A target detection method and system based on machine vision

Abstract

The application discloses a target detection method and system based on machine vision, relates to the technical field of target detection, and comprises the following steps: deploying collection devices and respectively acquiring collection images of the collection devices; acquiring calibration parameters of the collection devices by using the collection images; fusing the collection images of different collection devices to acquire a fused image; acquiring collection angles and collection distances of different collection points; acquiring measured distances of the corresponding collection points by using laser ranging; and constructing a distance correction model; constructing an object recognition model according to different target objects and appearance images of the target objects to acquire the target objects; and acquiring collection distances of different positions of the target objects, and acquiring corrected distances by using the distance correction model, so that the target objects can be recognized and the current distance conditions of the target objects can be judged only by using camera devices, and the efficiency and accuracy of target detection are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically a target detection method and system based on machine vision. Background Technology

[0002] Most existing target detection methods rely on a single type of sensor data, such as RGB images or depth data. Therefore, they often suffer from problems such as insufficient accuracy and difficulty in data fusion. Two-dimensional cameras provide high-precision RGB images but lack depth information, while three-dimensional depth cameras provide depth information but have lower accuracy and are prone to noise.

[0003] Since image acquisition is inevitably affected by environmental conditions, especially lighting conditions, the accuracy of target detection is always low when relying solely on image recognition. Introducing other technologies for comprehensive detection generates even more data, further reducing the efficiency of target detection. Therefore, existing technologies lack solutions to these problems. To address the shortcomings of existing technologies, this invention provides a target detection method and system based on machine vision. Summary of the Invention

[0004] The purpose of this invention is to provide a target detection method and system based on machine vision.

[0005] The objective of this invention can be achieved through the following technical solution: a target detection system based on machine vision, comprising the following modules:

[0006] The image acquisition module is used to deploy different acquisition devices and acquire corresponding images, and to obtain the calibration parameters of the acquisition devices using the acquired images;

[0007] The image fusion module is used to fuse images acquired by different acquisition devices according to the calibration parameters of different acquisition devices to obtain the corresponding fused image;

[0008] The distance correction module is used to obtain the acquisition angle and acquisition distance of different acquisition points in the fused image, and to obtain the measurement distance of the corresponding acquisition point using laser ranging. Based on the acquisition angle, acquisition distance and corresponding measurement distance of different acquisition points in different fused images, a distance correction model is constructed.

[0009] The object recognition module is used to construct an object recognition model based on different target objects and their appearance images to obtain the target objects in the acquired images, combine them with the fused images to obtain the acquisition distances at different positions of the target objects, and use the distance correction model to obtain the corrected distances at the corresponding positions and provide feedback.

[0010] Furthermore, the process of deploying different acquisition devices and acquiring corresponding images, and then using these images to obtain the calibration parameters of the acquisition devices, includes:

[0011] The acquisition device includes a 2D camera and a 3D depth camera, which are connected by a hardware synchronization line to ensure that the timestamps of the RGB image and the depth image are aligned. The 2D camera acquires RGB images of the application scene in real time, and the 3D depth camera acquires depth images of the application scene in real time. The acquired images include RGB images and depth images.

[0012] For a 2D camera, several RGB images of different poses are acquired by it against a checkerboard calibration board, and OpenCV is used for calibration to obtain the intrinsic parameter matrix and distortion coefficients of the 2D camera. For a 3D depth camera, the same number of depth images are acquired by it against the same checkerboard calibration board, and OpenCV is used to obtain the intrinsic parameter matrix and distortion coefficients of the 3D depth camera.

[0013] The relative position between the two acquisition devices is obtained by stereo calibration, and several sets of images containing the fields of view of both the 2D camera and the 3D depth camera are acquired. The corner coordinates detected by the 2D camera and the 3D depth camera are input using OpenCV to obtain the rotation matrix R and translation vector T between the two acquisition devices.

[0014] Furthermore, the process of fusing images acquired by different acquisition devices according to their calibration parameters to obtain a corresponding fused image includes:

[0015] The calibration parameters include the intrinsic parameter matrices and distortion coefficients of the 2D camera and the 3D depth camera, as well as the rotation matrix and translation vector between them, for each pixel in the depth image (u d v d ) and its depth value d, and transform them to the coordinate system of the 3D depth camera;

[0016] ;

[0017] (X d Y d Z d (c) represents the coordinates of the corresponding pixel in the coordinate system of the 3D depth camera. x c y ) are the principal point coordinates of the 3D depth camera, (f x f y ) is the focal length of the 3D depth camera. The rotation matrix R and translation vector T are used to transform the coordinate values ​​(X...)... d Y d Z dTransform to the coordinate system of a 2D camera;

[0018] ;

[0019] (X r Y r Z r ) represents the coordinates of the corresponding pixel in the coordinate system of the 2D camera. Finally, the coordinates (X, Y, F, Z) of each pixel in the coordinate system of the 2D camera are... r Y r Z r Projected onto the RGB image plane;

[0020] ;

[0021] (u r v r K represents the coordinates of the corresponding pixel in the RGB image plane. r Given the intrinsic parameter matrix of a 2D camera, a depth value is calculated for each pixel using bilinear interpolation to generate a fused image with the same resolution as the RGB image, and its timestamp is kept consistent with the depth image before conversion.

[0022] Furthermore, the process of acquiring the acquisition angle and acquisition distance of different acquisition points in the fused image, and using laser ranging to obtain the measurement distance of the corresponding acquisition points includes:

[0023] In a single fused image, any pixel is taken as the initial acquisition point, and the probability values ​​of each other pixel being selected as the next acquisition point are obtained. The pixel corresponding to the highest probability value is taken as the next acquisition point. This step is repeated until n acquisition points are selected.

[0024] The depth value of each acquisition point is used as its acquisition distance, and the angle between the line connecting each acquisition point and the principal point coordinates of the 3D depth camera and the camera optical axis of the 3D depth camera is used as its acquisition angle, and it is bound to the light intensity at the 3D depth camera at the corresponding timestamp.

[0025] A ranging unit is preset at the 3D depth camera, and the distance between each acquisition point and the 3D depth camera is obtained in real time using laser ranging technology through the ranging unit.

[0026] Furthermore, the process of constructing a distance correction model based on the acquisition angle, acquisition distance, and corresponding measurement distance of different acquisition points in different fused images includes:

[0027] Based on the acquisition angle, acquisition distance, measurement distance of different acquisition points in different fused images, and the illumination intensity under the corresponding timestamp, a distance correction set is generated and divided into the first training set and the first test set.

[0028] Construct a first convolutional neural network by using the acquisition angle, acquisition distance, and illumination intensity of different acquisition points in the first training set as input data for the first convolutional neural network, and using the corresponding measurement distance in the first training set as output data for the first convolutional neural network, and train the first convolutional neural network to obtain an initial first convolutional neural network.

[0029] The initial first convolutional neural network is validated using the first test set. The initial first convolutional neural network whose output is less than or equal to the preset first test error threshold is used as the distance correction model.

[0030] Furthermore, the process of constructing object recognition models based on different target objects and their appearance images to obtain target objects in the acquired images includes:

[0031] The target object refers to the object that is expected to be detected by the acquisition device, and the appearance image refers to a two-dimensional image of a single target object from different angles. An object recognition set is generated based on different target objects and their appearance images, and it is divided into a second training set and a second test set.

[0032] Construct a second convolutional neural network by using appearance images from different angles in the second training set as input data and the corresponding target objects in the second training set as output data, and train the second convolutional neural network to obtain an initial second convolutional neural network.

[0033] The initial second convolutional neural network is validated using the second test set. The initial second convolutional neural network whose output is less than or equal to the preset second test error threshold is used as the object recognition model. The RGB images captured by the 2D camera are input into the object recognition model in real time to identify the target objects.

[0034] Furthermore, the process of combining the fused images to obtain the acquisition distances at different locations of the target object, using a distance correction model to obtain the corrected distances at the corresponding locations, and then feeding them back includes:

[0035] When a target object is identified, a fused image is obtained based on the depth image under the corresponding timestamp. The acquisition angle and acquisition distance between the target object and the 3D depth camera are randomly obtained from several positions on the target object in the fused image.

[0036] The illumination intensity at the corresponding timestamp is input into the distance correction model to obtain the corrected distance at the corresponding location. The identified target object and the corrected distances at several locations are then fed back to the relevant personnel.

[0037] A machine vision-based object detection method includes the following steps:

[0038] Step S1: Deploy different acquisition devices and acquire corresponding acquisition images, and use the acquisition images to obtain the calibration parameters of the acquisition devices;

[0039] Step S2: Based on the calibration parameters of different acquisition devices, fuse the images acquired by different acquisition devices to obtain the corresponding fused image;

[0040] Step S3: Obtain the acquisition angle and acquisition distance of different acquisition points in the fused image, and use laser ranging to obtain the measurement distance of the corresponding acquisition point. Construct a distance correction model based on the acquisition angle, acquisition distance and corresponding measurement distance of different acquisition points in different fused images.

[0041] Step S4: Construct an object recognition model based on different target objects and their appearance images to obtain the target objects in the acquired images, combine them with the fused images to obtain the acquisition distances at different locations of the target objects, use the distance correction model to obtain the corrected distances at the corresponding locations and feed them back.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention addresses the disadvantage of low resolution in depth camera images by deploying both ordinary RGB cameras and depth cameras capable of acquiring depth values ​​in application scenarios, thereby converting the depth image into an RGB image plane. It achieves this by maintaining resolution while acquiring the depth value of each pixel. Laser ranging technology is used to obtain the actual distance between each pixel and the depth camera, which is then compared with the acquisition distance obtained from the image. A distance correction model is constructed based on the acquisition angle and corresponding lighting conditions, facilitating the direct output of the actual distance based on the acquisition distance and related parameters. This allows for the identification of target objects and determination of their current distance using only camera equipment. Furthermore, by combining images acquired at different timestamps, the movement of the target object can be determined, significantly improving the efficiency and accuracy of target detection. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0045] like Figure 1 As shown, a machine vision-based target detection system includes the following modules:

[0046] The image acquisition module is used to deploy different acquisition devices and acquire corresponding images, and to obtain the calibration parameters of the acquisition devices using the acquired images;

[0047] The image fusion module is used to fuse images acquired by different acquisition devices according to the calibration parameters of different acquisition devices to obtain the corresponding fused image;

[0048] The distance correction module is used to obtain the acquisition angle and acquisition distance of different acquisition points in the fused image, and to obtain the measurement distance of the corresponding acquisition point using laser ranging. Based on the acquisition angle, acquisition distance and corresponding measurement distance of different acquisition points in different fused images, a distance correction model is constructed.

[0049] The object recognition module is used to construct an object recognition model based on different target objects and their appearance images to obtain the target objects in the acquired images, combine them with the fused images to obtain the acquisition distances at different positions of the target objects, and use the distance correction model to obtain the corrected distances at the corresponding positions and provide feedback.

[0050] It should be further explained that, in the specific implementation process, different acquisition devices are deployed, and corresponding acquisition images are acquired respectively. The process of obtaining the calibration parameters of the acquisition devices using the acquired images includes:

[0051] The acquisition device includes a two-dimensional camera and a three-dimensional depth camera. The two acquisition devices are fixed on the same rigid bracket to minimize the baseline distance between them to reduce parallax. The two acquisition devices are connected by a hardware synchronization line to ensure that the timestamps of their RGB images and depth images are aligned.

[0052] The two-dimensional camera acquires RGB images of the application scene in real time, and the three-dimensional depth camera acquires depth images of the application scene in real time, and records the timestamps of both. The acquired images include RGB images and depth images.

[0053] For a 2D camera, at least 15 RGB images of different poses are acquired using a checkerboard calibration board to ensure that the images cover the entire field of view and have different rotations and translations. The cv2.calibrateCamera function in OpenCV is used for calibration to obtain the intrinsic parameter matrix (focal length, principal point coordinates) and distortion coefficients (radial and tangential distortion) of the 2D camera.

[0054] For a 3D depth camera, the same number of depth images are acquired using the same checkerboard calibration board, and the intrinsic parameter matrix and distortion coefficients of the 3D depth camera are obtained using the depth camera's SDK or OpenCV.

[0055] The relative position between the two acquisition devices is obtained through stereo calibration. Several sets of images containing the fields of view of both the 2D camera and the 3D depth camera are acquired. The stereoCalibrate function of OpenCV is used to obtain the rotation matrix R and translation vector T between the two acquisition devices by inputting the corner coordinates detected by the 2D camera and the 3D depth camera respectively.

[0056] It should be further explained that, in the specific implementation process, the process of fusing images acquired by different acquisition devices according to the calibration parameters of different acquisition devices to obtain the corresponding fused image includes:

[0057] The calibration parameters include the intrinsic parameter matrices and distortion coefficients of the 2D camera and the 3D depth camera, as well as the rotation matrix and translation vector between them. Based on the obtained rotation matrix and translation vector, a transformation matrix is ​​constructed from the coordinate system of the 3D depth camera to the coordinate system of the 2D camera.

[0058] For each pixel in the depth image (u d v d First, transform the coordinates of the 3D depth camera and its depth value d to the coordinate system of the 3D depth camera.

[0059] ;

[0060] Among them, (X) d Y d Z d (c) represents the coordinates of the corresponding pixel in the coordinate system of the 3D depth camera. x c y ) are the principal point coordinates of the 3D depth camera, (f x f y () is the focal length of the 3D depth camera;

[0061] Then, using the rotation matrix R and translation vector T, the coordinate values ​​(X) are... d Y d Z d Transform to the coordinate system of a 2D camera;

[0062] ;

[0063] Among them, (X) r Y r Z r ) represents the coordinates of the corresponding pixel in the coordinate system of the 2D camera. Finally, the coordinates (X, Y, F, Z) of each pixel in the coordinate system of the 2D camera are... r Y r Z r Projected onto the RGB image plane;

[0064] ;

[0065] Among them, (u r v r K represents the coordinates of the corresponding pixel in the RGB image plane. r This is the intrinsic parameter matrix of the two-dimensional camera;

[0066] Since the depth image and the RGB image have different resolutions, and the projected coordinate values ​​may be non-integer, bilinear interpolation is used to calculate the depth value for each pixel to generate a depth image with the same resolution as the RGB image, which is called the fused image. Its timestamp is kept consistent with the depth image before the conversion. Each depth value represents the depth of the corresponding RGB pixel.

[0067] It should be further explained that, in the specific implementation process, the process of obtaining the acquisition angle and acquisition distance of different acquisition points in the fused image, and using laser ranging to obtain the measurement distance of the corresponding acquisition points includes:

[0068] Taking any fused image as an example, on the fused image, any pixel is taken as the initial acquisition point, the shortest distance s between each other pixel and the current acquisition point is obtained, and the probability value R of each other pixel being selected as the next acquisition point is obtained.

[0069] ;

[0070] Where s represents the shortest distance, the pixel corresponding to the highest probability value is taken as the next acquisition point, and this step is repeated until n acquisition points are selected. The angle between the line connecting each acquisition point and the principal point coordinates of the 3D depth camera and the camera optical axis of the 3D depth camera (which is an acute angle) is taken as its acquisition angle.

[0071] Since the depth value of each pixel in the fused image represents its distance from the 3D depth camera, the depth value of each acquisition point is directly used as its acquisition distance and bound to the illumination intensity at the 3D depth camera at the corresponding timestamp.

[0072] Since the acquisition distance is obtained indirectly through image fusion, it is inevitably subject to errors due to environmental conditions. Therefore, a ranging unit is preset at the 3D depth camera. The actual distance between each acquisition point and the 3D depth camera is obtained in real time through the ranging unit using laser ranging technology, and is recorded as the measured distance.

[0073] It should be further explained that, in the specific implementation process, the process of constructing a distance correction model based on the acquisition angle, acquisition distance, and corresponding measurement distance of different acquisition points in different fused images includes:

[0074] Based on the acquisition angle, acquisition distance, measurement distance of different acquisition points in different fused images, and the illumination intensity under the corresponding timestamp, a distance correction set is generated and divided into the first training set and the first test set.

[0075] Construct a first convolutional neural network by using the acquisition angle, acquisition distance, and illumination intensity of different acquisition points in the first training set as input data for the first convolutional neural network, and using the corresponding measurement distance in the first training set as output data for the first convolutional neural network. Train the first convolutional neural network using the first training set to obtain an initial first convolutional neural network.

[0076] The initial first convolutional neural network is validated using the first test set. The initial first convolutional neural network whose output is less than or equal to the preset first test error threshold is used as the distance correction model.

[0077] It should be further explained that, in the specific implementation process, the process of constructing an object recognition model based on different target objects and their appearance images to obtain the target objects in the acquired images includes:

[0078] The target object refers to the object that is expected to be detected by the acquisition device. The appearance images of different target objects are obtained. The appearance images refer to two-dimensional images of a single target object from different angles, which can reflect the morphological characteristics of the corresponding target object.

[0079] An object recognition set is generated based on different target objects and their corresponding appearance images, and then divided into a second training set and a second test set.

[0080] Construct a second convolutional neural network by using appearance images from different angles in the second training set as input data and the corresponding target objects in the second training set as output data. Train the second convolutional neural network using the second training set to obtain an initial second convolutional neural network.

[0081] The initial second convolutional neural network is validated using the second test set. The initial second convolutional neural network whose output is less than or equal to the preset second test error threshold is used as the object recognition model. The RGB images captured by the 2D camera are input into the object recognition model in real time to identify the target objects.

[0082] It should be further explained that, in the specific implementation process, the process of obtaining the acquisition distance at different locations of the target object by combining the fused image, and using the distance correction model to obtain the corrected distance at the corresponding location and feeding it back includes:

[0083] When a target object is identified, a fused image is obtained based on the depth image under the corresponding timestamp. The acquisition angle and acquisition distance between the target object and the 3D depth camera are randomly obtained from several positions on the target object in the fused image.

[0084] The illumination intensity at the corresponding timestamp is input into the distance correction model to obtain the corrected distance at the corresponding location. The identified target object and the corrected distances at several locations are then fed back to the relevant personnel to inform them of the distance between the target object and the 3D depth camera at the corresponding timestamp.

[0085] The embodiments of the present invention also include a target detection method based on machine vision, comprising the following steps:

[0086] Step S1: Deploy different acquisition devices and acquire corresponding acquisition images, and use the acquisition images to obtain the calibration parameters of the acquisition devices;

[0087] Step S2: Based on the calibration parameters of different acquisition devices, fuse the images acquired by different acquisition devices to obtain the corresponding fused image;

[0088] Step S3: Obtain the acquisition angle and acquisition distance of different acquisition points in the fused image, and use laser ranging to obtain the measurement distance of the corresponding acquisition point. Construct a distance correction model based on the acquisition angle, acquisition distance and corresponding measurement distance of different acquisition points in different fused images.

[0089] Step S4: Construct an object recognition model based on different target objects and their appearance images to obtain the target objects in the acquired images, combine them with the fused images to obtain the acquisition distances at different locations of the target objects, use the distance correction model to obtain the corrected distances at the corresponding locations and feed them back.

[0090] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A target detection system based on machine vision, characterized in that, Includes the following modules: The image acquisition module is used to deploy different acquisition devices and acquire corresponding images, and to obtain the calibration parameters of the acquisition devices using the acquired images; The acquisition device includes a 2D camera and a 3D depth camera, which are connected by a hardware synchronization line to ensure that the timestamps of the RGB image and the depth image are aligned. The 2D camera acquires RGB images of the application scene in real time, and the 3D depth camera acquires depth images of the application scene in real time. The acquired images include RGB images and depth images. The calibration parameters include the intrinsic parameter matrices and distortion coefficients of the two-dimensional camera and the three-dimensional depth camera, as well as the rotation matrix and translation vector between them; The image fusion module is used to fuse images acquired by different acquisition devices according to the calibration parameters of different acquisition devices to obtain the corresponding fused image; The distance correction module is used to obtain the acquisition angle and acquisition distance of different acquisition points in the fused image, and to obtain the measurement distance of the corresponding acquisition point using laser ranging. Based on the acquisition angle, acquisition distance and corresponding measurement distance of different acquisition points in different fused images, a distance correction model is constructed. The object recognition module is used to build an object recognition model based on different target objects and their appearance images to obtain the target objects in the acquired images, combine them with the fused images to obtain the acquisition distances at different positions of the target objects, use the distance correction model to obtain the corrected distances at the corresponding positions and feed them back. The process of obtaining the acquisition angle, acquisition distance, and measurement distance of different acquisition points in the fused image includes: In a single fused image, any pixel is taken as the initial acquisition point, and the probability values ​​of each other pixel being selected as the next acquisition point are obtained. The pixel corresponding to the highest probability value is taken as the next acquisition point. This step is repeated until n acquisition points are selected. The depth value of each acquisition point is used as its acquisition distance, and the angle between the line connecting each acquisition point and the principal point coordinates of the 3D depth camera and the camera optical axis of the 3D depth camera is used as its acquisition angle, and it is bound to the light intensity at the 3D depth camera at the corresponding timestamp. A ranging unit is preset at the 3D depth camera, and the distance between each acquisition point and the 3D depth camera is obtained in real time using laser ranging technology through the ranging unit. The process of constructing a distance correction model includes: Based on the acquisition angle, acquisition distance, measurement distance of different acquisition points in different fused images, and the illumination intensity under the corresponding timestamp, a distance correction set is generated and divided into the first training set and the first test set. Construct a first convolutional neural network by using the acquisition angle, acquisition distance, and illumination intensity of different acquisition points in the first training set as input data for the first convolutional neural network, and using the corresponding measurement distance in the first training set as output data for the first convolutional neural network, and train the first convolutional neural network to obtain an initial first convolutional neural network. The initial first convolutional neural network is validated using the first test set. The initial first convolutional neural network whose output is less than or equal to the preset first test error threshold is used as the distance correction model.

2. The target detection system based on machine vision according to claim 1, characterized in that, The process of deploying acquisition equipment and obtaining acquired images and the calibration parameters of the acquisition equipment includes: For a 2D camera, several RGB images of different poses are acquired by it against a checkerboard calibration board, and OpenCV is used for calibration to obtain the intrinsic parameter matrix and distortion coefficients of the 2D camera. For a 3D depth camera, the same number of depth images are acquired by it against the same checkerboard calibration board, and OpenCV is used to obtain the intrinsic parameter matrix and distortion coefficients of the 3D depth camera. The relative position between the two acquisition devices is obtained by stereo calibration, and several sets of images containing the fields of view of both the 2D camera and the 3D depth camera are acquired. The corner coordinates detected by the 2D camera and the 3D depth camera are input using OpenCV to obtain the rotation matrix R and translation vector T between the two acquisition devices.

3. The target detection system based on machine vision according to claim 2, characterized in that, The process of fusing images acquired by different acquisition devices to obtain a fused image includes: The calibration parameters include the intrinsic parameter matrices and distortion coefficients of the 2D camera and the 3D depth camera, as well as the rotation matrix and translation vector between them, for each pixel in the depth image (u d v d ) and its depth value d, and transform them to the coordinate system of the 3D depth camera; ； (X d Y d Z d (c) represents the coordinates of the corresponding pixel in the coordinate system of the 3D depth camera. x c y ) are the principal point coordinates of the 3D depth camera, (f x f y ) is the focal length of the 3D depth camera. The rotation matrix R and translation vector T are used to transform the coordinate values ​​(X...)... d Y d Z d Transform to the coordinate system of a 2D camera; ； (X r Y r Z r ) represents the coordinates of the corresponding pixel in the coordinate system of the 2D camera. Finally, the coordinates (X, Y, F, Z) of each pixel in the coordinate system of the 2D camera are... r Y r Z r Projected onto the RGB image plane; ； (u r v r K represents the coordinates of the corresponding pixel in the RGB image plane. r Given the intrinsic parameter matrix of a 2D camera, a depth value is calculated for each pixel using bilinear interpolation to generate a fused image with the same resolution as the RGB image, and its timestamp is kept consistent with the depth image before conversion.

4. The target detection system based on machine vision according to claim 3, characterized in that, The process of building an object recognition model to obtain target objects in acquired images includes: The target object refers to the object that is expected to be detected by the acquisition device, and the appearance image refers to a two-dimensional image of a single target object from different angles. An object recognition set is generated based on different target objects and their appearance images, and it is divided into a second training set and a second test set. Construct a second convolutional neural network by using appearance images from different angles in the second training set as input data and the corresponding target objects in the second training set as output data, and train the second convolutional neural network to obtain an initial second convolutional neural network. The initial second convolutional neural network is validated using the second test set. The initial second convolutional neural network whose output is less than or equal to the preset second test error threshold is used as the object recognition model. The RGB images captured by the 2D camera are input into the object recognition model in real time to identify the target objects.

5. The target detection system based on machine vision according to claim 4, characterized in that, The process of obtaining and feeding back the corrected distances at different locations of the target object includes: When a target object is identified, a fused image is obtained based on the depth image under the corresponding timestamp. The acquisition angle and acquisition distance between the target object and the 3D depth camera are randomly obtained from several positions on the target object in the fused image. The illumination intensity at the corresponding timestamp is input into the distance correction model to obtain the corrected distance at the corresponding location. The identified target object and the corrected distances at several locations are then fed back to the relevant personnel.

6. A machine vision-based target detection method, implemented using the target detection system according to any one of claims 1-5, characterized in that, The method includes: Step S1: Deploy different acquisition devices and acquire corresponding acquisition images, and use the acquisition images to obtain the calibration parameters of the acquisition devices; Step S2: Based on the calibration parameters of different acquisition devices, fuse the images acquired by different acquisition devices to obtain the corresponding fused image; Step S3: Obtain the acquisition angle and acquisition distance of different acquisition points in the fused image, and use laser ranging to obtain the measurement distance of the corresponding acquisition point. Construct a distance correction model based on the acquisition angle, acquisition distance and corresponding measurement distance of different acquisition points in different fused images. Step S4: Construct an object recognition model based on different target objects and their appearance images to obtain the target objects in the acquired images, combine them with the fused images to obtain the acquisition distances at different locations of the target objects, use the distance correction model to obtain the corrected distances at the corresponding locations and feed them back.

Images