a method for deduplication and coordinate optimization of position points under the overlapping field of view of multiple cameras, so as to solve the problems of spatial distortion of single two-dimensional positioning and insufficient fusion of multi-source data. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for locating personnel trajectory in bidding evaluation bases based on multiple cameras and facial recognition.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition, comprising: Collect real-time monitoring video streams from various areas within the base, and bind identity tags to each face area successfully identified in the real-time monitoring video stream; Based on the camera's intrinsic parameters and the pre-calibrated pose of the camera in the world coordinate system, the pixel coordinates of the face region in the video frame are converted into the corresponding three-dimensional spatial coordinates of the person in the world coordinate system, which are used as the instantaneous position of the person. For individuals with the same identity, multiple instantaneous location points identified under different cameras are linked on consecutive timestamps to form a sequence of the individual's original location points under a single camera. Based on the spatial layout and overlapping field of view of multiple cameras, the original location point sequences of the same person under different cameras are integrated, and the location points in the overlapping area are deduplicated and the coordinates are optimized to generate the global location point sequence of the person in the bidding base. The global location point sequence is time-aligned and interpolated to eliminate the uneven time intervals of location points caused by differences in video frame rates or missed detections, forming a smooth personnel location sequence with continuous timestamps and uniform location intervals. Based on the smoothed personnel location sequence, the displacement vector of personnel between adjacent timestamps is calculated. The displacement vector is then corrected by combining the map information of the internal spatial structure of the bidding site to generate the final personnel movement trajectory.

[0007] As a further aspect of the present invention, real-time monitoring video streams from various areas within the base are collected, and identity tags are bound to each successfully identified facial region in the real-time monitoring video stream, including: Multiple high-definition cameras with overlapping fields of view are deployed within the bidding evaluation base to collect real-time monitoring video streams of various areas within the base; The real-time monitoring video streams captured by each camera are processed frame by frame. The face detection algorithm is used to locate the position of all face regions in the video frame and extract the image features corresponding to each face region. Using a facial recognition algorithm, the image features of the facial region are compared and matched with a pre-entered database of authorized personnel facial features to identify the identity information of the corresponding personnel, and an identity tag is bound to each successfully identified facial region; The process involves processing each frame of the real-time monitoring video stream captured by each camera, locating the position of all face regions in the video frames using a face detection algorithm, and extracting the image features corresponding to each face region, including: Decode the real-time monitoring video stream and extract video frames according to the preset frame rate; The extracted video frames are preprocessed with illumination normalization and contrast enhancement. The preprocessed video frames are input into a trained deep neural network face detection model, which outputs the bounding box coordinates and confidence scores of each face region in the video frame. Filter out bounding boxes of face regions whose confidence scores are below the detection threshold; For each retained face region, the corresponding image is cropped out, and a pre-trained feature extraction network model is used to extract high-dimensional feature vectors from the cropped face images as the image features of the face region.

[0008] As a further aspect of the present invention, the step of using a face recognition algorithm to compare and match the image features of the face region with a pre-entered authorized personnel face feature database to identify the identity information of the corresponding personnel, and to bind an identity tag to each successfully identified face region, including: Calculate the cosine similarity between the image features of the face region and the feature vector of each registered person in the authorized personnel face feature database; Set an identity matching threshold and filter out registered individuals whose cosine similarity is higher than the identity matching threshold as candidate identities; If there is a unique candidate identity whose cosine similarity is higher than the identity matching threshold, then the registered person's identity information will be bound to the corresponding face region. If multiple candidate identities have a cosine similarity higher than the identity matching threshold, the registrant identity with the highest cosine similarity will be selected for binding. If the cosine similarity of all registered individuals is lower than the identity matching threshold, then the face region is bound to the identity label of "unknown person".

[0009] As a further aspect of the present invention, the step of converting the pixel coordinates of the face region in the video frame into the corresponding three-dimensional spatial coordinates of the person in the world coordinate system based on the camera's intrinsic parameters and the pre-calibrated pose of the camera in the world coordinate system, as the instantaneous position point of the person, includes: Obtain the pixel coordinates of the midpoint of the bottom edge of the face region bounding box as the pixel coordinates of the face's position in the image; Based on the camera's intrinsic parameter matrix and distortion coefficients, the pixel coordinates of the foothold are distorted to obtain normalized camera planar coordinates. Assuming the ground where the person is standing is a horizontal plane, and combining the extrinsic parameter matrix of the camera relative to the world coordinate system, the normalized camera plane coordinates are back-projected onto the horizontal ground in the world coordinate system through the geometric relationship of monocular vision ranging, and the three-dimensional coordinates of the person's instantaneous position point are calculated.

[0010] As a further aspect of the present invention, for individuals with the same identity, the method of associating multiple instantaneous location points of the individual identified under different cameras on a continuous timestamp to form an original location point sequence of the individual under a single camera includes: For each camera, extract the instantaneous location points of all personnel bound to the same identity tag according to the timestamp order of the video frames; Sort and connect the instantaneous location points of people with the same camera, the same identity, and consecutive timestamps according to the chronological order; When the time interval between two consecutive frames exceeds the preset maximum allowable interval, a trajectory interruption marker is inserted between the instantaneous personnel position points corresponding to the two frames. The sorted, connected, and labeled set of instantaneous personnel location points is used as the original location point sequence of the same person under the camera.

[0011] As a further aspect of the present invention, the step of fusing the original location point sequences of the same person under different cameras based on the spatial layout and overlapping field of view of multiple cameras, deduplicating and optimizing the coordinates of the location points in the overlapping area, and generating a global location point sequence of the person within the bidding site includes: Establish a fusion mapping table that describes the spatial relationship of the fields of view of all cameras, wherein the fusion mapping table records the spatial range of the overlapping area of ​​the fields of view of any two cameras; For individuals with the same identity, align the original location sequence of all their cameras to a unified timeline by timestamp; On a unified timeline, for any given moment, check if there are original location points from different cameras at the current moment; If there are original location points from different cameras at a certain moment, the fusion mapping table is used to determine whether the original location points are located within the overlapping field of view of the corresponding cameras. If it is located within the overlapping field of view, the three-dimensional coordinates of the original location points are averaged to generate an optimized fused location point, and the original multiple location points are deleted. If not located within the overlapping field of view, all original location points are preserved; All the processed location points are arranged in chronological order to form a global location point sequence.

[0012] As a further aspect of the present invention, the step of performing time alignment and interpolation processing on the global location point sequence to eliminate uneven time intervals between location points caused by differences in video frame rates or missed detections, forming a smooth personnel location sequence with continuous timestamps and uniform location intervals, includes: Set the baseline time interval for the global location point sequence; Check the actual time interval between adjacent location points in the global location point sequence; If the actual time interval is greater than a preset multiple of the baseline time interval, it is determined that there is a missed detection between adjacent location points. Spline interpolation is used to generate new location points within the missed detection time period based on the known coordinates of the location points before and after the missed detection, so as to complete the trajectory. If the actual time intervals are uneven but do not exceed the threshold, the global location point sequence is uniformly resampled so that the time interval between all adjacent location points is equal to the reference time interval, thus obtaining a smooth personnel location sequence.

[0013] As a further aspect of the present invention, the step of calculating the displacement vector of personnel between adjacent time stamps based on a smoothed personnel location sequence, and correcting the displacement vector by combining map information of the internal spatial structure of the bidding site to generate the final personnel movement trajectory includes: Calculate the displacement vector between each pair of adjacent position points in a smoothed personnel position sequence; Obtain an indoor map of the bid evaluation site, which includes spatial coordinate information of walls, doors, windows, and fixed obstacles; Check whether each displacement vector spatially interferes with walls or fixed obstacles in the indoor map; If the movement path indicated by the displacement vector interferes with a wall or fixed obstacle, the shortest feasible path connecting adjacent locations is found in the allowed paths on the map based on the shortest path algorithm, and the straight path indicated by the original displacement vector is replaced with the shortest feasible path. Connect all the corrected displacement vectors to generate the final personnel movement trajectory.

[0014] As a further aspect of the present invention, after generating the final personnel movement trajectory, the method further includes: The final personnel movement trajectory is analyzed, and the dwell area in the trajectory is extracted. The dwell area is defined by the spatial distance between multiple consecutive location points being less than the dwell determination threshold. Record the center coordinates, entry time, and exit time of each stopping area; The designated area is linked to a functional area map within the bid evaluation base, which defines the spatial boundaries of the bid evaluation room, corridor, rest area, and entrances / exits. When the center coordinates of the area where the person stays are located within the boundary of a certain functional area, it is determined that the person has entered the functional area, and the duration of the person's stay in the functional area is recorded.

[0015] As a further aspect of the present invention, after identifying the identity information of the corresponding person, it further includes: The system binds the identified personnel's identity information, instantaneous location points, and timestamps in real time to form a real-time personnel location record; Write the real-time personnel location records into a time-series database; Provides a trajectory query interface to receive trajectory query requests for specific individuals or time ranges; Based on the trajectory query request, the corresponding personnel movement trajectory data is retrieved from the time series database and output.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By employing the camera's intrinsic parameters and pre-calibrated camera pose in the world coordinate system, and using the pinhole camera model back-projection formula, the pixel coordinates of the face region in the video frame are converted into normalized image plane coordinates. These coordinates are then substituted into the 3D reconstruction formula to solve for the 3D spatial coordinates in the world coordinate system. This technology overcomes the limitation of conventional face recognition, which only outputs two-dimensional planar information. It endows the instantaneous location of a person with true spatial depth and height attributes, providing a precise location benchmark that conforms to the physical scene for trajectory positioning and avoiding spatial location misjudgments caused by two-dimensional approximation.

[0017] Based on the spatial layout and overlapping field of view of multiple cameras, this technique eliminates duplicate recordings of the same person's location from different cameras across consecutive time stamps. It filters for detections occurring at nearly the same moment by matching time stamps, calculates the Euclidean distance between different camera locations within the overlapping field of view, and optimizes the coordinates using a weighted average or triangulation method before deduplication and merging. This technology eliminates duplicate recordings of the same person's location caused by overlapping viewpoints from multiple cameras, corrects coordinate deviations by fusing data from the same source, and ensures that the generated global location point sequence covers the entire base area without redundancy. This improves trajectory continuity and avoids trajectory breaks or deviations caused by data conflicts. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition, as described in this invention. Figure 2 A flowchart for converting pixel coordinates to 3D spatial coordinates; Figure 3 A flowchart for generating a global location point sequence through multi-camera fusion; Figure 4 The curves showing the impact of different baseline time intervals on trajectory point density and smoothness scores; Figure 5 This is a graph showing the path efficiency analysis across multiple scenarios. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 The system uses multiple cameras deployed within the evaluation site to capture real-time monitoring video streams, detects and recognizes faces in the videos, and assigns a corresponding identity tag to each recognized face region. Based on the camera's internal parameters and pre-calibrated pose in the world coordinate system, the pixel coordinates of the face in the image are transformed into three-dimensional spatial coordinates in the world coordinate system, which serve as the instantaneous location of the person. For individuals with the same identity tag, multiple instantaneous location points acquired from the same camera at consecutive timestamps are associated to form the original location point sequence of that person from that camera's perspective. Based on the spatial layout of the multiple cameras and their overlapping fields of view, the original location point sequences of the same person from different cameras are fused. Deduplication and coordinate optimization operations are performed on the location points within the overlapping field of view areas to generate a global location point sequence of that person in the global space of the evaluation site. This global location point sequence is then time-aligned and interpolated to eliminate uneven time intervals between location points caused by differences in video frame rates or missed face recognition, forming a smooth personnel location sequence with continuous timestamps and uniform location intervals. Based on this smooth personnel location sequence, the displacement vectors of personnel between adjacent timestamps are calculated. Combined with map information of the bidding site, including walls, doors, windows and other structures, these displacement vectors are reasonably corrected to finally generate personnel movement trajectories that conform to the actual physical space constraints.

[0022] In one embodiment of the present invention, two high-definition cameras are deployed at the entrance area of ​​the evaluation base. The fields of view of the two high-definition cameras partially overlap, and the two high-definition cameras capture real-time monitoring video streams of the entrance area. The real-time monitoring video streams captured by each high-definition camera are processed frame-by-frame. The processing includes decoding the real-time monitoring video stream and extracting video frames at a preset frame rate of 25 frames per second. The extracted video frames undergo preprocessing for illumination normalization and contrast enhancement. Illumination normalization is performed by subtracting the mean and dividing by the variance, and contrast enhancement uses a histogram equalization algorithm. The preprocessed video frames are input into a trained deep neural network face detection model. The deep neural network face detection model outputs the bounding box coordinates and confidence scores of each face region in the video frames. The deep neural network face detection model is based on the YOLOv5 architecture and trained on the Wideface dataset. Face regions with confidence scores below a detection threshold of 0.8 are filtered out. For each retained face region, the corresponding image is cropped, and a pre-trained feature extraction network model is used to extract high-dimensional feature vectors from the cropped face images as image features of the face region. The feature extraction network model uses a pre-trained FaceNet model and outputs a 512-dimensional feature vector. The confidence score is calculated using the following formula: in: This represents the confidence score. This indicates the probability that the bounding box contains an object. The confidence score represents the probability that an object within the bounding box belongs to the face category. The deep neural network face detection model directly outputs the fused confidence score. .

[0023] In some embodiments, comparing video frames before and after preprocessing, the preprocessed video frames show a more balanced grayscale distribution and more prominent facial details. The deep neural network face detection model outputs an increased number of face bounding boxes in the preprocessed video frames. For example, under low-light conditions, the confidence score of a certain face region in an unprocessed video frame is 0.75, which is below the detection threshold of 0.8, and the face region is filtered out. However, after preprocessing with illumination normalization and contrast enhancement, the confidence score of the same face region in the same video frame increases to 0.88, which is above the detection threshold of 0.8, and the face region is retained for subsequent feature extraction. Using a face recognition algorithm, the cosine similarity between the image features of the face region and the feature vectors of each registered person in a pre-entered authorized personnel face feature database is calculated. The authorized personnel face feature database contains the facial feature vectors of 100 registered persons. An identity matching threshold of 0.9 is set. Registered individuals with a cosine similarity higher than the threshold are selected as candidate identities. If a unique candidate identity has a cosine similarity higher than the threshold, the registered individual's identity information is bound to the corresponding face region. If multiple candidate identities have cosine similarities higher than the threshold, the registered individual with the highest cosine similarity is selected for binding. If all registered individuals have cosine similarities lower than the threshold, the face region is labeled "Unknown Person."

[0024] Optionally, comparing different identity matching scenarios, when the cosine similarity between the image features of the face region and the feature vector of a registered person in the authorized personnel face feature database is 0.95, the cosine similarity is higher than the identity matching threshold of 0.9, and the system performs an identity binding operation, assigning the corresponding registered person's identity information to the face region. When the cosine similarity between the image features of the face region and the feature vectors of two registered persons in the authorized personnel face feature database are 0.92 and 0.91 respectively, both cosine similarities are higher than the identity matching threshold of 0.9, and the system performs a selection operation, binding the registered person's identity corresponding to the cosine similarity of 0.92 to the face region. When the highest cosine similarity between the image features of the face region and the feature vectors of all registered persons in the authorized personnel face feature database is 0.85, the cosine similarity is lower than the identity matching threshold of 0.9, and the system performs a label binding operation, binding the face region with the "unknown person" identity label.

[0025] It is understandable that preprocessing operations directly affect the output performance of the deep neural network face detection model. Illumination normalization reduces the impact of illumination changes, contrast enhancement strengthens edge information, and the preprocessed video frames are fed into the deep neural network face detection model, causing the confidence scores output by the model to concentrate towards higher values. The identity binding process relies on cosine similarity calculation, which reflects the similarity between feature vectors. The identity matching threshold is set to 0.9, filtering out low-quality matches to ensure accurate identity binding. In some embodiments, the preset frame rate is adjustable, affecting the video frame extraction frequency. A preset frame rate of 25 frames per second balances real-time processing with computational load. A higher preset frame rate results in shorter video frame intervals and a denser sequence of original location points. Optionally, the detection threshold is adjustable, affecting the strictness of face region filtering. A detection threshold of 0.8 filters out face region bounding boxes with confidence scores below 0.8, reducing false detection interference.

[0026] In one embodiment of the present invention, high-definition cameras are deployed in the corridor area of ​​the bidding evaluation base. These cameras capture real-time monitoring video streams of the corridor area, and the video frames are processed to obtain facial region image features, which are 512-dimensional vectors. The cosine similarity between the facial region image features and the feature vectors of each registered person in a pre-entered authorized personnel facial feature database is calculated. This database stores the facial feature vectors of 50 bidding evaluation experts and 20 staff members. An identity matching threshold of 0.85 is set, and registered persons with cosine similarities higher than the threshold are selected as candidate identities. For example, if a person's facial region image features have a cosine similarity of 0.91 with the feature vector of registered person A, 0.72 with the feature vector of registered person B, and 0.89 with the feature vector of registered person C in the authorized personnel facial feature database, and the cosine similarity of registered persons A and C is higher than the identity matching threshold of 0.85, then registered persons A and C are listed as candidate identities. The calculation formula is: in: This represents the face region image feature vector extracted from a video frame. This represents the feature vector of a registered person in the authorized personnel facial feature database. The symbol "·" indicates the dot product operation of the vectors. Representing vectors The length of the mold, Representing vectors The modulus length, calculation results The value of is between -1 and 1. The closer the value is to 1, the more similar the two feature vectors are.

[0027] In some embodiments, when there is a unique candidate identity with a cosine similarity higher than the identity matching threshold, an identity binding operation is performed. For example, if a face region image feature has a cosine similarity of 0.90 with the feature vector of registered person D (which is higher than the identity matching threshold of 0.85), but a cosine similarity lower than 0.85 with the feature vectors of all other registered persons, then registered person D is the unique candidate identity, and the system binds the identity information (such as employee ID and name) of registered person D to the corresponding face region. When multiple candidate identities have cosine similarities higher than the identity matching threshold, a selection operation is performed. For example, if a face region image feature has a cosine similarity of 0.88 with the feature vector of registered person E and 0.92 with the feature vector of registered person F, both of which are higher than the identity matching threshold of 0.85, the system compares the two cosine similarity values ​​and selects the identity of registered person F with the highest cosine similarity for binding. When the cosine similarity of all face region image features with the feature vectors of all registered persons in the authorized personnel face feature database is lower than the identity matching threshold, a label binding operation is performed. For example, if the highest cosine similarity between a face region image feature and all feature vectors in the database is 0.80, which is lower than the identity matching threshold of 0.85, the system will assign the identity label "unknown person" to this face region.

[0028] After identifying the corresponding personnel's identity information, the system binds the identified personnel identity information, the corresponding instantaneous personnel location point, and the current timestamp in real time to form a real-time personnel location record. The real-time personnel location record includes fields: personnel identification, three-dimensional coordinates (x, y, z), and timestamp (year-month-day hour:minute:second.millisecond). The real-time personnel location record is written to a time-series database using InfluxDB, with each record stored as a data point indexed by time. The system provides a trajectory query interface, which receives trajectory query requests for specific personnel and specific time ranges. Query requests are sent in the form of HTTP POST requests, with the request body containing the target personnel's identity and the start and end times of the query. Based on the trajectory query request, the system retrieves all real-time personnel location records matching the target personnel's identity within the specified time range from the time-series database. The retrieved records are sorted in ascending order by timestamp and output as a JSON array, with each element containing coordinates and a timestamp.

[0029] Optionally, the identity matching threshold can be configured. A threshold of 0.85 provides more lenient filtering, while a threshold of 0.92 provides stricter filtering. For example, when the threshold is set to 0.85, registered individuals with a cosine similarity of 0.88 will be bound to the database. When the threshold is raised to 0.90, the same registered individual will not be bound because their cosine similarity is below the new threshold, and their face area will be marked as "unknown person." Optionally, the data retention policy for the time-series database can also be configured. The policy can be set to retain real-time personnel location records for 90 days, with records older than 90 days automatically archived or deleted to control the database storage size.

[0030] It is understandable that cosine similarity comparison and threshold determination are the core logic of identity binding. The system calculates the cosine similarity between the facial region image features and the feature vectors of each registered person in the authorized personnel facial feature database. Each calculation result is compared with the identity matching threshold, triggering different binding or labeling operations based on the comparison result. Real-time personnel location records are the basic unit of trajectory data. The structured fields of real-time personnel location records ensure information integrity, and writing them into a time-series database facilitates efficient retrieval by time dimension. The trajectory query interface defines the data access specifications, receiving query requests containing personnel identifiers and time ranges, retrieving and outputting the corresponding data sequences from the database, enabling on-demand acquisition of trajectory data. In some embodiments, database queries have millisecond-level latency; the time interval from submitting a trajectory query request to receiving a complete JSON response is within 100 milliseconds, meeting real-time query requirements. It is understood that identity binding operations occur in each frame of the video stream processing. The frequency of identity binding operations is related to the video frame rate. When the video frame rate is 25 frames per second, the system can perform a maximum of 25 identity binding attempts per second, forming 25 real-time personnel location records and writing them into the database.

[0031] See Figure 2In one embodiment of the present invention, a fixed-view high-definition camera is deployed in the evaluation room of the evaluation base. The intrinsic parameter matrix of the high-definition camera is known, and the pose of the high-definition camera in the world coordinate system is pre-calibrated. The bounding box coordinates of a specific face region in the face detection result are obtained. The bounding box coordinates are the upper left corner (300, 200) and the lower right corner (400, 320). The pixel coordinates of the midpoint of the bottom edge of the bounding box are calculated. The pixel coordinates of the midpoint of the bottom edge are ((300+400) / 2, 320), i.e., (350, 320). These pixel coordinates are used as the pixel coordinates representing the person's footing in the image. According to the intrinsic parameter matrix of the high-definition camera and the lens distortion coefficient, the footing pixel coordinates (350, 320) are subjected to distortion correction processing. The distortion correction processing uses the Brown-Conrady model, and the corrected coordinates are (-0.012, 0.085, 1) on the normalized camera plane. Assuming the ground where the personnel are standing is the horizontal floor of the evaluation room, and considering the extrinsic parameter matrix of the high-definition camera relative to the world coordinate system (which includes a rotation matrix and a translation vector), the normalized camera plane coordinates (-0.012, 0.085, 1) are back-projected onto the horizontal ground in the world coordinate system using the geometric relationship of monocular visual ranging. The back-projection calculates the three-dimensional coordinates of the personnel's instantaneous position, for example, (2.15, 3.78, 0.00). The coordinate transformation relationship is described by the following formula: in: This represents the pixel coordinates of the standpoint after distortion correction. This represents the intrinsic parameter matrix of a high-definition camera. This represents a scale factor, which is obtained by solving the constraint that the person's feet are located on the ground plane (Z=0). This represents the rotation matrix of a high-definition camera. This represents the translation vector of the high-definition camera. The inverse of the rotation matrix is ​​calculated as follows: This indicates the instantaneous horizontal position of a person in the world coordinate system.

[0032] For the high-definition camera in the evaluation room, based on the timestamp sequence of the video frames, the instantaneous location points of all personnel bound to the same identity tag "Expert A" are extracted. The timestamps start from 10:00:00.000. The extracted instantaneous personnel location point sequence includes coordinates (2.15, 3.78, 0.00)@10:00:00.000, coordinates (2.18, 3.80, 0.00)@10:00:00.040, and coordinates (2.22, 3.83, 0.00)@10:00:00.080. These instantaneous personnel location points from the high-definition camera, with the identity tag "Expert A" and consecutive timestamps, are sorted and connected in chronological order to form a preliminary point sequence. When the time interval between two consecutive frames exceeds a preset maximum allowable interval (set to 200 milliseconds), a trajectory interruption marker is inserted between these two instantaneous personnel location points. For example, if the time interval between the instantaneous personnel location coordinates (2.30, 3.90, 0.00) @ 10:00:01.000 and the next instantaneous personnel location coordinates (2.45, 4.05, 0.00) @ 10:00:01.350 is 350 milliseconds, which is greater than the preset maximum allowable interval of 200 milliseconds, the system inserts a trajectory interruption marker between coordinates (2.30, 3.90, 0.00) and coordinates (2.45, 4.05, 0.00). The sorted, connected set of instantaneous personnel location points containing trajectory interruption markers is used as the original location point sequence of the person identified as "Expert A" under this high-definition camera. The data structure of the original location point sequence is a list, and each element in the list contains coordinates, a timestamp, and a possible trajectory interruption marker.

[0033] In some embodiments, comparing the selection of the midpoint of the bottom edge of different bounding boxes, the world coordinates calculated by backprojection differ between selecting the pixel coordinates of the midpoint of the bottom edge of the face region bounding box as the foothold pixel coordinates and selecting the pixel coordinates of the center point of the face region bounding box as the foothold pixel coordinates. For example, the instantaneous person's position calculated based on the pixel coordinates of the center point of the bounding box (350, 260) is (2.50, 4.10, 0.00), while the instantaneous person's position calculated based on the pixel coordinates of the midpoint of the bottom edge of the bounding box (350, 320) is (2.15, 3.78, 0.00). The midpoint coordinates are closer to the projection position of the person's feet on the ground, therefore the midpoint pixel coordinates are selected. The preset maximum allowable interval can be adjusted. The preset maximum allowable interval is set to 200 milliseconds for a video stream of 25 frames per second. When the video frame rate is reduced to 10 frames per second, the preset maximum allowable interval may need to be increased accordingly to accommodate a larger base time interval.

[0034] It is understandable that the calculation of instantaneous personnel location points relies on camera intrinsic and extrinsic parameters and ground plane assumptions. Pixel coordinates are mapped to the world coordinate system through a series of geometric transformations such as distortion correction and back projection, resulting in the personnel's three-dimensional spatial coordinates in the world coordinate system. Constructing the original location point sequence is a process of organizing data in chronological order, connecting discrete observation points of the same identity under the same camera, and inserting trajectory interruption markers between point pairs with excessively large time intervals. These trajectory interruption markers indicate periods of discontinuity or missing data in the trajectory. In some embodiments, the accuracy of the timestamp affects the determination of sequence continuity. The timestamp accuracy is at the millisecond level, accurately calculating time intervals of 40 milliseconds or 100 milliseconds and comparing them with a preset maximum allowable interval. Optionally, the preset maximum allowable interval can be dynamically set according to the camera frame rate. The preset maximum allowable interval is set to twice the average interval time of video frames. For example, for a video stream of 25 frames per second with an average interval of 40 milliseconds, the preset maximum allowable interval can be set to 80 milliseconds. Optionally, identity tag binding may fail. When a face region in a frame is marked as "unknown person", this frame will not generate instantaneous person location points bound with clear identity tags. Therefore, it will not be extracted into the original location point sequence corresponding to the identity. This may lead to an increase in the actual interval in the sequence caused by recognition failure.

[0035] See Figure 3In one embodiment of the present invention, cameras A and B are deployed in the corridor area of ​​the bidding base. The fields of view of the two cameras overlap in the middle area of ​​the corridor. A fusion mapping table is established to describe the spatial relationship between the fields of view of cameras A and B. The fusion mapping table records the spatial range of the overlapping area of ​​the fields of view of cameras A and B in the world coordinate system. The spatial range is represented by the minimum bounding rectangle, and the coordinates of the vertices of the rectangle are (5.0,2.0), (5.0,8.0), (10.0,8.0), (10.0,2.0). For the person identified as "Staff Member A", extract the original location point sequences from camera A and camera B. The original location point sequence from camera A includes: point A1(4.5,3.1,0.0)@10:00:01.000, point A2(5.2,3.3,0.0)@10:00:01.040, and point A3(6.0,3.5,0.0)@10:00:01.080; the original location point sequence from camera B includes: point B1(5.8,3.4,0.0)@10:00:01.080 and point B2(6.5,3.7,0.0)@10:00:01.120. The original position point sequences of camera A and camera B are aligned to a unified timeline according to their respective timestamps. On the unified timeline, at timestamp 10:00:01.080, there exists an original position point A3 (6.0, 3.5, 0.0) from camera A and an original position point B1 (5.8, 3.4, 0.0) from camera B. Based on the fusion mapping table, it is determined whether the original position points are located within the overlapping field of view of the corresponding cameras. The coordinates of point A3 (6.0, 3.5, 0.0) are located within the rectangle (5.0, 2.0) to (10.0, 8.0) defined by the fusion mapping table, and the coordinates of point B1 (5.8, 3.4, 0.0) are also located within this rectangle. Therefore, points A3 and B1 are determined to be located within the overlapping field of view of cameras A and B. The 3D coordinates of the original location points A3 and B1, located within the overlapping field of view, are averaged to generate an optimized fused location point with coordinates ((6.0+5.8) / 2,(3.5+3.4) / 2,(0.0+0.0) / 2), i.e., (5.9,3.45,0.0). The original points A3 and B1 are then deleted. On the unified timeline, timestamps 10:00:01.000 and 10:00:01.040 only contain original location points from camera A, and timestamp 10:00:01.120 only contains original location points from camera B. Since the coordinates of these points are not located within the same overlapping area defined by the fusion mapping table, all original location points A1, A2, and B2 are retained.All the processed location points were arranged in chronological order to form the global location point sequence for the identity "Staff Member A" within the bidding base: Point A1 (4.5, 3.1, 0.0) @10:00:01.000, Point A2 (5.2, 3.3, 0.0) @10:00:01.040, merged point (5.9, 3.45, 0.0) @10:00:01.080, Point B2 (6.5, 3.7, 0.0) @10:00:01.120. Refer to Table 1 for a comparison between the original location points and the merged global location point sequence.

[0036] Table 1: Comparison of Original Location Point and Global Location Point Sequences In some embodiments, the baseline time interval of the global location point sequence is set to 40 milliseconds, and the actual time interval between adjacent location points in the global location point sequence is checked. For example, in a global location point sequence, point P1@10:00:02.000, point P2@10:00:02.150, and point P3@10:00:02.190, the actual time interval between point P1 and point P2 is 150 milliseconds. The preset multiple of the baseline time interval is set to 2. The actual time interval of 150 milliseconds is greater than twice the baseline time interval of 40 milliseconds (80 milliseconds), so it is determined that there is a missed detection between point P1 and point P2. Using spline interpolation, based on the coordinates of points P1 and P2, new location points are generated during the time interval between the timestamp of point P1 (10:00:02.000) and the timestamp of point P2 (10:00:02.150). Three new points are generated at times 10:00:02.040, 10:00:02.080, and 10:00:02.120 to complete the trajectory. The formula for calculating the coordinates of the new points using spline interpolation involves a piecewise cubic function. It is given by the following formula: in: Indicates the time to be interpolated. Represents the time at a known point P1, and the coefficients are... The coordinates of known point P1 and P2, along with the first-order derivative boundary conditions at the known points, are used to determine the smoothness of the position coordinates over time through interpolation. The actual time interval between points P2 and P3 is 40 milliseconds, which is no greater than twice the baseline time interval of 40 milliseconds (80 milliseconds). However, the time intervals between points P1 and the interpolation point, between the interpolation point and point P2, and between point P2 and point P3 are not completely uniform. The completed global position point sequence is uniformly resampled at 40-millisecond intervals, generating a series of equally spaced time points on the time axis, such as 10:00:02.000, 10:00:02.040, 10:00:02.080,... The position coordinates corresponding to these time points are calculated or used to ensure that the time interval between all adjacent position points equals the baseline time interval of 40 milliseconds, resulting in a smoothed personnel position sequence.

[0037] It is understandable that the fusion mapping table defines the spatial basis for multi-camera data fusion. Determining whether points from different cameras are in the same overlapping field of view based on the fusion mapping table is a prerequisite for performing coordinate averaging optimization and deduplication operations. Time alignment maps the original sequences from different time bases to a unified time axis, which is the basis for determining whether multi-source data exists at the same time. Coordinate averaging merges observations from different cameras within the overlapping field of view. Coordinate averaging can reduce single-camera positioning errors and improve the accuracy of the overlapping area position. In some embodiments, the choice of the reference time interval affects the trajectory smoothness. A reference time interval of 40 milliseconds corresponds to a virtual sampling rate of 25 Hz. A smaller reference time interval, such as 20 milliseconds, will result in a denser smooth position sequence, while a larger reference time interval, such as 100 milliseconds, will result in a sparser smooth position sequence. Optionally, the threshold multiple for determining missed detections can be adjusted. A preset multiple of 2 times for the reference time interval means that interpolation is triggered when the actual interval exceeds 80 milliseconds, while a preset multiple of 3 times means that interpolation is triggered when the actual interval exceeds 120 milliseconds. Optionally, spline interpolation can employ different boundary conditions. For example, natural spline boundary conditions set the second derivative to zero, while fixed slope boundary conditions set the first derivative to be known. Different boundary conditions affect the trend of the interpolated trajectory at the endpoints. Uniform resampling is the final step in trajectory smoothing. Uniform resampling ensures that the output sequence has strictly uniform time intervals, and it facilitates subsequent processing such as displacement calculation and velocity analysis.

[0038] See Figure 4In the personnel trajectory localization method for bidding bases based on multi-camera and facial recognition, the selection of the reference time interval is a key parameter affecting trajectory quality. Its impact on trajectory point density and trajectory smoothness score can be clearly observed in the figure. The horizontal axis represents the reference time interval (milliseconds), the left side of the vertical axis represents the trajectory point density (points / second), and the right side represents the trajectory smoothness score (0-100). As the reference time interval gradually increases from 20 milliseconds to 100 milliseconds, the trajectory point density shows a significant monotonically decreasing trend: at 20 milliseconds, the trajectory point density reaches 50 points / second, corresponding to a virtual sampling rate of 50Hz; when the reference time interval increases to 100 milliseconds, the trajectory point density drops to 10 points / second, corresponding to a virtual sampling rate of 10Hz. This change directly reflects the inverse relationship between the reference time interval and the trajectory point density; that is, the smaller the reference time interval, the denser the trajectory sampling points per unit time, and vice versa. Meanwhile, the trajectory smoothness score also decreased continuously with the increase of the baseline time interval: at 20 milliseconds, the smoothness score reached 95 points; at 40 milliseconds, the score dropped to 90 points; at 60 milliseconds, the score dropped to 75 points; at 80 milliseconds, the score dropped to 60 points; and at 100 milliseconds, the score dropped to 45 points. This trend indicates that a denser sequence of trajectory points can provide richer location information, thus generating a smoother trajectory for personnel movement through processes such as spline interpolation and uniform resampling; however, when the baseline time interval is too long, the original sequence of location points is too sparse, and even after interpolation completion, it is difficult to restore the true details of personnel movement, resulting in a significant decrease in trajectory smoothness.

[0039] In one embodiment of the present invention, the displacement vector of a person between adjacent timestamps is calculated based on a smoothed personnel position sequence. The smoothed personnel position sequence includes position points arranged in chronological order, including point S1(1.0,2.0,0.0)@10:05:00.000, point S2(1.2,2.3,0.0)@10:05:00.040, point S3(1.5,2.7,0.0)@10:05:00.080, and point S4(3.0,2.7,0.0)@10:05:00.120. Calculate the displacement vector between each pair of adjacent points in the smoothed personnel position sequence. The displacement vector consists of the starting point coordinates, the ending point coordinates, and the direction. For example, the displacement vector from point S1 to point S2 is from (1.0, 2.0) to (1.2, 2.3), and the displacement vector from point S3 to point S4 is from (1.5, 2.7) to (3.0, 2.7). Obtain the indoor map of the bidding site. The indoor map contains the spatial coordinate information of walls, doors, windows, and fixed obstacles. Walls are defined by line segments. For example, the endpoint coordinates of wall W1 are (1.8, 2.0) and (1.8, 4.0), and the center coordinates of a fixed obstacle such as column OBJ1 are (2.5, 2.7) with a radius of 0.3. Check whether each displacement vector spatially interferes with the walls or fixed obstacles in the indoor map. Spatial interference is checked by calculating whether the line segment represented by the displacement vector intersects with the wall line segment, or whether the distance from the line segment to the center of the fixed obstacle is less than the radius of the obstacle. The displacement vector from point S3 (1.5, 2.7, 0.0) to point S4 (3.0, 2.7, 0.0) indicates a horizontal straight path from (1.5, 2.7) to (3.0, 2.7). This straight path passes through the area of ​​column OBJ1 with center (2.5, 2.7) and radius 0.3. The distance from the straight path to the center (2.5, 2.7) is 0, which is less than the radius 0.3. Therefore, this displacement vector is determined to have spatial interference with the fixed obstacle. The straight path indicated by the displacement vector from point S1 to point S2 does not intersect any wall segment, and the distance to the center of column OBJ1 is greater than 0.3. Therefore, this displacement vector is determined not to have spatial interference. When the movement path indicated by the displacement vector interferes with a wall or fixed obstacle, the shortest path algorithm searches for the shortest feasible path connecting adjacent points within the map's allowed pathways. For the interfering displacement vector, due to the obstruction of pillar OBJ1 between the starting point S3 (1.5, 2.7) and the ending point S4 (3.0, 2.7), the shortest feasible path might be to go north from point S3 around the pillar, with the path being (1.5, 2.7) -> (1.5, 3.2) -> (3.0, 3.2) -> (3.0, 2.7). The total length of the shortest feasible path is... The calculation method is the sum of the Euclidean distances of each path segment: in: This represents the total number of path points traversed by the shortest feasible path. Indicates the first The coordinates of the path points This represents the total length of the shortest feasible path. The formula calculates the cumulative Euclidean distance from the starting point to the ending point, passing through each path point in sequence. The straight path indicated by the original displacement vector is replaced with the shortest feasible path. All the corrected displacement vectors are connected to generate the final personnel movement trajectory, which consists of a series of continuous path segments that conform to physical space constraints.

[0040] The final personnel movement trajectory is analyzed to extract the dwell areas within the trajectory. A dwell area is defined as a region where the spatial distance between multiple consecutive points in the trajectory is less than a dwell threshold, which is set to 0.5 meters. For example, if a trajectory segment contains consecutive points P1(5.0,5.0,0.0)@10:10:00.000, P2(5.1,5.0,0.0)@10:10:00.040, P3(5.0,5.1,0.0)@10:10:00.080, and P4(5.2,4.9,0.0)@10:10:00.120, the Euclidean distance between any two points P1, P2, P3, and P4 is calculated. If all distances are less than 0.5 meters, these four points are determined to constitute a dwell area. Record the center coordinates, entry time, and exit time of each dwelling area. The center coordinates are obtained by averaging the coordinates of all locations within the dwelling area. The entry time is the timestamp of the first location within the dwelling area, and the exit time is the timestamp of the last location within the dwelling area. Associate the dwelling areas with a functional area map within the bidding site. The functional area map defines the spatial boundaries of the bidding room, corridors, rest areas, and entrances / exits. For example, the boundary of bidding room Room 1 is a rectangular area from (4.0, 4.0) to (6.0, 6.0). When the center coordinates of a dwelling area are located inside the boundary of a functional area, it is determined that the person has entered that functional area. For example, if the center coordinates of a dwelling area (5.0, 5.0) are located inside the boundary of bidding room Room 1 (4.0, 4.0) to (6.0, 6.0), it is determined that the person has entered bidding room Room 1, and the duration of the person's stay in the functional area is recorded. The duration of stay is the exit time minus the entry time.

[0041] In some embodiments, spatial interference checks employ different methods for different types of obstacles. For linear obstacles such as walls, the checks examine whether the displacement vector line segment intersects with the wall line segment. For circular obstacles such as fixed obstacles, the shortest distance from the displacement vector line segment to the center of the obstacle is calculated, and it is determined whether the shortest distance is less than the obstacle radius. The shortest path algorithm can be Algorithm A, which searches for the shortest path from the starting point to the ending point on a map grid consisting of traversable and impassable areas (walls, obstacles). The dwell determination threshold is configurable. A dwell determination threshold of 0.5 meters is suitable for identifying people standing or sitting within a small area, while a dwell determination threshold of 1.0 meter allows people to have a larger range of movement but is still considered to be dwelling.

[0042] It is understandable that displacement vector correction ensures the trajectory conforms to physical space constraints. It checks whether the displacement vector interferes with indoor map elements and replaces interfering vectors with the shortest detour path, ensuring that the generated personnel movement trajectory does not pass through walls or fixed obstacles. Dwelling area extraction is based on the spatial clustering of location points; multiple consecutive location points close to each other indicate that a person is staying there. The dwelling determination threshold defines the spatial scale of "close." Functional area association maps abstract coordinate dwellings to semantically meaningful functional areas. By determining whether the center of the dwelling area falls within the functional area boundary, it establishes the association between personnel activities and the functional zoning of the bidding base (such as bidding rooms and rest areas). In some embodiments, the dwelling area may span the boundaries of multiple functional areas. When the center coordinates of the dwelling area are not uniquely located within any functional area, the determination can be based on the functional area to which the majority of location points within the dwelling area belong or on a preset priority rule. Optionally, recording the dwelling duration of personnel in functional areas can be used for subsequent behavior analysis. The dwelling duration is calculated from the entry time to the departure time, and the departure time is determined by the timestamp of the first location point whose displacement exceeds the threshold after the dwelling period ends. The generation of the shortest feasible path depends on the accuracy of the map and the efficiency of the algorithm. The map needs to accurately mark all impassable walls and fixed obstacles.

[0043] See Figure 5In multi-scenario path efficiency analysis, the quantitative evaluation of path optimization effects and scenario complexity relies on three core indicators: original path length, optimized path length, and path efficiency. Specifically, the original and optimized path lengths are expressed in meters, representing the uncorrected and corrected movement distances of a person in the corresponding scenario. Path efficiency is expressed as a percentage, calculated as the ratio of the optimized path length to the original path length, reflecting the degree to which spatial interference correction improves path rationality. In obstacle-free scenarios, both the original and optimized path lengths are at extremely low levels, with path efficiency approaching 100%, indicating that the displacement vector does not require correction when there are no physical constraints, and the trajectory is naturally optimal. In single-column obstacle scenarios, path efficiency drops to approximately 95%, indicating that the interference of a single fixed obstacle on the path is limited, and the increase in detour path length is small. In multi-wall obstacle scenarios, path efficiency further decreases to approximately 94%, demonstrating that linear obstacles such as walls exert stronger constraints on the path, increasing detour complexity. In complex environments, path efficiency drops to about 92%, while the lengths of both the original and optimized paths increase significantly. This indicates that the complex spatial structure with multiple obstacles interfering with the path most significantly, requiring correction through a more circuitous shortest feasible path.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition, characterized in that: include: Collect real-time monitoring video streams from various areas within the base, and bind identity tags to each face area successfully identified in the real-time monitoring video stream; Based on the camera's intrinsic parameters and the pre-calibrated pose of the camera in the world coordinate system, the pixel coordinates of the face region in the video frame are converted into the corresponding three-dimensional spatial coordinates of the person in the world coordinate system, which are used as the instantaneous position of the person. For individuals with the same identity, multiple instantaneous location points identified under different cameras are linked on consecutive timestamps to form a sequence of the individual's original location points under a single camera. Based on the spatial layout and overlapping field of view of multiple cameras, the original location point sequences of the same person under different cameras are integrated, and the location points in the overlapping area are deduplicated and the coordinates are optimized to generate the global location point sequence of the person in the bidding base. The global location point sequence is time-aligned and interpolated to eliminate the uneven time intervals of location points caused by differences in video frame rates or missed detections, forming a smooth personnel location sequence with continuous timestamps and uniform location intervals. Based on the smoothed personnel location sequence, the displacement vector of personnel between adjacent timestamps is calculated. The displacement vector is then corrected by combining the map information of the internal spatial structure of the bidding site to generate the final personnel movement trajectory.

2. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 1, characterized in that, Real-time surveillance video streams from various areas within the base are collected, and identity tags are assigned to each successfully identified facial region in the real-time surveillance video stream, including: Multiple high-definition cameras with overlapping fields of view are deployed within the bidding evaluation base to collect real-time monitoring video streams of various areas within the base; The real-time monitoring video streams captured by each camera are processed frame by frame. The face detection algorithm is used to locate the position of all face regions in the video frame and extract the image features corresponding to each face region. Using a facial recognition algorithm, the image features of the facial region are compared and matched with a pre-entered database of authorized personnel facial features to identify the identity information of the corresponding personnel, and an identity tag is bound to each successfully identified facial region; The process involves processing each frame of the real-time monitoring video stream captured by each camera, locating the position of all face regions in the video frames using a face detection algorithm, and extracting the image features corresponding to each face region, including: Decode the real-time monitoring video stream and extract video frames according to the preset frame rate; The extracted video frames are preprocessed with illumination normalization and contrast enhancement. The preprocessed video frames are input into a trained deep neural network face detection model, which outputs the bounding box coordinates and confidence scores of each face region in the video frame. Filter out bounding boxes of face regions whose confidence scores are below the detection threshold; For each retained face region, the corresponding image is cropped out, and a pre-trained feature extraction network model is used to extract high-dimensional feature vectors from the cropped face images as the image features of the face region.

3. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 2, characterized in that, The process involves using a facial recognition algorithm to compare and match the image features of the facial region with a pre-recorded database of authorized personnel facial features to identify the identity information of the corresponding personnel, and binding an identity tag to each successfully identified facial region, including: Calculate the cosine similarity between the image features of the face region and the feature vector of each registered person in the authorized personnel face feature database; Set an identity matching threshold and filter out registered individuals whose cosine similarity is higher than the identity matching threshold as candidate identities; If there is a unique candidate identity whose cosine similarity is higher than the identity matching threshold, then the registered person's identity information will be bound to the corresponding face region. If multiple candidate identities have a cosine similarity higher than the identity matching threshold, the registrant identity with the highest cosine similarity will be selected for binding. If the cosine similarity of all registered individuals is lower than the identity matching threshold, then the face region is bound to the identity label of "unknown person".

4. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 3, characterized in that, The step of converting the pixel coordinates of the face region in the video frame into the corresponding three-dimensional spatial coordinates of the person in the world coordinate system based on the camera's intrinsic parameters and the pre-calibrated pose of the camera in the world coordinate system, as the instantaneous position point of the person, includes: Obtain the pixel coordinates of the midpoint of the bottom edge of the face region bounding box as the pixel coordinates of the face's position in the image; Based on the camera's intrinsic parameter matrix and distortion coefficients, the pixel coordinates of the foothold are distorted to obtain normalized camera planar coordinates. Assuming the ground where the person is standing is a horizontal plane, and combining the extrinsic parameter matrix of the camera relative to the world coordinate system, the normalized camera plane coordinates are back-projected onto the horizontal ground in the world coordinate system through the geometric relationship of monocular vision ranging, and the three-dimensional coordinates of the person's instantaneous position point are calculated.

5. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 4, characterized in that, For individuals with the same identity, multiple instantaneous location points identified under different cameras are associated with continuous timestamps to form a sequence of original location points of the individual under a single camera, including: For each camera, extract the instantaneous location points of all personnel bound to the same identity tag according to the timestamp order of the video frames; Sort and connect the instantaneous location points of people with the same camera, the same identity, and consecutive timestamps according to the chronological order; When the time interval between two consecutive frames exceeds the preset maximum allowable interval, a trajectory interruption marker is inserted between the instantaneous personnel position points corresponding to the two frames. The sorted, connected, and labeled set of instantaneous personnel location points is used as the original location point sequence of the same person under the camera.

6. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 5, characterized in that, The method, based on the spatial layout and overlapping field of view of multiple cameras, fuses the original location point sequences of the same person under different cameras, performs deduplication and coordinate optimization on the location points in the overlapping area, and generates a global location point sequence of the person within the bidding base, including: Establish a fusion mapping table that describes the spatial relationship of the fields of view of all cameras, wherein the fusion mapping table records the spatial range of the overlapping area of ​​the fields of view of any two cameras; For individuals with the same identity, align the original location sequence of all their cameras to a unified timeline by timestamp; On a unified timeline, for any given moment, check if there are original location points from different cameras at the current moment; If there are original location points from different cameras at a certain moment, the fusion mapping table is used to determine whether the original location points are located within the overlapping field of view of the corresponding cameras. If it is located within the overlapping field of view, the three-dimensional coordinates of the original location points are averaged to generate an optimized fused location point, and the original multiple location points are deleted. If not located within the overlapping field of view, all original location points are preserved; All the processed location points are arranged in chronological order to form a global location point sequence.

7. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 6, characterized in that, The step of performing time alignment and interpolation processing on the global location point sequence to eliminate uneven time intervals between location points caused by differences in video frame rates or missed detections, forming a smooth personnel location sequence with continuous timestamps and uniform location intervals, includes: Set the baseline time interval for the global location point sequence; Check the actual time interval between adjacent location points in the global location point sequence; If the actual time interval is greater than a preset multiple of the baseline time interval, it is determined that there is a missed detection between adjacent location points. Spline interpolation is used to generate new location points within the missed detection time period based on the known coordinates of the location points before and after the missed detection, so as to complete the trajectory. If the actual time intervals are uneven but do not exceed the threshold, the global location point sequence is uniformly resampled so that the time interval between all adjacent location points is equal to the reference time interval, thus obtaining a smooth personnel location sequence.

8. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 7, characterized in that, The process involves calculating the displacement vectors of personnel between adjacent time stamps based on a smoothed personnel location sequence, correcting the displacement vectors using map information of the internal spatial structure of the bidding site, and generating the final personnel movement trajectory, including: Calculate the displacement vector between each pair of adjacent position points in a smoothed personnel position sequence; Obtain an indoor map of the bid evaluation site, which includes spatial coordinate information of walls, doors, windows, and fixed obstacles; Check whether each displacement vector spatially interferes with walls or fixed obstacles in the indoor map; If the movement path indicated by the displacement vector interferes with a wall or fixed obstacle, the shortest feasible path connecting adjacent locations is found in the allowed paths on the map based on the shortest path algorithm, and the straight path indicated by the original displacement vector is replaced with the shortest feasible path. Connect all the corrected displacement vectors to generate the final personnel movement trajectory.

9. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 8, characterized in that, After generating the final personnel movement trajectory, the following is also included: The final personnel movement trajectory is analyzed, and the dwell area in the trajectory is extracted. The dwell area is defined by the spatial distance between multiple consecutive location points being less than the dwell determination threshold. Record the center coordinates, entry time, and exit time of each stopping area; The designated area is linked to a functional area map within the bid evaluation base, which defines the spatial boundaries of the bid evaluation room, corridor, rest area, and entrances / exits. When the center coordinates of the area where the person stays are located within the boundary of a certain functional area, it is determined that the person has entered the functional area, and the duration of the person's stay in the functional area is recorded.

10. The method for locating personnel trajectory in a bidding evaluation base based on multiple cameras and facial recognition as described in claim 9, characterized in that, After identifying the identity information of the corresponding person, it also includes: The system binds the identified personnel's identity information, instantaneous location points, and timestamps in real time to form a real-time personnel location record; Write the real-time personnel location records into a time-series database; Provides a trajectory query interface to receive trajectory query requests for specific individuals or time ranges; Based on the trajectory query request, the corresponding personnel movement trajectory data is retrieved from the time series database and output.

Images