A method for identifying and reconstructing feature points of moving object images based on multi-viewing
Through multi-eye vision technology and feature point matching method, traditional binocular stereo vision has solved the limited field of view and matching difficulties in moving objects and texture monotonic scenes, achieving wider applicability and high accuracy of three-dimensional reconstruction.
Patent Information
- Application Number
- CN202111210015.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-18
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-10-18
AI Technical Summary
Traditional binocular stereo vision has problems such as limited field of view and difficulty in matching when dealing with moving objects and monotonic texture scenes, making it difficult to achieve accurate three-dimensional reconstruction.
Multi-eye vision is used for image capture, the camera is calibrated through Zhang Zhengyou calibration method, distortion correction and color space conversion are performed, feature points and auxiliary recognition points in the ROI area are extracted, and marking straight lines are constructed to sort and match feature points. Reconstruct the three-dimensional coordinates.
It realizes coverage of a larger field of view and accurate capture of target objects, improves recognition effectiveness and accuracy, and is suitable for three-dimensional reconstruction and dynamic target recognition of complex scenes.
Smart Images

Figure CN114004895B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for identifying and reconstructing target points in an image, and in particular to a method for identifying and reconstructing feature points of a moving object image based on multi-viewing. Background Art
[0002] Stereo vision technology is an important technology in the field of computer vision. The purpose of this technology is to reconstruct the three-dimensional geometric information of the scene. At present, this technology has been applied to the fields of industrial automation, medicine, construction, etc., and its application range is very wide. The main principle of this technology is to obtain camera parameters through calibration, shoot the target image, and convert the two-dimensional information in the image into three-dimensional information in space. This technology has many advantages, such as high computational efficiency, fast processing speed, relatively appropriate accuracy, low cost, no need to contact the target object, etc. It is very suitable for product recognition and detection, industrial automation equipment recognition and detection, etc.
[0003] However, as the scope of application becomes wider and wider, the requirements for stereo vision technology are also increasing. For scenes where the target object is a moving object, traditional binocular stereo vision has some shortcomings. Traditional binocular stereo vision uses two cameras to shoot the same target object, and uses the parallax between the two pictures taken at the same time to calculate the coordinates. However, when applied to moving objects, it is not applicable. The main shortcomings are as follows:
[0004] (1) The general camera is fixedly installed and uses binocular vision, which has a limited field of view and cannot obtain the situation outside the field of view. Therefore, it is generally only used to identify static objects. If the target object is a moving object, the traditional method is difficult to use normally;
[0005] (2) Traditional binocular vision methods are not suitable for scenes that are monotonous and lack texture. If the texture of the captured target is single, matching will be difficult and the three-dimensional situation of the scene and target cannot be accurately restored. In moving scenes, the contour and texture will change, which is even more unsuitable for traditional methods.
[0006] In view of the above situation, for moving objects, a larger field of view is needed to ensure that the photographed object is always within the measurement range, and a new recognition method is needed to better identify the target object, which is impossible to achieve with traditional stereo vision. Summary of the invention
[0007] In order to solve the problems existing in the background technology, the purpose of the present invention is to provide a method for identifying and reconstructing feature points of moving object images based on multi-eye vision, which solves the problem that the field of view is limited in general traditional stereo vision and it is almost impossible to apply it to moving objects with lack of texture. The recognition method of the present invention uses multi-eye vision to capture images, which can significantly improve the recognition acquisition rate of effective images.
[0008] In order to achieve the above object, the technical solution adopted by the present invention is:
[0009] Step (1): Use Zhang Zhengyou calibration method to calibrate the multi-view camera;
[0010] Step (2): read the image captured by the multi-camera of the moving object, perform distortion correction on the image, pre-process the Gaussian noise reduction, and convert the image color space into Lab space;
[0011] Step (3): obtaining color sampling data of the background and ROI area, dividing the image into the background area and the ROI area according to the color sampling data, and segmenting the ROI area from the image;
[0012] Step (4): According to human-computer interaction, different feature points and auxiliary recognition points in the recognition ROI area are selected to obtain color sampling values of the feature points and the auxiliary recognition points, and the feature point area and the auxiliary recognition point area are identified according to the color sampling values of the feature points and the auxiliary recognition points, and then the number of feature points and auxiliary recognition points is obtained, and the number is compared to obtain all feature points and auxiliary recognition points;
[0013] Step (5): construct a marked line using the auxiliary recognition points, and sort the feature points from near to far according to the distance from the point to the line;
[0014] Step (6): The actual coordinates of the feature points are calculated from the different image coordinates of the same feature point in different images taken by the multi-camera at the same time to perform three-dimensional reconstruction.
[0015] The present invention completes the accurate extraction of feature points in the ROI area in the image and the matching between multiple cameras through the processing of the above steps (3) to (5), thereby achieving accurate capture of the target object.
[0016] In the step (2), the conversion from RGB color space to Lab color space uses XYZ color space, and the specific calculation formula is as follows:
[0017]
[0018] R * =R / 255, G * =G / 255,B * =B / 255
[0019]
[0020]
[0021]
[0022]
[0023] Among them, R, G, and B represent the RGB values of the pixels in the original image, respectively. * , G * , B * Respectively represent the scaling of RGB values to values between 0 and 1, X, Y, and Z represent the XYZ values of the imaginary three primary colors in the XYZ color space; L, a, and b represent the three values in the Lab color space, where L represents brightness, a represents the component from green to red, and b represents the component from blue to yellow; H represents the coefficient matrix, X n , Y n , Z n is the sum of the coefficients in the first row, the second row, and the third row of the coefficient matrix H in the matrix operation; f(t) is the correction function required to calculate the three values of L, a, and b, t represents the independent variable, and in the formula Bring in.
[0024] In the step (3), several landmark pixels of the ROI region and the background region are selected through human-computer interaction, and the color values of the landmark pixels of the ROI region and the background region are obtained and averaged as their respective color sampling values; after the color sampling values are obtained, the ROI region is extracted from the image:
[0025] For each pixel p, the following Euclidean distance dpk is used to determine whether the pixel belongs to the ROI area or the background area:
[0026]
[0027] Among them, d pk represents the sampling value of pixel p in region k, where k represents the sequence number of the region, indicating whether it is a background region or a ROI region; a p 、b p They represent the components of the a and b channels of pixel p in the Lab color space, respectively. k 、b k They represent the components of the a and b channels of the color sampling values of region k in the Lab color space respectively;
[0028] If the sampling value d of pixel p in the background area is p0≤ the sampling value d of pixel p in the ROI area p1 , then pixel p belongs to the background area; if pixel p is in the background area, the sampling value d p0 >The sampling value d of pixel p in the ROI area p1 , then the pixel p belongs to the ROI area.
[0029] The auxiliary identification points are another color point pre-set on the moving object in order to match the points identified in different images. The auxiliary identification points are set in the ROI area. Each moving object is pre-set with auxiliary identification points. In a specific implementation, a label is attached to the moving object, and the identification pattern on the label is used as the auxiliary identification point. The position of the auxiliary identification point on the moving object is known.
[0030] The feature points are color marking points set to capture the three-dimensional features of the target object. Surface positions with structural features are additionally set or selected as feature points on the moving object, and the positions of the feature points on the moving object are unknown.
[0031] The feature points and auxiliary identification points are all located in the ROI area.
[0032] The step (4) is specifically:
[0033] (4.1) For the feature points and auxiliary identification points in the ROI area, the following method is used to obtain the color sampling value: first, a pixel in the feature point / auxiliary identification point in the ROI area is determined as a seed pixel through the human-computer interaction method, added to the empty stack, and then the following stack algorithm is used to obtain it:
[0034] 1) The top pixel of the stack is popped out of the stack as the popped pixel;
[0035] 2) Mark the pixels out of the stack as pixels belonging to feature points / auxiliary recognition points;
[0036] 3) Calculate the Euclidean distance between each adjacent pixel in the four neighborhoods of the stacked pixel and the stacked pixel: If the Euclidean distance of the adjacent pixel is less than the preset similarity metric value T e , and the adjacent pixel is not marked as a feature point or auxiliary recognition point, then the adjacent pixel is pushed into the stack;
[0037] 4) Continue to iterate the above process 1) to 3) until the Euclidean distance of all adjacent pixels in the current iteration is not less than the preset similarity measurement value T e , or all adjacent pixels are marked as feature points or auxiliary recognition points, the iteration stops;
[0038] The color average of all pixels belonging to the feature point is taken as the color sampling value L of the feature point.2 、a 2 、b 2 , take the average color of all pixels belonging to the auxiliary recognition point as the color sampling value L of the auxiliary recognition point 3 、a 3 、b 3 .
[0039] (4.2) For each pixel p, the Euclidean distance d is calculated using the following formula: pk To determine whether the pixel belongs to the feature point area or the auxiliary identification point area:
[0040]
[0041] Among them, d pk represents the sampling value of pixel p in region k, where k represents the sequence number of the region, indicating whether it is a background region or a ROI region; a p 、b p They represent the components of the a and b channels of pixel p in the Lab color space, respectively. k 、b k They represent the components of the a and b channels of the color sampling values of region / point k in the Lab color space.
[0042] If d p1 <d p2 And d p1 <d p3 , then the pixel belongs to the non-feature point and non-auxiliary identification point area;
[0043] If d p2 <d p1 And d p2 <d p3 , then the pixel belongs to the feature point area;
[0044] If d p3 <d p1 And d p3 <d p2 , then the pixel belongs to the auxiliary recognition point area;
[0045] (4.3) Perform circle fitting on all identified feature point areas / auxiliary identification point areas, and use the center of the circle as the center coordinate of the feature point area / auxiliary identification point area, that is, as the coordinate of the feature point / auxiliary identification point; count the number of feature points / auxiliary identification points based on the number of coordinates and compare it with the actual number of feature points / auxiliary identification points on the known moving object. If the numbers are consistent, the recognition is successful and proceed to the next step.
[0046] In the step (5), the feature points are sorted from near to far according to the distance to the marked straight line to determine whether the feature points in different images taken by the multi-eye camera at the same time are the same, and the feature points with the same sorting in different images are considered to be the same feature point.
[0047] The distance and positional relationship of the feature point relative to the marking line basically does not change regardless of the camera shooting angle, so it is adopted as the feature point matching method.
[0048] In the step (6), the same feature point in different images taken by the multi-camera at the same time is matched, and the three-dimensional world coordinates of the feature point are calculated using the camera parameters calibrated in step (1).
[0049] The present invention first calibrates the camera, reads in the captured image and converts it into Lab color space, first extracts the ROI area in the image according to the sampling values of different areas, then obtains the color sampling values of the feature points and auxiliary recognition points, performs color recognition on the ROI area again, obtains the feature point area, determines whether the recognition is successful by comparing the quantity, and finally uses the auxiliary recognition points for matching, reversely calculates the three-dimensional coordinates, and reconstructs.
[0050] When identifying feature points, the present invention first obtains the image captured by the multi-camera, samples the color values of different image regions and color points, selects the cluster center through human-computer interaction during sampling, and takes the average color value of the pixel as the sampled color after approximate color region growth. Then, the ROI (Region of Interest) area is segmented using the nearest neighbor principle, and the target color recognition and segmentation are continued in the ROI area. Then, the identified target feature points are compared in quantity and sorted in position, and finally the shape recognition and coordinate acquisition of the feature points are realized.
[0051] Due to the adoption of the above technical solution, the beneficial effects achieved by the present invention are:
[0052] The present invention uses multi-eye vision to achieve greater coverage of the field of view, and can also accurately capture moving objects.
[0053] The method of the present invention has wider applicability and adopts more direct color features, so the recognition effectiveness is improved. In a moving object, the color features will not change, and the accuracy is also guaranteed.
[0054] The present invention has the characteristics of simple operation, effective recognition, and high accuracy. It is suitable for three-dimensional reconstruction of complex scenes, static and dynamic target recognition and tracking, such as the feedback-free network system of drones, the intelligent car tracking system in complex terrain, or the animal tracking system. BRIEF DESCRIPTION OF THE DRAWINGS
[0055] Figure 1 is a flow chart of the method of the present invention;
[0056] Figure 2 A schematic diagram of a shooting model for an embodiment of the method of the present invention;
[0057] Figure 3 It is a recognition result diagram of an embodiment of the method of the present invention;
[0058] in, Figure 2 1 and 2 are feature points of different colors, 3 is an auxiliary recognition point, and 4 is the camera. Figure 3 1 and 2 are different identified feature points, and 3 is the identified auxiliary identification point. DETAILED DESCRIPTION
[0059] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:
[0060] like Figure 1 As shown, the embodiment and implementation process of the present invention are as follows:
[0061] Step (1): multi-eye vision camera calibration, using Zhang Zhengyou calibration method to calibrate the multi-eye vision camera;
[0062] Step (2): read the image captured by the multi-camera for the moving object. In this embodiment, the aircraft is taken as the target object and the shooting model is simplified as follows: Figure 2 As shown in the figure, the image is subjected to distortion correction, pre-processed Gaussian noise reduction, and the image color space is converted into Lab space, and the image in the pre-processed Lab color space is taken as the solution target.
[0063] Converting from RGB color space to Lab color space borrows XYZ color space. The specific calculation formula is as follows:
[0064]
[0065] R * =R / 255, G * =G / 255,B * =B / 255
[0066]
[0067]
[0068]
[0069]
[0070] Among them, R, G, and B represent the RGB values of the pixels in the original image, respectively.* , G * , B * Respectively represent the RGB values scaled to the value between 0 and 1, X, Y, and Z represent the XYZ values of the three imaginary primary colors in the XYZ color space, which can be calculated by the above formula. L, a, and b represent the three values in the Lab color space, where L represents brightness, a represents the component from green to red, and b represents the component from blue to yellow. H represents the coefficient matrix, X n , Y n , Z n is the sum of the coefficients in the first row, the second row, and the third row of the coefficient matrix H in the matrix operation. The purpose is to ensure that the XYZ value is mapped within the same range as the RGB value; f(t) is the correction function required to calculate the three values of L, a, and b. t represents the independent variable. Bring in.
[0071] In a specific implementation, the coefficient matrix is
[0072] If the calculated value of L is negative, it is taken as 0.
[0073] Step (3): obtaining color sampling data of the background and ROI area, dividing the image into the background area and the ROI area according to the color sampling data, and segmenting the ROI area from the image;
[0074] Through human-computer interaction, several landmark pixels of the ROI area and the background area are selected, and the color values of the landmark pixels of the ROI area and the background area are obtained and the average value is taken as the respective color sampling values; in the embodiment, 10-20 points are randomly selected from different areas, and their color values are obtained and the average value is taken as the sampling value.
[0075] After obtaining the color sampling value, extract the ROI area from the image:
[0076] Set the color sampling value of the background area to L 0 、a 0 、b 0 , the color sampling value of the ROI area is L 1 、a 1 、b 1 , the pixel set of ROI area is U i (0<i<m), the pixels of the whole image constitute the whole set U m , then U i ∈U m At the same time, each pixel in the image stores the three-channel Lab value. Let the three-channel Lab color value of pixel p be L p 、a p 、bp .
[0077] For each pixel p, the Euclidean distance d is calculated using the following formula: pk To determine whether the pixel belongs to the ROI area or the background area:
[0078]
[0079] Among them, d pk represents the sampling value of pixel p in region k, where k represents the sequence number of the region, indicating whether it is a background region or a ROI region; a p 、b p They represent the components of the a and b channels of pixel p in the Lab color space, respectively. k 、b k They represent the components of the color sampling values of area k in the a and b channels in the Lab color space, specifically a 0 、b 0 Respectively represent the color sampling values of the background area in the Lab color space a, b two channel components, a 1 、b 1 Represents the color sampling value of the ROI area in the Lab color space, in the two channels a and b;
[0080] If the sampling value d of pixel p in the background area is p0 ≤ the sampling value d of pixel p in the ROI area p1 , then pixel p belongs to the background area; if pixel p is in the background area, the sampling value d p0 >The sampling value d of pixel p in the ROI area p1 , then the pixel p belongs to the ROI area.
[0081] Step (4): According to human-computer interaction, different feature points and auxiliary recognition points in the recognition ROI area are selected to obtain color sampling values of the feature points and the auxiliary recognition points, and the feature point area and the auxiliary recognition point area are identified according to the color sampling values of the feature points and the auxiliary recognition points, and then the number of feature points and auxiliary recognition points is obtained, and the number is compared to obtain all feature points and auxiliary recognition points;
[0082] (4.1) For the feature points and auxiliary identification points in the ROI area, the following method is used to obtain the color sampling value: first, a pixel in the feature point / auxiliary identification point in the ROI area is determined as a seed pixel through the human-computer interaction method, added to the empty stack, and then the following stack algorithm is used to obtain it:
[0083] The specific method is to repeat the following steps:
[0084] 1) The top pixel of the stack is popped out of the stack as the popped pixel;
[0085] 2) Mark the pixels out of the stack as pixels belonging to feature points / auxiliary recognition points;
[0086] 3) Calculate the Euclidean distance between each adjacent pixel in the four neighborhoods of the stacked pixel and the stacked pixel: If the Euclidean distance of the adjacent pixel is less than the preset similarity metric value T e , and the adjacent pixel is not marked as a feature point or auxiliary recognition point, then the adjacent pixel is pushed into the stack;
[0087] 4) Continue to iterate the above process 1) to 3) until the Euclidean distance of all adjacent pixels in the current iteration is not less than the preset similarity measurement value T e , or all adjacent pixels are marked as feature points or auxiliary recognition points, the iteration stops;
[0088] The calculation method of the Euclidean distance in step (3) is the same as that in step (3).
[0089] The color average of all pixels belonging to the feature point is taken as the color sampling value L of the feature point. 2 、a 2 , b 2 , take the average color of all pixels belonging to the auxiliary recognition point as the color sampling value L of the auxiliary recognition point 3 、a 3 , b 3 .
[0090] (4.1) The calculation process is a preprocessing process, which only needs to be processed once during the experiment. The purpose is to obtain the color sampling values of feature points and auxiliary recognition points.
[0091] (4.2) The color sampling value of the ROI area is L 1 、a 1 , b 1 , the color sampling value of the feature point is L 2 、a 2 , b 2 , the color sampling value of the auxiliary recognition point is L 3 、a 3 , b 3 For each pixel p, the Euclidean distance d is calculated using the following formula: pk To determine whether the pixel belongs to the feature point area or the auxiliary identification point area:
[0092]
[0093] Among them, d pk represents the sampling value of pixel p in region k, where k represents the sequence number of the region, indicating whether it is a background region or a ROI region; a p、b p They represent the components of the a and b channels of pixel p in the Lab color space, respectively. k 、b k They represent the components of the a and b channels of the color sampling values of region / point k in the Lab color space, specifically a 1 、b 1 Represents the color sampling value of the ROI area in the Lab color space, the components of the two channels a and b; a 2 、b 2 Represents the color sampling value of the feature point in the Lab color space, the components of the two channels a and b; a 3 、b 3 Represents the components of the a and b channels of the color sampling value of the auxiliary recognition point in the Lab color space;
[0094] If d p1 <d p2 And d p1 <d p3 , then the pixel belongs to the non-feature point and non-auxiliary identification point area;
[0095] If d p2 <d p1 And d p2 <d p3 , then the pixel belongs to the feature point area;
[0096] If d p3 <d p1 And d p3 <d p2 , then the pixel belongs to the auxiliary recognition point area;
[0097] (4.3) After continuous pixels are processed into connected domains, circle fitting is performed on all identified feature point areas / auxiliary identification point areas, and the center of the circle is used as the central coordinate of the feature point area / auxiliary identification point area, that is, as the coordinate of the feature point / auxiliary identification point; the number of feature points / auxiliary identification points is counted according to the number of coordinates and compared with the actual number of feature points / auxiliary identification points on the known moving object. If the number is consistent, the recognition is successful and the next step is performed; otherwise, the recognition fails, the image is invalidated, and the process returns to step (1) to continue.
[0098] In this embodiment, two sets of feature points with different colors are used, so the obtained sampling values are L 2 、a 2 、b 2 , L 3 、a 3 、b 3 ; Take the average color of all the pixel values belonging to the auxiliary recognition point, which is the color sampling value L of the auxiliary recognition point. 4 、a4 、b 4 .
[0099] This calculation process is a preprocessing process and only needs to be processed once during the experiment. The purpose is to obtain the color sampling values of feature points and auxiliary recognition points.
[0100] Then the ROI area is calculated and processed again in a similar way to obtain the feature point area. The sampling mean of the ROI area is L 1 、a 1 、b 1 , the color sampling value of the feature point is L 2 、a 2 、b 2 , L 3 、a 3 、b 3 , the color sampling value of the auxiliary recognition point is L 4 、a 4 、b 4 , the Euclidean distance dpk is also used to determine the subordinate area of a certain pixel. The calculation formula is:
[0101]
[0102] Among them, pixel p needs to traverse the image. p1 <d p2 , d p1 <d p3 And d p1 <d p4 Then the pixel belongs to the non-feature point area; if d p2 <d p1 , d p2 <d p3 And d p2 <d p4 Then the pixel belongs to feature point 1 area; if d p3 <d p1 , d p3 <d p2 And d p3 <d p4 , then the pixel belongs to feature point 2 area; if d p4 <d p1 , d p4 <d p And d p4 <d p3 , then the pixel belongs to the auxiliary recognition point area. A circle is fitted for the identified feature points and the auxiliary recognition point area, and the center of the circle can be regarded as the coordinates of the feature point.
[0103] Next, the number of identified feature points needs to be compared. In this embodiment, 4 feature points are set for each color. If the number of identified feature points is exactly 4, it can be considered that the recognition is successful, and the coordinates of all identified feature points are calculated; if the number of feature points is not 4, it is considered that the recognition fails, the image is invalidated, and the process returns to step (1) to continue;
[0104] Step (5): construct a marked line using the auxiliary recognition points, and sort the feature points from near to far according to the distance from the point to the line;
[0105] The feature points are sorted from near to far according to the distance to the marked straight line to determine whether the feature points in different images taken by the multi-eye camera at the same time are the same. The feature points with the same sorting in different images are considered to be the same feature point.
[0106] Step (6): The actual coordinates of the feature points are calculated from the different image coordinates of the same feature point in different images taken by the multi-camera at the same time to perform three-dimensional reconstruction.
[0107] Match the same feature point in different images taken by multiple cameras at the same time, and use the camera parameters calibrated in step (1) to calculate the three-dimensional world coordinates of the feature point.
Claims
1. A method for identifying and reconstructing feature points of moving object images based on multi-view vision, characterized in that: The method comprises the following steps: Step (1): Use Zhang Zhengyou calibration method to calibrate the multi-view camera; Step (2): read the image captured by the multi-camera of the moving object, perform distortion correction on the image, pre-process the Gaussian noise reduction, and convert the image color space into Lab space; Step (3): obtaining color sampling data of the background and ROI area, dividing the image into the background area and the ROI area according to the color sampling data, and segmenting the ROI area from the image; Step (4): According to human-computer interaction, different feature points and auxiliary recognition points in the recognition ROI area are selected to obtain color sampling values of the feature points and the auxiliary recognition points, and the feature point area and the auxiliary recognition point area are identified according to the color sampling values of the feature points and the auxiliary recognition points, and then the number of feature points and auxiliary recognition points is obtained, and the number is compared to obtain all feature points and auxiliary recognition points; Step (5): construct a marked line using the auxiliary recognition points, and sort the feature points from near to far according to the distance from the point to the line; Step (6): The actual coordinates of the feature points are calculated from the different image coordinates of the same feature point in different images taken by the multi-camera at the same time to perform three-dimensional reconstruction.
2. The method for identifying and reconstructing feature points of moving object images based on multi-viewing vision as claimed in claim 1, characterized in that: In the step (2), the conversion from RGB color space to Lab color space uses XYZ color space, and the specific calculation formula is as follows: R * =R / 255,G * =G / 255,B * =B / 255 Among them, R, G, and B represent the RGB values of the pixels in the original image, respectively. * , G * , B * Respectively represent the scaling of RGB values to values between 0 and 1, X, Y, and Z represent the XYZ values of the imaginary three primary colors in the XYZ color space; L, a, and b represent the three values in the Lab color space, where L represents brightness, a represents the component from green to red, and b represents the component from blue to yellow; H represents the coefficient matrix, X n , Y n , Z n is the sum of the coefficients in the first row, the second row, and the third row of the coefficient matrix H in the matrix operation; f(t) is the correction function required to calculate the three values of L, a, and b, t represents the independent variable, and in the formula Bring in.
3. The method for identifying and reconstructing feature points of moving object images based on multi-viewing vision as claimed in claim 1, characterized in that: In the step (3), several landmark pixels of the ROI region and the background region are selected through human-computer interaction, and the color values of the landmark pixels of the ROI region and the background region are obtained and averaged as their respective color sampling values; after the color sampling values are obtained, the ROI region is extracted from the image: For each pixel p, the Euclidean distance d is calculated using the following formula: pk To determine whether the pixel belongs to the ROI area or the background area: Among them, d pk represents the sampling value of pixel p in region k, where k represents the sequence number of the region, indicating whether it is a background region or a ROI region; a p 、b p They represent the components of the a and b channels of pixel p in the Lab color space, respectively. k 、b k They represent the components of the a and b channels of the color sampling values of region k in the Lab color space respectively; If the sampling value d of pixel p in the background area is p0 ≤ the sampling value d of pixel p in the ROI area p1 , then pixel p belongs to the background area; if pixel p is in the background area, the sampling value d p0 >The sampling value d of pixel p in the ROI area p1 , then the pixel p belongs to the ROI area.
4. The method for identifying and reconstructing feature points of moving object images based on multi-viewing vision as claimed in claim 1, characterized in that: The step (4) is specifically as follows: (4.1) For the feature points and auxiliary identification points in the ROI area, the following method is used to obtain the color sampling value: first, a pixel in the feature point / auxiliary identification point in the ROI area is determined as a seed pixel through the human-computer interaction method, added to the empty stack, and then the following stack algorithm is used to obtain it: 1) The top pixel of the stack is popped out of the stack as the popped pixel; 2) Mark the pixels out of the stack as pixels belonging to feature points / auxiliary recognition points; 3) Calculate the Euclidean distance between each adjacent pixel in the four neighborhoods of the stacked pixel and the stacked pixel: If the Euclidean distance of the adjacent pixel is less than the preset similarity metric value T e , and the adjacent pixel is not marked as a feature point or auxiliary recognition point, then the adjacent pixel is pushed into the stack; 4) Continue to iterate the above process 1) to 3) until the Euclidean distance of all adjacent pixels in the current iteration is not less than the preset similarity measurement value T e , or all adjacent pixels are marked as feature points or auxiliary recognition points, the iteration stops; The colors of all pixels belonging to the feature points are averaged as the color sampling values L2, a2, b2 of the feature points, and the colors of all pixels belonging to the auxiliary recognition points are averaged as the color sampling values L3, a3, b3 of the auxiliary recognition points; (4.2) For each pixel p, the Euclidean distance d is calculated using the following formula: pk To determine whether the pixel belongs to the feature point area or the auxiliary identification point area: Among them, d pk represents the sampling value of pixel p in region k, where k represents the sequence number of the region, indicating whether it is a background region or a ROI region; a p 、b p They represent the components of the a and b channels of pixel p in the Lab color space, respectively. k 、b k They represent the components of the a and b channels of the color sampling values of region / point k in the Lab color space. If d p1 <d p2 And d p1 <d p3 , then the pixel belongs to the non-feature point and non-auxiliary identification point area; If d p2 <d p1 And d p2 <d p3 , then the pixel belongs to the feature point area; If d p3 <d p1 And d p3 <d p2 , then the pixel belongs to the auxiliary recognition point area; (4.3) Perform circle fitting on all identified feature point areas / auxiliary identification point areas, and use the center of the circle as the center coordinate of the feature point area / auxiliary identification point area, that is, as the coordinate of the feature point / auxiliary identification point; count the number of feature points / auxiliary identification points based on the number of coordinates and compare it with the actual number of feature points / auxiliary identification points on the known moving object. If the numbers are consistent, the recognition is successful and proceed to the next step.
5. The method for identifying and reconstructing feature points of moving object images based on multi-viewing vision as claimed in claim 1, characterized in that: In the step (5), the feature points are sorted from near to far according to the distance to the marked straight line to determine whether the feature points in different images taken by the multi-eye camera at the same time are the same, and the feature points with the same sorting in different images are considered to be the same feature point.
6. The method for identifying and reconstructing feature points of moving object images based on multi-viewing vision as claimed in claim 1, characterized in that: In the step (6), the same feature point in different images taken by the multi-camera at the same time is matched, and the three-dimensional world coordinates of the feature point are calculated using the camera parameters calibrated in step (1).
