A multi-camera-based invigilator-free examination room examinee positioning method

Abstract

The application provides a multi-camera-based unmanned invigilation examinee positioning method in an examination room, relates to the technical field of computer vision, and solves the technical problems of a single camera scheme being mostly used in the prior art for unmanned invigilation, being limited to single-view image information, resulting in the problems of a view angle blind area, occlusion and incomplete information when detecting examinees; the technical scheme is as follows: a plurality of monocular cameras at fixed positions are arranged to realize all-around coverage of the view angle of the examination room; and the application can accurately realize three-dimensional positioning of the examinees through multi-frame image combination, instance segmentation and depth estimation algorithm.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a method for positioning examinees in an unmanned examination room based on multiple cameras. Background Art

[0002] In modern exam environments, ensuring exam fairness and security is a significant challenge. Traditional manual proctoring methods have numerous limitations, such as high labor costs and limited monitoring range. With the advancement of computer vision technology, unmanned proctoring has become an important means of improving exam management efficiency and accuracy.

[0003] Existing unmanned exam proctoring systems often rely on a single camera, limited to image information from a single perspective. This leads to blind spots, occlusions, and incomplete information when detecting examinees. This is especially true in complex exam environments with multiple examinees taking the test simultaneously, where a single camera struggles to comprehensively and accurately detect all examinees. Furthermore, the depth estimation accuracy of a single camera is limited, making it impossible to achieve three-dimensional positioning of examinees. Summary of the Invention

[0004] Therefore, the present invention solves the technical problems in the prior art that unmanned examinations mostly adopt a single camera solution, which is limited to image information from a single perspective, resulting in blind spots, occlusion problems and incomplete information when detecting candidates; the present invention provides a method for positioning candidates in an unmanned examination room based on multiple cameras, which sets up multiple monocular cameras in fixed positions to achieve full coverage of the examination room's perspective, and can accurately achieve three-dimensional positioning of candidates through the combination of multi-frame images with instance segmentation and depth estimation algorithms.

[0005] The inventive concept of the present invention is to achieve accurate three-dimensional positioning of the examinee through multiple steps. First, the left and right monocular cameras are fixed and the intrinsic parameters, extrinsic parameters and relative extrinsic parameters of the cameras are obtained through checkerboard calibration. Subsequently, the collected images are preprocessed and instance segmented to remove low-confidence and repeated examinee masks to ensure the accuracy of examinee detection. Then, the ZoeDepth algorithm is used to estimate the depth of the examination room image, and the segmented examinee mask is converted into a three-dimensional point cloud. Finally, the examinees under the two cameras are matched through the matrix relationship between the cameras and the three-dimensional coordinates of the examinees are obtained.

[0006] To achieve the above-mentioned purpose, the present invention adopts the following technical solutions: a 3D candidate positioning method based on multiple fields of view in an examination room, comprising the following steps:

[0007] S1: Two monocular cameras (front left and front right) are fixed in position, a checkerboard is placed in the field of view of the two cameras, and the intrinsic and extrinsic parameters of the two monocular cameras and their relative extrinsic parameters are obtained by capturing images of the calibration plate taken at different positions and directions;

[0008] S2: Perform image preprocessing on the examination room images captured by the left and right cameras, including image size unification, image pixel normalization, and image denoising. The pixel value range is converted, for example, from 0 to 255 to 0 to 1. A 3D block matching filter algorithm is then used for denoising. While retaining the original information of the examination room image, a series of noises (such as the blackboard) in the image are filtered out, making the processed examination room image smoother.

[0009] S3: Perform instance segmentation on the examinees in the images captured by the left and right cameras, and perform deduplication processing on the generated examinee masks to achieve examinee identification and segmentation, ensuring the accuracy and uniqueness of the segmentation results;

[0010] S4: Use the ZoeDepth algorithm to estimate the depth of the captured left and right examination room images, and save the generated relative depth map in PFM format;

[0011] S5: Process the obtained depth map of the examination room and convert the examinees segmented from the left and right images into a three-dimensional point cloud using the camera's intrinsic reference;

[0012] S6: Calculate the center position of the point cloud of each examinee segmented from the left and right images and optimize the center position;

[0013] S7: Convert the point cloud of the right camera to the coordinate system of the left camera, match the point clouds of the candidates under the left and right cameras, and obtain the three-dimensional coordinates of each candidate.

[0014] Furthermore, the specific steps of step S1 are as follows:

[0015] S11: Place a checkerboard in the left and right front fields of view of the two cameras, and capture images of the checkerboard from multiple angles to ensure that it is clearly visible at different positions and angles;

[0016] S12: Read the checkerboard images captured by each camera using the OpenCV computer vision library, detect the corner points (feature points) of the checkerboard in each image, and calculate the intrinsic parameter K, rotation matrix R, and translation vector T of each camera;

[0017] S13: Calculate the relative external parameters of the two cameras through a pair of checkerboard images taken by the two cameras to obtain the relative rotation matrix R between the two cameras rel and the relative translation vector T rel .

[0018] Furthermore, the specific steps of step S3 are as follows:

[0019] S31: Read the examination room images captured by the left and right cameras and perform instance segmentation to obtain the mask, label, and confidence score of each detected examinee. Filter out examinee instances with a confidence score lower than 0.5. From the filtered results, only examinee masks and corresponding confidence scores are selected to ensure that only examinee instances with higher confidence scores are retained.

[0020] S32: Use the improved IoU algorithm to calculate the IoU between masks and retain the candidate masks with high confidence and no duplication;

[0021] S33: Create a completely black background of the same size as the original image, superimpose the mask onto the background, and use random colors to display the mask area. Then save each individual mask image for individual inspection and verification of each candidate's mask.

[0022] S34: All the deduplicated candidate masks are superimposed on the original image one by one, and different mask areas are distinguished by different colors, so that different detected candidates can be clearly distinguished in the final image.

[0023] Furthermore, the specific steps of step S5 are as follows:

[0024] S51: Reading depth data from the obtained relative depth map, and multiplying the estimated relative depth information by a scaling factor to convert it into actual depth information;

[0025] S52: The depth value of each pixel is converted into three-dimensional point cloud data through the camera internal reference, and the invalid point cloud data is filtered out through the candidate mask segmented by instance segmentation, and only the candidate's point cloud data is retained.

[0026] Furthermore, the specific steps of step S6 are as follows:

[0027] S61: Calculate the spatial position mean of all points in the point cloud corresponding to each candidate and determine the initial center position of each point cloud;

[0028] S62: Calculate the Euclidean distance from each point in the point set to the center position;

[0029] S63: Find the 20% points with the greatest distance and remove these outliers;

[0030] S64: Repeat the process from S61 to S63 multiple times, and further eliminate outliers in each iteration;

[0031] S65: Calculate and save the final point cloud center position.

[0032] Furthermore, the specific steps of step S7 are as follows:

[0033] S71: transforming the point cloud of the right camera into the coordinate system of the left camera according to the relative extrinsic parameter matrix calibrated in step S1;

[0034] S72: If the number of candidate point cloud instances detected by the left and right cameras is the same, pair all point clouds in the converted left camera coordinate system, calculate the average position of the center point of each pair of matched point clouds using the center point of each point cloud optimized in step S6, and output the average position of the candidate's three-dimensional coordinates;

[0035] S73: If the number of candidate point cloud instances detected by the left and right cameras is different, then in the converted left camera coordinate system, for the point cloud in the original left camera coordinate system, find the nearest point cloud from the original right camera point cloud converted to this coordinate system as the matching target, and calculate the average position of the center point of each pair of optimized point clouds, and output it as the three-dimensional coordinates of the candidate.

[0036] Furthermore, the specific steps of step S32 are as follows:

[0037] S321: Calculate the improved IoU, taking into account whether one mask is mostly contained in the other mask;

[0038] S322: Arrange the masks in descending order according to their confidence scores, and then calculate the IoU between each mask and other masks one by one. When the IoU value exceeds a threshold, such as 0.5, these masks are duplicates and are deduplicated.

[0039] Furthermore, the specific steps of step S7 are as follows:

[0040] S71: transforming the point cloud of the right camera into the coordinate system of the left camera according to the relative extrinsic parameter matrix calibrated in step S1;

[0041] S72: When the number of candidate point cloud instances detected by the left and right cameras is the same, all point clouds in the converted left camera coordinate system are paired, and the average position of the center point of each pair of matched point clouds is calculated using the center point of each point cloud optimized in step S6, and the output is the three-dimensional coordinates of the candidate;

[0042] S73: When the number of candidate point cloud instances detected by the left and right cameras is different, in the converted left camera coordinate system, for the point cloud in the original left camera coordinate system, find the nearest point cloud from the original right camera point cloud converted to this coordinate system as the matching target, and calculate the average position of the center point of each pair of optimized point clouds, and output it as the three-dimensional coordinates of the candidate.

[0043] Furthermore, the specific process of step S321 is as follows:

[0044] S3211: Calculate the intersection and union between masks:

[0045] The intersection computes the overlapping portion of two masks:

[0046] Intersection=∑(mask1∩mask2)

[0047] Union computes the union of two masks:

[0048] Union=∑(mask1∪mask2)

[0049] The standard IoU calculation formula is:

[0050]

[0051] S3212: Calculate coverage:

[0052] Coverage measures the overlap of one mask within another:

[0053] Coverage of mask 1:

[0054]

[0055] Coverage of mask 2:

[0056]

[0057] S3213: When the coverage of a mask is greater than 80%, it means that the mask is mostly contained by another mask. In this case, the value with higher coverage is used as the improved IoU value:

[0058] Improved IoU=max(Coverage1,Coverage2)

[0059] Otherwise, return the standard IoU value:

[0060] Improved IoU=Standard IoU.

[0061] Furthermore, the specific process of step S322 is as follows:

[0062] S3221: Sort masks according to confidence scores;

[0063] S3222: Use the improved IoU algorithm to calculate the IoU value of each mask with other masks. When the IoU exceeds a threshold, such as 0.5, these masks are marked as duplicates and masks with lower confidence are removed.

[0064] S3223: Keep the candidate mask after deduplication.

[0065] Compared with the prior art, the present application has the following advantages:

[0066] 1. The multi-camera-based invigilation-free examination room candidate positioning method provided by the present application uses an improved IoU algorithm to calculate the IoU value between masks, thereby enhancing the accuracy of candidate detection.

[0067] 2. The multi-camera-based invigilation-free examination room candidate positioning method provided by the present application realizes matching and positioning of candidates under multiple monocular cameras, thereby improving the intelligent level of the invigilation-free examination room.

[0068] 3. The multi-camera-based invigilation-free examination room candidate positioning method provided by the present application solves the problem that a single camera cannot comprehensively detect all candidates and realizes accurate positioning of candidates in the invigilation-free examination room. BRIEF DESCRIPTION OF DRAWINGS

[0069] In order to more clearly illustrate the technical solutions in the specific embodiments or the prior art, the drawings needed in the specific embodiments or the prior art description will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0070] Figure 1 The flowchart of the multi-camera-based invigilation-free examination room candidate positioning method. DETAILED DESCRIPTION

[0071] The technical solutions of the present application will be described below in conjunction with the drawings. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0072] Embodiment 1

[0073] This embodiment is a specific application of the multi-camera-based invigilation-free examination room candidate positioning method, as shown in Figure 1 The present application discloses a multi-camera-based invigilation-free examination room candidate positioning method, and the specific steps are as follows:

[0074] Step 1): Place the calibration board (checkerboard) in the field of view of two cameras (left front and right front). To ensure calibration accuracy, the calibration board should be clearly visible at different positions and angles. The camera captures images of the calibration board at different angles to capture enough checkerboard feature points, and these images are used to calculate the intrinsic parameters, extrinsic parameters and relative extrinsic parameters of the camera;

[0075] Step 2): Use the OpenCV library to read the calibration images captured by each camera and detect the corners (feature points) of the chessboard in each image. This is done using the cv2.findChessboardCorners function, which returns the sub-pixel coordinates of each corner point on the chessboard.

[0076] Step 3): Use OpenCV's cv2.calibrateCamera function to read the checkerboard images captured by each camera, detect the checkerboard corners (feature points) in each image, and calculate the intrinsic parameters K, rotation matrix R, and translation vector T of each camera.

[0077] Step 4): Calculate the relative extrinsic parameters (relative rotation matrix and relative translation vector) between the two cameras using a pair of checkerboard images captured by the two cameras. These extrinsic parameters can be calculated using the rotation and translation relationship between the two cameras. The relative transformation relationship can be expressed as:

[0078] P left =R rel ×P right +T rel

[0079] Among them, P left It represents the three-dimensional point coordinates in the left camera coordinate system, P right It represents the three-dimensional point coordinates in the right camera coordinate system;

[0080] Step 5): Use OpenCV's cv2.resize function to resize the images captured by the left and right cameras to ensure that subsequent image analysis steps are performed at the same resolution;

[0081] Step 6): Convert the pixel value range from 0 to 255 to 0 to 1 to achieve the purpose of standardizing the input data. The data normalization formula is as follows:

[0082]

[0083] Among them, I represents the original data, I normalized represents the normalized data, min(I) and max(I) represent the minimum and maximum values ​​in data set I, respectively;

[0084] Step 7): Use a three-dimensional block matching filter algorithm to denoise the image, filtering out a range of noise (such as blackboards and trash cans) while preserving the original information of the exam room image as much as possible. The BM3D algorithm uses the self-similarity of images to separate noise from real signals, thereby improving image clarity and quality.

[0085] Step 8) Read the images taken by the left and right cameras, perform instance segmentation on the images, and obtain the mask, label, and confidence score of each detected examinee. Filter out instances with a confidence score below 0.5 and retain high-confidence examinee instances to ensure the reliability and accuracy of the segmentation results.

[0086] Step 9) Calculate the IoU value between the masks using the improved IoU algorithm, remove duplicate examinee masks, and retain high-confidence and non-duplicate examinee masks. The specific steps are as follows:

[0087] Step 9-1) Calculate the improved IoU, considering whether one mask is mostly contained in another mask. The specific process is as follows:

[0088] Step 9-1-1) Calculate the intersection and union between the masks:

[0089] Intersection calculates the overlapping part of two masks:

[0090] Intersection = ∑(mask1 ∩ mask2)

[0091] Union calculates the union of two masks:

[0092] Union = ∑(mask1 ∪ mask2)

[0093] Standard IoU calculation formula:

[0094]

[0095] Step 9-1-2) Calculate the coverage:

[0096] Coverage is used to measure the proportion of overlap of one mask in another mask:

[0097] Coverage of mask 1:

[0098]

[0099] Coverage of mask 2:

[0100]

[0101] Step 9-1-3) If the coverage of one mask is greater than 80%, it means that this mask is mostly contained in another mask. At this time, use the higher coverage value as the improved IoU value:

[0102] Improved IoU = max(Coverage1, Coverage2)

[0103] Otherwise, return the standard IoU value:

[0104] ImprovedIoU=StandardIoU

[0105] Step 9-2): Arrange the masks in descending order based on their confidence scores, and then calculate their IoU with other masks one by one. If the IoU exceeds the threshold (0.5), the masks are considered duplicates and deduplication is performed. The specific process is as follows:

[0106] Step 9-2-1): To avoid low-confidence masks covering or interfering with high-confidence masks, the masks are sorted in descending order according to their confidence scores, ensuring that high-confidence masks are retained first during the deduplication process.

[0107] Step 9-2-2): Use the improved IoU algorithm to calculate the IoU value of each mask with other masks one by one. If the IoU value of two masks exceeds the preset threshold (0.5), these masks are considered highly similar and one of the masks is marked as duplicate and removed;

[0108] Step 9-2-3): Return the candidate mask after deduplication. This mask will reduce the interference of repeated tests and improve the reliability of the overall test results.

[0109] Step 10): Use the ZoeDepth algorithm to estimate the depth of the captured left and right examination room images, and save the generated relative depth map in PFM format. The ZoeDepth algorithm uses the MiDaS encoding-decoding architecture, performs relative depth pre-training through multiple data sets, and learns depth features from images of multiple scenes (indoor and outdoor), but these depths are relative depths, that is, the depth of each pixel is only consistent with the depth of other pixels in the image, and does not have an absolute unit of measurement. Step 11): Convert the obtained relative depth information of the examination room into actual depth information, and convert the examinees segmented from the left and right images into a three-dimensional point cloud through the camera's internal reference. The specific steps are as follows:

[0110] Step 11-1): Read the depth data from the relative depth map obtained by the ZoeDepth algorithm, and multiply the estimated relative depth information by the scaling factor to convert it into actual depth information. The formula is as follows:

[0111] D actual =D relative ×scale

[0112] Among them D actual is the actual depth, D relative is the relative depth, and scale is the scaling factor.

[0113] Step 11-2): Convert the depth value of each pixel into three-dimensional point cloud data through the camera internal reference, and filter out invalid point cloud data based on the candidate mask segmented by instance segmentation, and only retain the point cloud data of the candidate captured in the left and right images.

[0114] Step 12): Calculate the center position of the point cloud of each candidate instance segmented from the left and right images, and continuously optimize the center position. The specific steps are as follows:

[0115] Step 12-1): Calculate the spatial position mean of all points in the point cloud corresponding to each candidate and determine the initial center position of each point cloud;

[0116] Step 12-2): Calculate the distance from each point (x, y, z) in each point cloud to the center position (x center ,y center ,z center ) is calculated as follows:

[0117]

[0118] Step 12-3): Find the 20% of points that are farthest away, then remove these outliers and keep only the 80% of points that are closer to the center of the point cloud;

[0119] Step 12-4): Update the point set to the point set after removing outliers;

[0120] Step 12-5): Repeat the above process 5 times, further eliminating outliers in each iteration;

[0121] Step 12-6): Calculate and save the final point cloud center position.

[0122] Step 13): Convert the point cloud from the right camera to the coordinate system of the left camera, match the point clouds of the candidates under the left and right cameras, and obtain the 3D coordinates of each candidate. The specific steps are as follows:

[0123] Step 13-1): Convert the point cloud of the right camera to the coordinate system of the left camera according to the relative extrinsic parameter matrix calibrated in step 4);

[0124] Step 13-2): If the number of point cloud instances detected by the left and right cameras is the same for the candidate, all point clouds in the converted left camera coordinate system are paired, and the average position of the center points of each pair of matched point clouds is calculated using the center point of each point cloud optimized in step 12), and the output is the three-dimensional coordinates of the candidate;

[0125] Step 13-3): If the number of candidate point cloud instances detected by the left and right cameras is different, then in the converted left camera coordinate system, for the point cloud in the original left camera coordinate system, find the nearest point cloud from the original right camera point cloud converted to this coordinate system as the matching target, and calculate the average position of the center point of each pair of optimized point clouds, and output it as the three-dimensional coordinates of the candidate.

[0126] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will readily appreciate that other variations or modifications based on the above descriptions are possible. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for positioning examinees in an unsupervised examination room based on multiple cameras, characterized in that: The steps include: S1: Fix the positions of the left and right monocular cameras, place a checkerboard in the field of view of the two cameras, and obtain the intrinsic and extrinsic parameters of the two monocular cameras and their relative extrinsic parameters by capturing images of the calibration plate taken at different positions and directions; S2: Perform image preprocessing on the examination room images captured by the left and right cameras, including image resizing, pixel normalization, and image denoising. This includes converting the pixel value range and using a three-dimensional block matching filter algorithm for denoising. This removes a range of noise from the images while preserving the original information of the images. S3: Perform instance segmentation on the examinees in the images captured by the left and right cameras, and perform deduplication processing on the generated examinee masks to achieve examinee recognition and segmentation; S4: Use the ZoeDepth algorithm to estimate the depth of the captured left and right examination room images, and save the generated relative depth map in PFM format; S5: Process the obtained depth map of the examination room and convert the examinees segmented from the left and right images into a three-dimensional point cloud using the camera's intrinsic reference; S6: Calculate the center position of the point cloud of each examinee segmented from the left and right images and optimize the center position; S7: Convert the point cloud of the right camera to the coordinate system of the left camera, match the point clouds of the examinees under the left and right cameras, and obtain the three-dimensional coordinates of each examinee; The specific steps of step S7 are as follows: S71: transforming the point cloud of the right camera into the coordinate system of the left camera according to the relative extrinsic parameter matrix calibrated in step S1; S72: If the number of candidate point cloud instances detected by the left and right cameras is the same, pair all point clouds in the converted left camera coordinate system, calculate the average position of the center point of each pair of matched point clouds using the center point of each point cloud optimized in step S6, and output the average position of the candidate's three-dimensional coordinates; S73: If the number of candidate point cloud instances detected by the left and right cameras is different, then in the converted left camera coordinate system, for the point cloud in the original left camera coordinate system, find the nearest point cloud from the original right camera point cloud converted to this coordinate system as the matching target, and calculate the average position of the center point of each pair of optimized point clouds, and output it as the three-dimensional coordinates of the candidate.

2. The method for positioning examinees in an unmanned examination room based on multiple cameras according to claim 1 is characterized in that: The specific steps of step S1 are as follows: S11: Place a checkerboard in the left and right front fields of view of the two cameras, and capture images of the checkerboard from multiple angles. S12: Read the checkerboard images captured by each camera using the OpenCV computer vision library, detect the corner points of the checkerboard in each image, and calculate the intrinsic parameter K, rotation matrix R, and translation vector T of each camera; S13: Calculate the relative external parameters of the two cameras through a pair of checkerboard images taken by the two cameras to obtain the relative rotation matrix R between the two cameras rel and the relative translation vector T rel .

3. The method for positioning examinees in an unmanned examination room based on multiple cameras according to claim 1 is characterized in that: The specific steps of step S3 are as follows: S31: Read the examination room images captured by the left and right cameras and perform instance segmentation to obtain the mask, label, and confidence score of each detected examinee. Filter out examinee instances with a confidence score lower than 0.

5. From the filtered results, only examinee masks and corresponding confidence scores are selected to ensure that only examinee instances with higher confidence scores are retained. S32: Use the improved IoU algorithm to calculate the IoU between masks and retain the candidate masks with high confidence and no duplication; S33: Create a completely black background of the same size as the original image, superimpose the mask onto the background, and use random colors to display the mask area. Then save each individual mask image for individual inspection and verification of each candidate's mask. S34: Superimpose all the duplicated candidate masks on the original image one by one, and use different colors to distinguish different mask areas.

4. The method for positioning examinees in an unsupervised examination room based on multiple cameras according to claim 1 is characterized in that: The specific steps of step S5 are as follows: S51: Reading depth data from the obtained relative depth map, and multiplying the estimated relative depth information by a scaling factor to convert it into actual depth information; S52: The depth value of each pixel is converted into three-dimensional point cloud data through the camera internal reference, and the invalid point cloud data is filtered out through the candidate mask segmented by instance segmentation, and only the candidate's point cloud data is retained.

5. The method for positioning examinees in an unmanned examination room based on multiple cameras according to claim 1 is characterized in that: The specific steps of step S6 are as follows: S61: Calculate the spatial position mean of all points in the point cloud corresponding to each candidate and determine the initial center position of each point cloud; S62: Calculate the Euclidean distance from each point in the point set to the center position; S63: Find the 20% points with the greatest distance and remove these outliers; S64: Repeat the process from S61 to S63 multiple times, and further eliminate outliers in each iteration; S65: Calculate and save the final point cloud center position.

6. The method for positioning examinees in an unmanned examination room based on multiple cameras according to claim 3 is characterized in that: The specific steps of step S32 are as follows: S321: Calculate the improved IoU, taking into account whether one mask is mostly contained in the other mask; S322: Arrange the masks in descending order according to their confidence scores, and then calculate the IoU between each mask and other masks one by one. When the IoU value exceeds a threshold, these masks are duplicated and deduplication is performed.

7. The method for positioning examinees in an unsupervised examination room based on multiple cameras according to claim 6 is characterized in that: The specific process of step S321 is as follows: S3211: Calculate the intersection and union between masks: The intersection computes the overlapping portion of two masks: Intersection=∑(mask1∩mask2) Union computes the union of two masks: Union=∑(mask1∪mask2) The standard IoU calculation formula is: S3212: Calculate coverage: Coverage measures the overlap of one mask within another: Coverage of mask 1: Coverage of mask 2: S3213: When the coverage of a mask is greater than 80%, it means that the mask is mostly contained by another mask. In this case, the value with higher coverage is used as the improved IoU value: Improved IoU=max(Coverage1,Coverage2) Otherwise, return the standard IoU value: Improved IoU=Standard IoU.

8. The method for positioning examinees in an unsupervised examination room based on multiple cameras according to claim 6 is characterized in that: The specific process of step S322 is as follows: S3221: Sort masks according to confidence scores; S3222: Use the improved IoU algorithm to calculate the IoU value of each mask with other masks. When the IoU exceeds the threshold, these masks are marked as duplicates and masks with lower confidence are removed; S3223: Keep the candidate mask after deduplication.

Images