Image registration-based underground ROI monitoring area dynamic calibration method and system

Abstract

The application provides a kind of based on image registration's underground garage ROI monitoring area dynamic calibration method and system, method includes: step S1: the monitoring picture of the camera obtained is regarded as reference image, and ROI region is demarcated;Step S2: the feature points of current picture and reference image are extracted and matched, and conversion matrix is calculated based on the feature point pair of successful matching;Step S3: the ROI region of reference image is mapped to current picture based on the conversion matrix to complete the dynamic calibration of ROI region.The application calculates the transformation relationship of current picture and reference picture in real time by image matching technology, realizes the automatic correction of ROI region, improves the accuracy and robustness of monitoring system, reduces the demand of artificial maintenance, solves the problem of monitoring area inaccuracy caused by the position change of underground garage entrance camera.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method and system for dynamic calibration of ROI monitoring areas in underground parking lots based on image registration. Background Technology

[0002] With the advancement of smart city construction, computer vision-based intelligent traffic monitoring systems have been widely applied in underground parking garage entrance and exit management. These systems typically rely on pre-defined Regions of Interest (ROIs) within the monitored area for vehicle detection, tracking, and behavior analysis. Most existing technologies employ a fixed ROI approach, where the monitoring area is manually calibrated during system initialization, and subsequent analyses are based on this fixed area. This method works effectively when the camera installation location remains unchanged and has become the mainstream technical solution for intelligent monitoring of underground parking garage entrances and exits.

[0003] While existing parking garage entrance / exit monitoring systems achieve vehicle monitoring through fixed Regions of Interest (ROIs), their technical limitations inevitably lead to the following problems: First, existing solutions manually calibrate the ROI during initialization and then store its coordinates, directly retrieving these fixed coordinates for subsequent analysis. This fixed coordinate usage means that slight displacement or angular shifts in the camera due to wind, equipment adjustments, or long-term vibration can cause a deviation between the pre-calibrated ROI and the actual monitored area. Second, existing technologies lack automatic calibration mechanisms to detect and compensate for this offset. Continuing to use the deviated ROI for vehicle detection and tracking directly results in inaccurate vehicle counting, increased speed calculation errors, and false or missed alarms, severely impacting the accuracy and reliability of the monitoring system. Furthermore, to address this issue, existing technologies rely on periodic manual inspection and recalibration, increasing system maintenance costs and the potential for human error.

[0004] The purpose of this invention is to provide an intelligent method capable of automatically detecting camera offset and dynamically calibrating the Region of Interest (ROI), solving the problem of inaccurate monitoring areas caused by changes in the position of cameras at parking garage entrances and exits. This invention aims to use image matching technology to calculate the transformation relationship between the current image and the reference image in real time, achieving automatic correction of the ROI, improving the accuracy and robustness of the monitoring system, and reducing the need for manual maintenance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a dynamic calibration method and system for monitoring ROI areas in underground parking lots based on image registration.

[0006] A dynamic calibration method for monitoring ROI (Region of Interest) areas of a basement based on image registration, provided by the present invention, includes:

[0007] Step S1: Use the acquired camera footage as a reference image to define the ROI region;

[0008] Step S2: Extract feature points from the current image and the reference image and match them. Calculate the transformation matrix based on the successfully matched feature point pairs.

[0009] Step S3: Based on the transformation matrix, map the ROI region of the reference image to the current image to complete the dynamic calibration of the ROI region.

[0010] Preferred options also include:

[0011] Step S4: When the number of matching points is lower than the preset threshold or the mapped ROI exceeds the image boundary, a calibration anomaly alarm is issued, and the previous valid transformation result is adopted.

[0012] Preferably, step S2 includes the following sub-steps:

[0013] Step S2.1: Use the improved FAST algorithm to extract stable feature points from the current image and the reference image;

[0014] Step S2.2: Cross-validation is introduced by calculating the Euclidean distance between the feature descriptors of the reference image and the current image;

[0015] Step S2.3: Randomly select a certain number of successfully matched feature point pairs as sample data, and calculate and obtain the transformation matrix.

[0016] Preferably, step S2.1 includes:

[0017] The size of the input image is unified by bilinear interpolation, and the threshold T is obtained by the maximum inter-class variance algorithm.

[0018] The proportion of foreground pixels less than the threshold T is calculated using the following formula. and its average gray value The proportion of background pixels greater than the threshold T and its average gray value The total average gray value of the image and inter-class variance Finally, the optimal threshold was calculated. This maximizes the inter-class variance:

[0019]

[0020] For each pixel P, compare it with the surrounding pixels. If the absolute value of the difference between the surrounding pixels and pixel P exceeds the threshold T, then it is classified into the candidate key point set.

[0021] By calculating the response value of each candidate keypoint and comparing it with its adjacent left and right points, if the keypoint has the largest response value, it is considered a maximum and is retained. Otherwise, the keypoint is discarded.

[0022] Then, by combining the SIFT operator to generate feature descriptors, each keypoint is determined. The main direction is determined by selecting the main peak value as the main direction of the feature point and storing the magnitude. and direction :

[0023]

[0024]

[0025] in, The spatial value of the key point.

[0026] Preferably, step S2.2 includes:

[0027] Assume a feature point in the reference image is It uses brute force matching to traverse all feature points in the current image and calculates the sum of the points. Given the distance, find the point with the smallest distance, and denote it as . ;

[0028] Traverse all feature points in the reference image and calculate the sum of the points. The distance, if the point corresponding to the minimum distance returned is If the cross-validation passes, then... and Treat them as a candidate matching pair; otherwise, discard them, and repeat this process until all matching pairs have been cross-validated.

[0029] For the above candidate matching pairs, assume The shortest distance to return. For the next closest distance, the following formula is used to filter out matches with a higher probability of mismatch:

[0030]

[0031] in, This is an empirical threshold.

[0032] Preferably, step S2.3 includes:

[0033] Four pairs of successfully matched feature points are randomly selected as sample data, and the transformation matrix between each matching pair is calculated. and corresponding models Substitute the remaining matching pairs into the obtained model. In this process, the error values ​​between them are calculated, and those with differences less than the empirical values ​​are recorded as inliers. This process is repeated until a specified number of iterations is reached. The model with the most inliers is selected as the best matching model. The least squares method is used to perform regression fitting on the inlier set corresponding to the best matching model to obtain the optimized transformation matrix. .

[0034] Preferably, step S3 includes:

[0035] Assume that one of the ROI points in the baseline image is The current image pixels after perspective transformation are Then their relationship is expressed by the following formula:

[0036]

[0037] Calculations show that:

[0038]

[0039] in, and For rotation transformation parameters, and These are the scaling transformation parameters. and These are the parameters for the translation transformation. and These are the perspective transformation parameters. The value is 1; based on this, the ROI coordinates of the current image after transformation are obtained.

[0040] A dynamic calibration system for monitoring ROI (Region of Interest) areas of a basement based on image registration, provided by the present invention, includes:

[0041] Module M1: Uses the acquired camera footage as a reference image to define the ROI region;

[0042] Module M2: Extracts feature points from the current image and the reference image and performs matching; calculates the transformation matrix based on the successfully matched feature point pairs.

[0043] Module M3: Based on the transformation matrix, it maps the ROI region of the reference image to the current image to complete the dynamic calibration of the ROI region.

[0044] Preferred options also include:

[0045] Module M4: When the number of matching points is lower than the preset threshold or the mapped ROI exceeds the image boundary, a calibration abnormality alarm is issued, and the previous valid transformation result is adopted.

[0046] Preferably, module M2 includes the following sub-modules:

[0047] Module M2.1: Uses an improved FAST algorithm to extract stable feature points from the current image and the reference image;

[0048] Module M2.2: Introduces cross-validation by calculating the Euclidean distance between the feature descriptors of the baseline image and the current image;

[0049] Module M2.3: Randomly select a certain number of successfully matched feature point pairs as sample data and calculate the transformation matrix.

[0050] Preferably, module M2.1 includes:

[0051] The size of the input image is unified by bilinear interpolation, and the threshold T is obtained by the maximum inter-class variance algorithm.

[0052] The proportion of foreground pixels less than the threshold T is calculated using the following formula. and its average gray value The proportion of background pixels greater than the threshold T and its average gray value The total average gray value of the image and inter-class variance Finally, the optimal threshold was calculated. This maximizes the inter-class variance:

[0053]

[0054] For each pixel P, compare it with the surrounding pixels. If the absolute value of the difference between the surrounding pixels and pixel P exceeds the threshold T, then it is classified into the candidate key point set.

[0055] By calculating the response value of each candidate keypoint and comparing it with its adjacent left and right points, if the keypoint has the largest response value, it is considered a maximum and is retained. Otherwise, the keypoint is discarded.

[0056] Then, by combining the SIFT operator to generate feature descriptors, each keypoint is determined. The main direction is determined by selecting the main peak value as the main direction of the feature point and storing the magnitude. and direction :

[0057]

[0058]

[0059] in, The spatial value of the key point.

[0060] Preferably, the module M2.2 includes:

[0061] Assume a feature point in the reference image is It uses brute force matching to traverse all feature points in the current image and calculates the sum of the points. Given the distance, find the point with the smallest distance, and denote it as . ;

[0062] Traverse all feature points in the reference image and calculate the sum of the points. The distance, if the point corresponding to the minimum distance returned is If the cross-validation passes, then... and Treat them as a candidate matching pair; otherwise, discard them, and repeat this process until all matching pairs have been cross-validated.

[0063] For the above candidate matching pairs, assume The shortest distance to return. For the next closest distance, the following formula is used to filter out matches with a higher probability of mismatch:

[0064]

[0065] in, This is an empirical threshold.

[0066] Preferably, the module M2.3 includes:

[0067] Four pairs of successfully matched feature points are randomly selected as sample data, and the transformation matrix between each matching pair is calculated. and corresponding models Substitute the remaining matching pairs into the obtained model. In this process, the error values ​​between them are calculated, and those with differences less than the empirical values ​​are recorded as inliers. This process is repeated until a specified number of iterations is reached. The model with the most inliers is selected as the best matching model. The least squares method is used to perform regression fitting on the inlier set corresponding to the best matching model to obtain the optimized transformation matrix. .

[0068] Preferably, the module M3 includes:

[0069] Assume that one of the ROI points in the baseline image is The current image pixels after perspective transformation are Then their relationship is expressed by the following formula:

[0070]

[0071] Calculations show that:

[0072]

[0073] in, and For rotation transformation parameters, and These are the scaling transformation parameters. and These are the parameters for the translation transformation. and These are the perspective transformation parameters. The value is 1; based on this, the ROI coordinates of the current image after transformation are obtained.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] 1. This invention uses image matching technology to calculate the transformation relationship between the current image and the reference image in real time, realizes automatic correction of the ROI area, improves the accuracy and robustness of the monitoring system, reduces the need for manual maintenance, and solves the problem of inaccurate monitoring area caused by changes in the position of cameras at the entrance and exit of the parking lot.

[0076] 2. This invention achieves precise dynamic calibration of the ROI region through automated feature matching and transformation matrix calculation, controlling the deviation of the monitoring area within the pixel range, and greatly improving the accuracy of vehicle detection and traffic parameter calculation.

[0077] 3. This invention possesses strong adaptive capabilities, effectively compensating for camera misalignment caused by various factors, reducing system performance degradation due to equipment displacement, and enhancing system robustness and reliability. Furthermore, this invention avoids frequent manual intervention and maintenance, significantly reducing system operating costs and improving the level of intelligence.

[0078] 4. While achieving a calibration accuracy of over 95%, this invention only increases processing time by 50-60ms, fully meeting the needs of real-time monitoring of underground parking garage entrances and exits, and providing a more reliable technical guarantee for intelligent traffic management systems. Attached Figure Description

[0079] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0080] Figure 1 This is a schematic diagram of the dynamic calibration method for monitoring ROI areas in a basement based on image registration according to the present invention.

[0081] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0082] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0083] Reference Figure 1 and Figure 2 As shown, a dynamic calibration method for monitoring ROI (Region of Interest) areas in a basement based on image registration includes:

[0084] Step S1: Use the acquired camera footage as a reference image to define the ROI region;

[0085] Step S2: Extract feature points from the current image and the reference image and match them. Calculate the transformation matrix based on the successfully matched feature point pairs.

[0086] Step S3: Based on the transformation matrix, map the ROI region of the reference image to the current image to complete the dynamic calibration of the ROI region.

[0087] Step S4: When the number of matching points is lower than the preset threshold or the mapped ROI exceeds the image boundary, a calibration anomaly alarm is issued, and the previous valid transformation result is adopted.

[0088] In one specific embodiment, the specific steps are as follows:

[0089] S1. Reference Image Selection: During the system initialization phase, the monitoring images from the cameras are acquired as reference images, and the ROI area is manually calibrated.

[0090] S2. Establishing baseline features: Based on the current image and the baseline image, use the improved FAST algorithm to extract stable feature points (such as underground building structures, fixed landmarks, ground textures, etc.) in the image.

[0091] 1) First, the image size is unified using bilinear interpolation. Assume the source image size is... The target image size is The transformation relation is: Therefore, according to this correspondence, the pixels of the transformed image can be obtained.

[0092] 2) The threshold T is obtained using the Otsu's Maximum Inter-Class Variance (OTSU) algorithm. To address the empirical and fixed threshold issues of the traditional FAST algorithm, the OTSU algorithm is introduced to adaptively obtain the threshold T for different frames of the image. This threshold can optimally distinguish between foreground and background. The proportion of foreground pixels smaller than the threshold T is calculated using the following formula. and its average gray value The proportion of background pixels greater than the threshold T and its average gray value The total average gray value of the image and inter-class variance Finally, the optimal threshold was calculated. This maximizes the inter-class variance.

[0093]

[0094] 3) Obtain candidate corner points. For each pixel P, compare it with the surrounding pixels. If there are at least 3 / 4 surrounding pixels whose absolute difference with pixel P exceeds the threshold T, then it is included in the candidate key point set.

[0095] 4) Non-maximum suppression yields local optima. By calculating the response value of each candidate keypoint and comparing it with its adjacent left and right points, if the keypoint has the largest response value, it is a maximum and is retained. Otherwise, the point is discarded.

[0096] 5) Combine the SIFT operator to generate feature descriptors and determine each key point. The main direction is determined by selecting the main peak value as the main direction of the feature point and storing the magnitude. and direction .

[0097]

[0098]

[0099] in, The spatial value of the key point.

[0100] S3. Real-time Feature Matching: By calculating the Euclidean distance between the baseline image and the feature descriptors of the currently captured image, cross-validation is introduced to reduce false matching. The specific steps are as follows:

[0101] Assume a feature point in the reference image is First, a brute-force matching process is used to traverse all feature points in the currently captured image and calculate the sum of the points. Find the point with the smallest distance from the given distance, i.e., the point with the highest similarity, and denote it as . Traverse all feature points in the reference image and calculate the relationship between them and the points. The distance, if the point corresponding to the minimum distance returned is If the cross-validation passes, then... and Treat each pair as a candidate match. Otherwise, discard them, and repeat this process until all match pairs have undergone cross-validation. For the above candidate match pairs, assume... The shortest distance to return. For the next closest distance, the following formula is used to filter out matches with a higher probability of mismatch:

[0102]

[0103] in, This is an empirical threshold. In this example, the value is 0.65. and The larger the ratio, the closer the nearest and second-nearest feature points are, and the greater the possibility of a mismatch. Conversely, the smaller the ratio, the farther the nearest and second-nearest feature points are, and the less likely a mismatch is. This is a threshold value; if it's too large, it increases the likelihood of false matches. If it's too small, it may result in fewer matching pairs. It needs to be set appropriately. The magnitude of the value.

[0104] S4. Transformation Matrix Calculation: Based on the successfully matched feature point pairs, randomly select 4 pairs as sample data. Note that these four sample pairs cannot be collinear. Calculate the transformation matrix between these 4 matched pairs. To obtain the corresponding model Substitute the remaining matching pairs into the obtained model. In this process, the error value between them is calculated, and those with a difference less than the empirical value are recorded as inliers. This process is repeated until a specified number of iterations is reached, and the model with the most inliers is selected as the best matching model. Finally, the least squares method is used to perform regression fitting on the inlier set corresponding to the best matching model to obtain the optimized transformation matrix. .

[0105] S5, ROI Dynamic Mapping: Assuming that one of the ROI points in the baseline image is... The current image pixels after perspective transformation are Then their relationship can be expressed by the following formula:

[0106]

[0107] Calculations show that:

[0108]

[0109] in, and For rotation transformation parameters, and These are the scaling transformation parameters. and These are the parameters for the translation transformation. and These are the perspective transformation parameters. The value is 1; based on this, the ROI coordinates of the current image after transformation are obtained.

[0110] S6. Exception Handling and Update: When the number of matching points is lower than the threshold N. min If the mapped ROI exceeds the image boundary, the system issues a calibration anomaly alarm and adopts the previous valid transformation result. The system supports periodic updates of baseline features to adapt to environmental changes.

[0111] The present invention also provides a dynamic calibration system for parking lot ROI monitoring areas based on image registration. The dynamic calibration system for parking lot ROI monitoring areas based on image registration can be implemented by executing the process steps of the dynamic calibration method for parking lot ROI monitoring areas based on image registration. That is, those skilled in the art can understand the dynamic calibration method for parking lot ROI monitoring areas based on image registration as a preferred embodiment of the dynamic calibration system for parking lot ROI monitoring areas based on image registration.

[0112] Specifically, a dynamic calibration system for monitoring the ROI of a basement parking lot based on image registration includes:

[0113] Module M1: Uses the acquired camera footage as a reference image to define the ROI region;

[0114] Module M2: Extracts feature points from the current image and the reference image and performs matching; calculates the transformation matrix based on the successfully matched feature point pairs.

[0115] Module M3: Based on the transformation matrix, it maps the ROI region of the reference image to the current image to complete the dynamic calibration of the ROI region.

[0116] Also includes:

[0117] Module M4: When the number of matching points is lower than the preset threshold or the mapped ROI exceeds the image boundary, a calibration abnormality alarm is issued, and the previous valid transformation result is adopted.

[0118] Module M2 includes the following sub-modules:

[0119] Module M2.1: Uses an improved FAST algorithm to extract stable feature points from the current image and the reference image;

[0120] Module M2.2: Introduces cross-validation by calculating the Euclidean distance between the feature descriptors of the baseline image and the current image;

[0121] Module M2.3: Randomly select a certain number of successfully matched feature point pairs as sample data and calculate the transformation matrix.

[0122] The module M2.1 includes:

[0123] The size of the input image is unified by bilinear interpolation, and the threshold T is obtained by the maximum inter-class variance algorithm.

[0124] The proportion of foreground pixels less than the threshold T is calculated using the following formula. and its average gray value The proportion of background pixels greater than the threshold T and its average gray value The total average gray value of the image and inter-class variance Finally, the optimal threshold was calculated. This maximizes the inter-class variance:

[0125]

[0126] For each pixel P, compare it with the surrounding pixels. If the absolute value of the difference between the surrounding pixels and pixel P exceeds the threshold T, then it is classified into the candidate key point set.

[0127] By calculating the response value of each candidate keypoint and comparing it with its adjacent left and right points, if the keypoint has the largest response value, it is considered a maximum and is retained. Otherwise, the keypoint is discarded.

[0128] Then, by combining the SIFT operator to generate feature descriptors, each keypoint is determined. The main direction is determined by selecting the main peak value as the main direction of the feature point and storing the magnitude. and direction :

[0129]

[0130]

[0131] in, The spatial value of the key point.

[0132] The module M2.2 includes:

[0133] Assume a feature point in the reference image is It uses brute force matching to traverse all feature points in the current image and calculates the sum of the points. Given the distance, find the point with the smallest distance, and denote it as . ;

[0134] Traverse all feature points in the reference image and calculate the sum of the points. The distance, if the point corresponding to the minimum distance returned is If the cross-validation passes, then... and Treat them as a candidate matching pair; otherwise, discard them, and repeat this process until all matching pairs have been cross-validated.

[0135] For the above candidate matching pairs, assume The shortest distance to return. For the next closest distance, the following formula is used to filter out matches with a higher probability of mismatch:

[0136]

[0137] in, This is an empirical threshold.

[0138] The module M2.3 includes:

[0139] Four pairs of successfully matched feature points are randomly selected as sample data, and the transformation matrix between each matching pair is calculated. and corresponding models Substitute the remaining matching pairs into the obtained model. In this process, the error values ​​between them are calculated, and those with differences less than the empirical values ​​are recorded as inliers. This process is repeated until a specified number of iterations is reached. The model with the most inliers is selected as the best matching model. The least squares method is used to perform regression fitting on the inlier set corresponding to the best matching model to obtain the optimized transformation matrix. .

[0140] The module M3 includes:

[0141] Assume that one of the ROI points in the baseline image is The current image pixels after perspective transformation are Then their relationship is expressed by the following formula:

[0142]

[0143] Calculations show that:

[0144]

[0145] in, and For rotation transformation parameters, and These are the scaling transformation parameters. and These are the parameters for the translation transformation. and These are the perspective transformation parameters. The value is 1; based on this, the ROI coordinates of the current image after transformation are obtained.

[0146] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0147] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for dynamic calibration of a monitoring region of interest (ROI) in a mine based on image registration, characterized in that, include: Step S1: Use the acquired camera footage as a reference image to define the ROI region; Step S2: Extract feature points from the current image and the reference image and match them. Calculate the transformation matrix based on the successfully matched feature point pairs. Step S3: Based on the transformation matrix, map the ROI region of the reference image to the current image to complete the dynamic calibration of the ROI region; Step S2 includes the following sub-steps: Step S2.1: Use the improved FAST algorithm to extract stable feature points from the current image and the reference image; Step S2.2: Cross-validation is introduced by calculating the Euclidean distance between the feature descriptors of the reference image and the current image; Step S2.3: Randomly select a certain number of successfully matched feature point pairs as sample data, and calculate and obtain the transformation matrix; Step S2.1 includes: The size of the input image is unified by bilinear interpolation, and the threshold T is obtained by the maximum inter-class variance algorithm. The proportion of foreground pixels less than the threshold T is calculated using the following formula. and its average gray value The proportion of background pixels greater than the threshold T and its average gray value The total average gray value of the image and inter-class variance Finally, the optimal threshold was calculated. This maximizes the inter-class variance: Then, by combining the SIFT operator to generate feature descriptors, each keypoint is determined. The main direction is determined by selecting the main peak value as the main direction of the feature point and storing the magnitude. and direction : wherein spatial value of a key point; Step S2.3 includes: Four pairs of successfully matched feature points are randomly selected as sample data, and the transformation matrix between each matching pair is calculated. and corresponding models Substitute the remaining matching pairs into the obtained model. In this process, the error values ​​between them are calculated, and those with differences less than the empirical values ​​are recorded as inliers. This process is repeated until a specified number of iterations is reached. The model with the most inliers is selected as the best matching model. The least squares method is used to perform regression fitting on the inlier set corresponding to the best matching model to obtain the optimized transformation matrix. .

2. The image registration based dynamic calibration method of a geocell ROI monitoring region according to claim 1, wherein, Also includes: Step S4: When the number of matching points is lower than the preset threshold or the mapped ROI exceeds the image boundary, a calibration abnormality alarm is issued, and the previous valid transformation result is adopted.

3. The image registration based in-mine ROI monitoring area dynamic calibration method according to claim 1, characterized in that, Step S2.2 includes: Assume one feature point in the reference image is , by brute force matching all the feature points in the current image and calculating the distance with point , find out the minimum distance point, recorded as ; Traverse all feature points in the reference image and calculate the sum of the points. The distance, if the point corresponding to the minimum distance returned is If the cross-validation passes, then... and Treat them as a candidate matching pair; otherwise, discard them, and repeat this process until all matching pairs have been cross-validated. For the above candidate matched pairs, assume is the nearest distance returned, is the next nearest distance, further filter out false matched pairs by the following equation: wherein is an empirical threshold value.

4. The image registration based in-mine ROI monitoring area dynamic calibration method according to claim 1, characterized in that, Step S3 includes: Assume one of the ROI points in the reference image is , and the pixel point of the current image after perspective transformation is , then the relationship is represented by the following formula: Calculations show that: wherein and are rotation transformation parameters, and are scale transformation parameters, and are translation transformation parameters, and are perspective transformation parameters, the value is 1; according to which the ROI coordinates of the transformed current picture image are obtained.

5. A dynamic calibration system for monitoring ROI (Region of Interest) areas in underground parking lots based on image registration, characterized in that, include: Module M1: Uses the acquired camera footage as a reference image to define the ROI region; Module M2: Extracts feature points from the current image and the reference image and performs matching; calculates the transformation matrix based on the successfully matched feature point pairs. Module M3: Based on the transformation matrix, the ROI region of the reference image is mapped to the current image to complete the dynamic calibration of the ROI region; The module M2 includes the following sub-steps: Module M2.1: Uses an improved FAST algorithm to extract stable feature points from the current image and the reference image; Module M2.2: Introduces cross-validation by calculating the Euclidean distance between the feature descriptors of the baseline image and the current image; Module M2.3: Randomly select a certain number of successfully matched feature point pairs as sample data, and calculate and obtain the transformation matrix; The module M2.1 includes: The size of the input image is unified by bilinear interpolation, and the threshold T is obtained by the maximum inter-class variance algorithm. The proportion of foreground pixel points less than the threshold T is calculated by the following formula and the average gray value thereof The proportion of background pixel points greater than the threshold T is calculated by the following formula and the average gray value thereof The total average gray value of the image is calculated by the following formula and the inter-class variance is calculated by the following formula The optimal threshold T is finally calculated so as to maximize the inter-class variance ​ Then, by combining the SIFT operator to generate feature descriptors, each keypoint is determined. The main direction is determined by selecting the main peak value as the main direction of the feature point and storing the magnitude. and direction : in, Spatial values ​​for key points; The module M2.3 includes: Four pairs of successfully matched feature points are randomly selected as sample data, and the transformation matrix between each matching pair is calculated. and corresponding models Substitute the remaining matching pairs into the obtained model. In this process, the error values ​​between them are calculated, and those with differences less than the empirical values ​​are recorded as inliers. This process is repeated until a specified number of iterations is reached. The model with the most inliers is selected as the best matching model. The least squares method is used to perform regression fitting on the inlier set corresponding to the best matching model to obtain the optimized transformation matrix. .

6. The image registration-based dynamic calibration system for monitoring ROI areas in underground parking lots according to claim 5, characterized in that, Also includes: Module M4: When the number of matching points is lower than the preset threshold or the mapped ROI exceeds the image boundary, a calibration abnormality alarm is issued, and the previous valid transformation result is adopted.

Images