Indoor visual identification positioning device and method

Abstract

The invention discloses an indoor visual identification positioning device and method, and relates to the technical field of computer visual positioning. The device comprises two fixing rods of a large arm and a small arm and a camera holder. The positioning method comprises the following steps: 1) fixing the upper end of the fixing device on a ceiling; 2) fixing the camera to the lower end of the fixing device; 3) correcting internal parameters and external parameters of the industrial camera; 4) identifying and positioning the object by adopting a two-stage target detection positioning process; and 5) performing information fusion and positioning precision optimization by using Kalman filtering. According to the device and the method, the technical problems of low precision and poor applicability in the prior art are solved, real-time positioning of a small target object in an indoor scene can be realized, and high-precision positioning information is provided for a control and command simulation model.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision positioning, and in particular to an indoor vision positioning system for commanding and controlling a simulation model. Background Art

[0002] Indoor positioning systems are technologies that locate and track targets indoors. They are widely used in command and control simulation models to accurately locate and track personnel, equipment, and supplies in real time. In recent years, with advances in deep learning and hardware performance, the field of computer vision has rapidly developed, and computer vision technology has also been applied to positioning. Compared to traditional indoor positioning technologies such as RSSI and TOA / TDOA, indoor visual positioning offers the advantages of high precision, low cost, and strong adaptability to complex environments.

[0003] Although indoor visual positioning technology has made significant progress, it still faces some challenges and problems. First, many positioning targets in indoor environments are usually small in size. Especially when the camera is far away from the target or the viewing angle is large, these targets may only occupy a small area of ​​the image. To improve the accuracy of detecting small targets, higher resolution input images and deeper and more complex neural network models are required. This significantly increases the computational overhead, making it difficult to simultaneously meet the real-time and high-precision requirements of the positioning system. In addition, multi-camera systems are often used to expand the field of view or provide redundant information to improve the accuracy and robustness of positioning. However, the perspective information of different cameras needs to be integrated through an efficient data fusion algorithm to generate accurate positioning results. Summary of the Invention

[0004] In order to realize indoor target detection and positioning, the present invention provides an indoor visual recognition and positioning device and method.

[0005] The present invention is achieved through the following technical solutions: an indoor visual recognition and positioning device, which is provided with a large arm, a small arm, a pan-tilt head, and an industrial camera in sequence from top to bottom; the top of the large arm is installed on the ceiling through expansion screws, and the bottom is connected to the small arm through a buckle; the bottom of the small arm is installed with a pan-tilt head, which includes a horizontal rotation mechanism and a pitch adjustment mechanism; the bottom of the pan-tilt head is installed with an industrial camera.

[0006] The method for indoor visual recognition and positioning by the device for indoor visual recognition and positioning comprises the following steps:

[0007] Step 1) Fix the upper end of the fixing device to the ceiling;

[0008] Step 2) Fix the camera to the lower end of the fixing device;

[0009] Step 3) Calibrate the internal and external parameters of the industrial camera;

[0010] The internal parameter calibration method for industrial cameras is:

[0011] 3a.1) Under stable lighting conditions, place a high-precision checkerboard calibration plate within the camera's field of view.

[0012] 3a.2) Take at least 10 images of the calibration plate from different directions and distances;

[0013] 3a.3) Calculate the camera's intrinsic parameter matrix and distortion coefficients using a computer program based on the Zhang Zhengyou calibration method. The intrinsic parameter matrix includes the focal length and principal point coordinates, and the distortion coefficients include radial and tangential distortion parameters.

[0014] The external parameter calibration method for industrial cameras is:

[0015] 3b.1) Place a checkerboard calibration plate of known dimensions on the positioning plane, aligning the plate with the center of the camera's field of view.

[0016] 3b.2) Use computer vision algorithms to detect corner points on the calibration plate and obtain pixel coordinates of at least four feature points;

[0017] 3b.3) Calculate the world coordinates of the corresponding feature points based on the actual size of the calibration plate;

[0018] 3b.4) Combine the camera intrinsic parameters and distortion coefficients to solve the camera pose using the PnP algorithm, and output the rotation vector and translation vector.

[0019] Step 4) Use a two-stage target detection and positioning process to identify and locate the object;

[0020] Two-stage target detection and positioning process: The real-time video stream is transmitted to the computer program via a data line or RTSP protocol, and the video image is corrected in real time using an image correction algorithm based on the distortion coefficient.

[0021] For each frame of the real-time video stream, two-stage target detection is performed:

[0022] In the first stage, the target detection algorithm is used to preliminarily identify the positioning target in the picture and generate the corresponding detection frame;

[0023] In the second stage, the image area within the detection frame is extracted as input, the distinguishing identification information on the target object is detected through the recognition algorithm, and the specific target information and the center pixel coordinates are output.

[0024] The model used in the two-stage target detection and positioning process is a neural network model based on deep learning, including but not limited to the YOLO series model.

[0025] In the two-stage target detection and positioning process, based on the camera intrinsic parameter matrix, the camera extrinsic parameter pose information and the obtained target pixel coordinates, the three-dimensional spatial coordinates of the target in the real world are obtained by calculating the transformation matrix from the camera coordinate system to the world coordinate system.

[0026] Step 5) Use Kalman filtering to perform information fusion and positioning accuracy optimization.

[0027] The number of industrial cameras is greater than or equal to one; when multiple industrial cameras are used, each camera is installed according to a preset spatial layout plan to ensure full coverage of the monitoring area; the video stream acquisition of multiple industrial cameras is based on a video stream reading program, and a concurrent processing mechanism is adopted to realize parallel reading and processing of the video stream.

[0028] In the method of indoor visual recognition and positioning, when there are multiple industrial cameras, each camera independently generates positioning information; an information fusion algorithm is used to process the positioning information, and the algorithm is:

[0029] 1) When the target object appears in the overlapping area of ​​the field of view of at least two cameras at the same time, obtain the independent positioning coordinate set {P1(x1,y1),P2(x2,y2),...,P n (x n ,y n )}; where P(x i ,y i ) is the independent positioning coordinate of the target object by the i-th camera, x i and y i are the horizontal and vertical coordinates of the target in the camera coordinate system respectively;

[0030] 2) Based on the relative spatial relationship between each camera and the target, calculate the confidence weight of each positioning result {w1,w2,...,w n}, where the weight calculation function includes the Euclidean distance between the camera and the target object, the cosine value of the angle between the camera optical axis and the target line, and the historical positioning error statistics of the camera; i is the confidence weight of the positioning result of the i-th camera. The larger the weight, the more reliable the positioning result of the camera.

[0031] 3) Use weighted least squares method to fuse multi-source positioning data and establish target position optimization model: min∑w i ||PP i || 2 , solve and get the optimal estimated coordinate P after fusion * (x * ,y * ); where P(x,y) is the position coordinate of the target object to be optimized, P i (x i ,yi ) is the independent positioning coordinate of the target object by the i-th camera, w i is the confidence weight of the positioning result of the i-th camera, P*(x*,y*) is the optimal estimated coordinate after weighted least squares fusion, which is the joint result of multi-camera positioning;

[0032] 4) Input the fused coordinates into the Kalman filter, correct the predicted value through the measurement update step, output the final positioning result, and synchronously update the state vector and covariance matrix of the Kalman filter corresponding to each camera.

[0033] The beneficial effects of the present invention are:

[0034] 1. The camera fixing device implemented by the upper arm, the lower arm and the pan / tilt head of the present invention can achieve full coverage of the positioning area by the camera field of view and is easy to debug.

[0035] 2. Through the two-stage target detection and positioning process, a general target detection model can be used to simultaneously meet the real-time and accuracy requirements of small target detection. The method is simple and has low requirements on computing equipment.

[0036] 3. By converting pixel coordinates to world coordinates and combining it with Kalman filtering, high-precision positioning information of the target can be obtained, and the positioning information can be transmitted to other functional modules (such as the target control module) through network communication to provide positioning information reference. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0038] The present invention is achieved through the following technical solutions: a device for indoor visual recognition and positioning, which is provided with a large arm 1, a small arm 2, a pan-tilt head 3, and an industrial camera 4 from top to bottom; the top of the large arm 1 is installed on the ceiling through expansion screws, and the bottom is connected to the small arm 2 through a buckle; the pan-tilt head 3 is installed at the bottom of the small arm 2, and the pan-tilt head 3 includes a horizontal rotation mechanism and a pitch adjustment mechanism; the industrial camera 4 is installed at the bottom of the pan-tilt head 3.

[0039] The method for indoor visual recognition and positioning by the device for indoor visual recognition and positioning comprises the following steps:

[0040] Step 1) Fix the upper end of the fixing device to the ceiling;

[0041] Step 2) Fix the camera to the lower end of the fixing device;

[0042] Step 3) Calibrate the internal and external parameters of the industrial camera 4;

[0043] The internal parameter calibration method for industrial cameras is:

[0044] 3a.1) Under stable lighting conditions, place a high-precision checkerboard calibration plate within the camera's field of view.

[0045] 3a.2) Take at least 10 images of the calibration plate from different directions and distances;

[0046] 3a.3) Calculate the camera's intrinsic parameter matrix and distortion coefficients using a computer program based on the Zhang Zhengyou calibration method. The intrinsic parameter matrix includes the focal length and principal point coordinates, and the distortion coefficients include radial and tangential distortion parameters.

[0047] The external parameter calibration method for industrial cameras is:

[0048] 3b.1) Place a checkerboard calibration plate of known dimensions on the positioning plane, aligning the plate with the center of the camera's field of view.

[0049] 3b.2) Use computer vision algorithms to detect corner points on the calibration plate and obtain pixel coordinates of at least four feature points;

[0050] 3b.3) Calculate the world coordinates of the corresponding feature points based on the actual size of the calibration plate;

[0051] 3b.4) Combine the camera intrinsic parameters and distortion coefficients to solve the camera pose using the PnP algorithm, and output the rotation vector and translation vector.

[0052] Step 4) Use a two-stage target detection and positioning process to identify and locate the object;

[0053] Two-stage target detection and positioning process: The real-time video stream is transmitted to the computer program via a data line or RTSP protocol, and the video image is corrected in real time using an image correction algorithm based on the distortion coefficient.

[0054] For each frame of the real-time video stream, two-stage target detection is performed:

[0055] In the first stage, the target detection algorithm is used to preliminarily identify the positioning target in the picture and generate the corresponding detection frame;

[0056] In the second stage, the image area within the detection frame is extracted as input, the distinguishing identification information on the target object is detected through the recognition algorithm, and the specific target information and the center pixel coordinates are output.

[0057] The model used in the two-stage target detection and positioning process is a neural network model based on deep learning, including but not limited to the YOLO series model.

[0058] In the two-stage target detection and positioning process, based on the camera intrinsic parameter matrix, the camera extrinsic parameter pose information and the obtained target pixel coordinates, the three-dimensional spatial coordinates of the target in the real world are obtained by calculating the transformation matrix from the camera coordinate system to the world coordinate system.

[0059] Step 5) Use Kalman filtering to perform information fusion and positioning accuracy optimization.

[0060] The number of the industrial cameras 4 is greater than or equal to one; when multiple industrial cameras 4 are used, each camera is installed according to a preset spatial layout plan to ensure full coverage of the monitoring area; the video stream acquisition of multiple industrial cameras 4 is based on a video stream reading program, and a concurrent processing mechanism is adopted to realize parallel reading and processing of the video stream.

[0061] In the method for indoor visual recognition and positioning, when there are multiple industrial cameras 4, each camera independently generates positioning information; an information fusion algorithm is used to process the positioning information, and the algorithm is:

[0062] 1) When the target object appears in the overlapping area of ​​the field of view of at least two cameras 4 at the same time, the independent positioning coordinate sets {P1(x1,y1),P2(x2,y2),...,P n (x n ,y n )}; where P(x i ,y i ) is the independent positioning coordinate of the target object by the i-th camera, x i and y i are the horizontal and vertical coordinates of the target in the camera coordinate system respectively;

[0063] 2) Based on the relative spatial relationship between each camera and the target, calculate the confidence weight of each positioning result {w1,w2,...,w n}, where the weight calculation function includes the Euclidean distance between the camera and the target object, the cosine value of the angle between the camera optical axis and the target line, and the historical positioning error statistics of the camera; i is the confidence weight of the positioning result of the i-th camera. The larger the weight, the more reliable the positioning result of the camera.

[0064] 3) Use weighted least squares method to fuse multi-source positioning data and establish target position optimization model: min∑w i ||PP i || 2 , solve and get the optimal estimated coordinate P after fusion * (x * ,y * ); where P(x,y) is the position coordinate of the target object to be optimized, P i (x i,y i ) is the independent positioning coordinate of the target object by the i-th camera, w i is the confidence weight of the positioning result of the i-th camera, P*(x*,y*) is the optimal estimated coordinate after weighted least squares fusion, which is the joint result of multi-camera positioning;

[0065] 4) Input the fused coordinates into the Kalman filter, correct the predicted value through the measurement update step, output the final positioning result, and synchronously update the state vector and covariance matrix of the Kalman filter corresponding to each camera.

[0066] When used specifically:

[0067] 1. System hardware and mechanical components installation

[0068] AFixture Installation

[0069] Step 1: Fix the arm 1 vertically to the predetermined position of the ceiling with expansion screws to ensure its stability as the main supporting structure.

[0070] Step 2: Connect the small arm 2 to the end of the large arm 1 through the snap-on connection mechanism, adjust the length of the small arm to the target coverage area, and lock the snap to secure it.

[0071] Step 3: Install the multi-degree-of-freedom gimbal 3 at the end of the forearm 2, adjust the azimuth angle of the gimbal through the horizontal rotation mechanism, and adjust the pitch angle through the pitch adjustment mechanism to ensure that the field of view of the industrial camera 4 covers the target area.

[0072] Step 4: Fix the industrial camera 4 to the bottom of the pan / tilt head 3, ensure that it is firmly installed, and connect the power and data cables to the computer terminal.

[0073] B Multi-camera system configuration

[0074] Choose a single camera or multi-camera layout based on the area of ​​the positioning area and the occlusion situation. If using multiple cameras, deploy them according to the following rules:

[0075] The cameras 4 are arranged in a ring or matrix, with a field of view overlap ratio of ≥20%, ensuring that the target object can be captured in the overlapping area of ​​at least two cameras.

[0076] Synchronous acquisition of multi-camera video streams is achieved through hardware trigger signals or software timestamps, with an error of ≤1ms.

[0077] 2. Industrial Camera Calibration Method

[0078] A Internal reference calibration

[0079] Step 1: In a stable lighting environment, place a high-precision checkerboard calibration plate within the camera's field of view.

[0080] Step 2: Take at least 15 images of the calibration plate at different tilt angles and azimuths, ranging from 0.5m to 3m along the camera's optical axis. Ensure that the calibration plate occupies at least 1 / 4 of the entire image, that there is no obstruction on the calibration plate surface, and that all corners are clear.

[0081] Step 3: Use the Zhang Zhengyou calibration method in the OpenCV library to calculate the camera intrinsic parameter matrix K and distortion coefficients (radial distortion k1, k2, k3 and tangential distortion p1, p2.

[0082]

[0083] B external reference correction

[0084] Step 1: Place a checkerboard calibration plate of known size on the positioning plane, ensuring it is located in the center of the camera's field of view.

[0085] Step 2: Extract the pixel coordinates of all corner points of the calibration plate through the corner detection algorithm.

[0086] Step 3: Calculate the world coordinates (X, Y, Z = 0) of the corresponding corner point based on the actual size of the calibration plate (e.g., the grid side length is 30 mm).

[0087] Step 4: Use the solvePnP algorithm in the OpenCV library, combined with the intrinsic parameters and distortion coefficients, to solve the extrinsic parameters of the camera relative to the calibration plate (rotation vector R and translation vector T), and convert them into the pose matrix [R|Τ] in the world coordinate system.

[0088] 3. Two-stage target detection and localization process

[0089] A. Video stream preprocessing: The video stream is transmitted to the computer in real time via the RTSP protocol or USB3.0 data cable. The image distortion correction algorithm in the OpenCV library (based on intrinsic parameters and distortion coefficients) is used to eliminate lens distortion frame by frame.

[0090] Stage 1: Coarse-grained object detection

[0091] Use the YOLOv10 model to infer the rectified video frame and output the detection box (x1, y1, x2, y2) and category label of the target object, where (x1, y1) is the coordinate of the upper left corner of the detection box and (x2, y2) is the coordinate of the lower right corner of the detection box.

[0092] B. Second stage: fine-grained object detection

[0093] Step 1: Capture the image area within the detection frame, input it into the YOLO model, identify specific marks on the target surface (such as numbers), and output the detection frame (x'1, y'1, x'2, y'2) and category label of the specific mark, where (x'1, y'1) is the coordinate of the upper left corner of the detection frame, and (x'2, y'2) is the coordinate of the lower right corner of the detection frame.

[0094] Step 2: Calculate the center point c x =(x'1+x'2) / 2,c y =(y'1+y'2) / 2, and converted to global image coordinates (i.e. pixel coordinates) (u c ,v c ):

[0095] u c =x1+c x

[0096] v c =y1+c y

[0097] C Three-dimensional coordinate transformation

[0098] Based on the camera pose matrix [R|T], the pixel coordinates (u c ,v c ) is converted to world coordinate X w ,Y w .

[0099] 4. Kalman filter positioning optimization

[0100] A Kalman filter state model definition

[0101] The state vector of the Kalman filter is defined as the position and velocity of the positioning target, that is, Among them, x k ,y k is the position of the target in the world coordinate system, is the velocity component of the target.

[0102] The state transition model is

[0103] x k+1 =F k x k +w k

[0104] Among them, x k+1 Represents the state vector of the target at the next moment, F k is the state transfer matrix, w k is the process noise, assuming it is zero-mean Gaussian noise, with covariance Q k .

[0105] For the uniform motion model, the state transfer matrix is

[0106]

[0107] Where Vt is the time step.

[0108] The observation model is

[0109] z k =H k x k +v k

[0110] Among them, z k is the observation vector, H k is the observation matrix, v k is the observation noise. Assume it is zero-mean Gaussian noise with covariance R k Assuming the position is observed directly, the observation matrix is

[0111]

[0112] B. Kalman filter prediction and update steps

[0113] Prediction stage:

[0114] (1) State prediction:

[0115]

[0116] in, is the predicted state, is the posterior estimate of the state vector at time k-1.

[0117] (2) Covariance prediction:

[0118]

[0119] Among them, P k|k-1 is the prediction covariance, P k-1|k-1 is the posterior estimated covariance matrix at time k-1.

[0120] Update phase:

[0121] (1) Calculate the Kalman gain:

[0122]

[0123] (2) Status update:

[0124]

[0125] (3) Covariance update:

[0126] P k|k =(IK k H k )P k|k-1

[0127] Among them, K k is the Kalman gain, is the updated state estimate, P k|k is the updated covariance estimate.

[0128] C output optimization results

[0129] The final output positioning coordinates are

[0130]

[0131] Its covariance matrix is

[0132]

[0133] 5. Multi-camera information fusion

[0134] A Confidence Weight Calculation

[0135] The positioning results of each camera {P1(x1,y1),P2(x2,y2),...,P n (x n ,y n )}, calculate the weight:

[0136]

[0137] Where: d i is the Euclidean distance between the camera and the target; θ i is the angle between the camera optical axis and the target line; σ i is the standard deviation of the historical positioning error of the camera; α, β, γ are normalization coefficients (α+β+γ=11).

[0138] B-weighted least squares fusion

[0139] Construct the objective function:

[0140]

[0141] Solve for the optimal estimated coordinates

[0142]

[0143] C Kalman filter synchronous update

[0144] P opt Input the global Kalman filter, update the state vector and covariance matrix, and output the final fusion positioning result.

[0145] Parts not described in detail in the present invention belong to the well-known technology in the field.

[0146] The above are only specific steps of the present invention and do not constitute any limitation to the scope of protection of the present invention; it can be extended to the field of indoor target detection and positioning, and any technical solutions formed by equivalent transformation or equivalent replacement fall within the scope of protection of the present invention.

Claims

1. A device for indoor visual recognition and positioning, characterized by: The invention is provided with a big arm (1), a small arm (2), a pan-tilt platform (3), and an industrial camera (4) in sequence from top to bottom; the top of the big arm (1) is mounted on the ceiling by means of expansion screws, and the bottom is connected to the small arm (2) by means of buckles; the bottom of the small arm (2) is mounted with a pan-tilt platform (3), which includes a horizontal rotation mechanism and a pitch adjustment mechanism; the bottom of the pan-tilt platform (3) is mounted with an industrial camera (4).

2. A method for performing indoor visual recognition and positioning using the device for indoor visual recognition and positioning according to claim 1, characterized in that: The steps are: Step 1) Fix the upper end of the fixing device to the ceiling; Step 2) Fix the camera to the lower end of the fixing device; Step 3) calibrating the internal and external parameters of the industrial camera (4); Step 4) Use a two-stage target detection and positioning process to identify and locate the object; Step 5) Use Kalman filtering to perform information fusion and positioning accuracy optimization.

3. The method for indoor visual recognition and positioning according to claim 2, characterized in that: In step 3), the industrial camera internal parameter calibration method is: 3a.1) Under stable lighting conditions, place a high-precision checkerboard calibration plate within the camera's field of view. 3a.2) Take at least 10 images of the calibration plate from different directions and distances; 3a.3) Calculate the camera's intrinsic parameter matrix and distortion coefficients using a computer program based on the Zhang Zhengyou calibration method. The intrinsic parameter matrix includes the focal length and principal point coordinates, and the distortion coefficients include radial and tangential distortion parameters.

4. The method for indoor visual recognition and positioning according to claim 2, characterized in that: In step 3), the industrial camera external parameter calibration method is: 3b.1) Place a checkerboard calibration plate of known dimensions on the positioning plane, aligning the plate with the center of the camera's field of view. 3b.2) Use computer vision algorithms to detect corner points on the calibration plate and obtain pixel coordinates of at least four feature points; 3b.3) Calculate the world coordinates of the corresponding feature points based on the actual size of the calibration plate; 3b.4) Combine the camera intrinsic parameters and distortion coefficients to solve the camera pose using the PnP algorithm, and output the rotation vector and translation vector.

5. The method for indoor visual recognition and positioning according to claim 2, characterized in that: In step 4), the two-stage target detection and positioning process: the real-time video stream is transmitted to the computer program via a data line or RTSP protocol, and the video image is subjected to real-time distortion correction using an image correction algorithm based on the distortion coefficient.

6. The method for indoor visual recognition and positioning according to claim 5, characterized in that: For each frame of the real-time video stream, two-stage target detection is performed: In the first stage, the target detection algorithm is used to preliminarily identify the positioning target in the picture and generate the corresponding detection frame; In the second stage, the image area within the detection frame is extracted as input, the distinguishing identification information on the target object is detected through the recognition algorithm, and the specific target information and the center pixel coordinates are output.

7. The method for indoor visual recognition and positioning according to claim 2, characterized in that: In step 4), the model used in the two-stage target detection and positioning process is a neural network model based on deep learning, including but not limited to the YOLO series model.

8. The method for indoor visual recognition and positioning according to claim 2, characterized in that: In step 4), in the two-stage target detection and positioning process, based on the camera intrinsic parameter matrix, the camera extrinsic parameter pose information and the obtained target pixel coordinates, the three-dimensional spatial coordinates of the target in the real world are obtained by calculating the conversion matrix from the camera coordinate system to the world coordinate system.

9. The method for indoor visual recognition and positioning according to claim 2, characterized in that: The number of the industrial cameras (4) is greater than or equal to one; when multiple industrial cameras (4) are used, each camera is installed according to a preset spatial layout plan to ensure full coverage of the monitoring area; the video stream acquisition of the multiple industrial cameras (4) is based on a video stream reading program, and a concurrent processing mechanism is used to achieve parallel reading and processing of the video stream.

10. The method for indoor visual recognition and positioning according to claim 9, characterized in that: When there are multiple industrial cameras (4), each camera generates positioning information independently; an information fusion algorithm is used to process the positioning information, and the algorithm is: 1) When the target object appears in the overlapping area of ​​the field of view of at least two cameras (4) at the same time, the independent positioning coordinate sets {P1(x1,y1),P2(x2,y2),...,P n (x n ,y n )}; Among them, P(x i ,y i ) is the independent positioning coordinate of the target object by the i-th camera, x i and y i are the horizontal and vertical coordinates of the target in the camera coordinate system respectively; 2) Based on the relative spatial relationship between each camera and the target, calculate the confidence weight of each positioning result {w1,w2,...,w n }, where the weight calculation function includes the Euclidean distance between the camera and the target object, the cosine value of the angle between the camera optical axis and the target line, and the historical positioning error statistics of the camera; i is the confidence weight of the positioning result of the i-th camera. The larger the weight, the more reliable the positioning result of the camera. 3) Use weighted least squares method to fuse multi-source positioning data and establish target position optimization model: min∑w i ||PP i || 2 , solve and get the optimal estimated coordinate P after fusion * (x * ,y * ); where P(x,y) is the position coordinate of the target object to be optimized, P i (x i ,y i ) is the independent positioning coordinate of the target object by the i-th camera, w i is the confidence weight of the positioning result of the i-th camera, P*(x*,y*) is the optimal estimated coordinate after weighted least squares fusion, which is the joint result of multi-camera positioning; 4) Input the fused coordinates into the Kalman filter, correct the predicted value through the measurement update step, output the final positioning result, and synchronously update the state vector and covariance matrix of the Kalman filter corresponding to each camera.

Images