WO2019148453A1 - Procédé d'apprentissage de modèle de reconnaissance de cible, procédé de reconnaissance de cible, appareil et robot - Google Patents

Procédé d'apprentissage de modèle de reconnaissance de cible, procédé de reconnaissance de cible, appareil et robot Download PDF

Info

Publication number
WO2019148453A1
WO2019148453A1 PCT/CN2018/075134 CN2018075134W WO2019148453A1 WO 2019148453 A1 WO2019148453 A1 WO 2019148453A1 CN 2018075134 W CN2018075134 W CN 2018075134W WO 2019148453 A1 WO2019148453 A1 WO 2019148453A1
Authority
WO
WIPO (PCT)
Prior art keywords
identified
target
image
recognition model
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2018/075134
Other languages
English (en)
Chinese (zh)
Inventor
张�浩
吴启帆
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen Dorabot Inc
Original Assignee
Shenzhen Dorabot Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen Dorabot Inc filed Critical Shenzhen Dorabot Inc
Priority to PCT/CN2018/075134 priority Critical patent/WO2019148453A1/fr
Priority to CN201880002216.8A priority patent/CN109313710A/zh
Publication of WO2019148453A1 publication Critical patent/WO2019148453A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour

Definitions

  • the present application relates to the field of target recognition, and in particular to a target recognition model training method, a target recognition method, a device, and a robot.
  • the recognition model when the recognition model is trained, the acquired target sample image is first input into the recognition model, and the marked image is output, and the labeled image is compared with the standard image, and the error is fed back to the recognition.
  • the model is modified, and after the target position is changed, the above process is cycled until the recognition model is trained, and then the recognition model is used for target recognition.
  • the standard image needs to manually outline the target after acquiring the target image, and the manual drawing process takes a long time and has low efficiency.
  • the present application provides a target recognition model training method, a target recognition method, a device, and a robot, which can solve the problem of long time and low efficiency in acquiring standard images by manually sketching contours.
  • the first technical solution adopted by the present application is: providing a target recognition model training method, comprising: acquiring a sample image of a target to be identified; inputting the sample image into a recognition model, and outputting the segmented data after the recognition; Comparing the identified segmentation data with the standard data of the standard image to obtain the recognition error; feeding the recognition error to the recognition model and correcting the recognition model; wherein, when the standard image is acquired, the edge of the target to be identified is coated with the fluorescent material, according to the standard The color development of the fluorescent material in the image, and the standard data of the target to be identified is obtained.
  • the second technical solution adopted by the present application is: providing a target recognition method, comprising: acquiring an image of the goods to be identified; inputting the image into the target recognition model of the training completion, and outputting the identified goods to be identified. Segmentation data; using the segmentation data to obtain data of the goods to be identified, the data of the goods to be identified is used to plan the position and/or posture of the goods to be identified; wherein the target recognition model passes the target recognition model as described above Training methods are trained.
  • the third technical solution adopted by the present application is: providing a target identification device, comprising: a communication circuit and a processor connected to each other; the communication circuit is configured to acquire a sample image of the target to be identified; the processor is configured to The sampled image is input into the recognition model, and the recognized segmentation data is output, the identified segmentation data is compared with the standard image standard data, the recognition error is obtained, the recognition error is fed back to the recognition model, and the recognition model is corrected; wherein the standard image is acquired; The edge of the object to be identified is coated with a fluorescent material, and standard data of the object to be identified is obtained according to the color development of the fluorescent material in the standard image.
  • the fourth technical solution adopted by the present application is to provide a robot comprising: an interconnected mechanical arm and a target recognition device as described above; the mechanical arm is used for the target identified by the target recognition device Data, planning the location and/or posture of the target to operate the target object.
  • the beneficial effects of the present application are: different from the prior art, in some embodiments of the present application, in the process of training the target recognition model, when the standard image is acquired, the target edge is coated with a fluorescent material according to the fluorescent material in the standard image. Color development, the standard data of the target to be identified can be obtained, so that the standard data can be directly obtained according to the collected standard image, and the standard image can be obtained by manually sketching after acquiring the image, thereby obtaining standard data, thereby saving manual sketching. Time, which in turn increases the speed and efficiency of model training.
  • FIG. 1 is a schematic flow chart of a first embodiment of a target recognition model training method of the present application
  • FIG. 2 is a schematic flow chart of a second embodiment of a target recognition model training method of the present application
  • FIG. 3 is a schematic diagram of an application scenario when a standard image is acquired in a second embodiment of the target recognition model training method of the present application;
  • FIG. 4 is a schematic diagram of another application scenario when acquiring a standard image in the second embodiment of the target recognition model training method of the present application;
  • FIG. 5 is a schematic flowchart of an embodiment of an object identification method according to the present application.
  • FIG. 6 is a schematic structural diagram of a first embodiment of a target recognition device of the present application.
  • FIG. 7 is a schematic structural diagram of a second embodiment of the target recognition device of the present application.
  • FIG. 8 is a schematic structural diagram of a third embodiment of the target recognition device of the present application.
  • FIG. 9 is a schematic structural view of an embodiment of a robot of the present application.
  • the first embodiment of the target recognition model training method of the present application includes:
  • S11 Acquire a sample image of the target to be identified.
  • the target to be identified includes, but is not limited to, a cargo, a person, or an animal, and may be another object that needs to be identified.
  • the number of the objects to be identified may be one or two or more, and is not specifically limited herein. .
  • the object to be identified is described by taking the cargo as an example.
  • S12 Input the sampled image into the recognition model, and output the identified segmentation data.
  • the identification model is a model for identifying the target to be identified, and the type of the recognition model may be determined according to actual needs, such as a neural network recognition model.
  • the identified segmentation data may be a region coordinate sequence of the object to be identified segmented from the sample image, or an edge coordinate sequence of the target to be identified, or a sequence of orthographic projection edge coordinates of the object to be identified, or may be The image after the target to be identified is marked on the sampled image, or the image of the region to be identified is divided, and is not specifically limited herein.
  • the standard image is a pre-acquired image that has been marked with a target to be identified, or an image that has been segmented out to be identified.
  • the target edge is coated with a fluorescent material, and according to the color development of the fluorescent material in the standard image, standard data of the actual target to be identified can be obtained, and the description form corresponding to the segmentation data can be used for the target to be identified.
  • the sequence of regional coordinates, or the sequence of edge coordinates of the target to be identified, or the sequence of orthographic projection edges of the target to be identified may also be the obtained standard image with fluorescent material color development, or the fluorescent material color identification Identify the area image of the target.
  • the fluorescent material may be a material that exhibits a preset color under natural light, and the preset color may be a color different from a non-edge region, such as black or red, and the fluorescent material may also be colorless under natural light, but In some frequency bands, the light is illuminated under the illumination, and the material of the preset color is displayed, for example, the color is red under the infrared light.
  • the specific material of the fluorescent material may be determined according to actual needs, and is not specifically limited herein.
  • the identification error may be a difference between the identified segmentation data output by the recognition model and the standard data of the target to be identified in the standard image, for example, the segmentation data and the standard data are cargo area data, cargo edge data or cargo projection edge data. , the difference may be the phase difference of the coordinate sequence of the segmentation data and the standard data.
  • modifying the recognition model may be a parameter of modifying the recognition model.
  • the recognition model is a neural network model
  • the weight of the neural network is corrected.
  • a sampling image of a target to be identified may be captured in real time by using a photographing device, such as a camera, a camera, a vision sensor, or the like, for example, taking a sample image of a pile of goods stacked by manual/machine, and then The sampled image is input to a recognition model, which can perform segmentation processing on the sampled image to identify a plurality of goods to be divided into a plurality of goods included therein to facilitate subsequent operations on the respective goods. For example, using edge recognition, surface recognition, region recognition, and the like, the target to be identified in the sampled image is identified, thereby obtaining segmentation data of the object to be identified, such as a coordinate sequence of the edge or a coordinate sequence of the region.
  • the segmentation data is then compared with the standard data of the same pile of goods acquired in advance and the standard image of the same acquisition angle of view.
  • the edge of the goods to be identified has been previously smeared with a fluorescent material, and when the pile of goods is photographed, the standard data of the preset color displayed on the edge of the goods to be identified can be directly obtained.
  • the standard data is a recognition result that the recognition model is expected to be trained, that is, the identification model is provided as a standard standard data for segmenting the target to be identified.
  • the difference between the two is calculated, for example, the phase difference of the coordinate sequence, and the feedback is sent to the identification.
  • the model corrects the parameters of the recognition model to improve the recognition accuracy of the recognition model.
  • the above steps may be repeated to continue to correct the recognition model until the recognition model reaches a preset requirement, for example, the accuracy is greater than a preset threshold (for example, 80%). It means that the recognition model has been trained, and the modified recognition model can be directly used for target recognition.
  • the object to be identified may be all objects included in the image, or a specific object, or a specific object, and may be flexibly set according to the requirements of the application scenario.
  • the target edge is coated with a fluorescent material, and the standard data of the target to be identified is obtained according to the color development of the fluorescent material in the standard image. Therefore, the standard data can be directly obtained according to the collected standard image, and after the image is acquired, the standard image is obtained by manual sketching, thereby obtaining standard data. Therefore, the method of the embodiment saves the time of manual sketching, thereby improving the speed and efficiency of the model training. In this way, the sampling image and the standard image can be obtained by using the above-mentioned photographing device, and the training of the target recognition model can be directly performed. Without the manual operation mode, after obtaining the sampled image, it is necessary to manually extract the standard image after the goods to be identified, which results in a large amount of time.
  • the second embodiment of the target recognition model training method of the present application is based on the first embodiment of the target recognition model training method of the present application.
  • the method includes:
  • S131 Acquire an image of a target to be recognized whose edge is smeared with a fluorescent material by using a photographing device to obtain the standard image.
  • the fluorescent material may directly be a material that exhibits a certain color (for example, black) under natural light, and the fluorescent material may also be a colorless fluorescent material, which is colorless under natural light, but will be light under certain frequency bands. Display a specific color, such as red.
  • the photographing device may be a general camera, and the fluorescent material is rendered in a desired color by being equipped with a specific light source device. It may also be a camera with a special function, such as a camera with a light source that causes the fluorescent material to display a color.
  • the light source device may be turned on to present a specific light source, so that the fluorescent material displays a specific color, and is directly photographed by a normal camera.
  • An image of the object to be identified for example, an image of a pile of goods to be identified, wherein the edge of the goods to be identified is coated with a fluorescent material, and therefore, in the captured image, the edge of the goods to be identified will display the color of the fluorescent material So that the image can be used directly as a standard image.
  • the method further includes:
  • S130 illuminating the fluorescent material with light of a preset frequency band, so that the fluorescent material displays a preset color.
  • the preset frequency band is a frequency band that causes the fluorescent material to display a predetermined color of light
  • the same fluorescent material may display different colors under different light beams, and different kinds of fluorescent materials are illuminated by light of the same frequency band. Different colors can also be displayed, so the specific value of the preset frequency band can be selected according to the type and characteristics of the fluorescent material.
  • the light source 301 when acquiring a standard image of the goods to be identified, the light source 301 may be used to illuminate a set of to-be-identified goods 302 that are placed in a predetermined frequency band, thereby making it possible to make
  • the same fluorescent material applied to the edge of the stack of to-be-identified goods 302 displays a preset color (for example, red), and the standard image with the border color can be directly obtained by the photographing device 303 (for example, a camera).
  • the number of targets to be identified in the standard image is at least two, and different fluorescent materials may be applied to the edges of adjacent targets to be identified, including, under the illumination of a specific frequency band, different fluorescent materials respectively. Displaying different colors; or different fluorescent materials are separately colored under illumination of different frequency bands, so that the edges of the adjacent objects to be identified display different colors, thereby comparing edge contours of adjacent objects to be identified It is relatively easy to distinguish adjacent targets to be identified, and further improve the recognition accuracy of the target recognition model.
  • the robot is determined to be the face of a cargo based on a closed line, assuming that the closed dashed line 402 is the face of a cargo A, but the closed dashed line 402 is essentially only the face of the cargo A, which is obscured by a piece if the system or robot is based on This plan is planned to be held. It is possible to obtain the face that is directly perpendicular to the closed dotted line 402 and move it out in the vertical direction. This may cause the occluded part of the cargo A to collide with the front cargo C, which may cause the task to fail.
  • different types of inks may be applied on the edges of adjacently placed goods (for example, A, B, and C), so that the light of the same frequency band emitted by the light source 403 illuminates the goods A, B, and C.
  • the edge part displays different colors, so that the edges of different goods in the standard image can be distinguished by different colors, so as to obtain more accurate standard data of each goods, so that the recognition error can be more accurately obtained, and then the model is improved when training the recognition model. Identification accuracy.
  • the edge of the adjacently placed goods may also be coated with a fluorescent material that is colored under the illumination of different frequency bands, and when the standard image is acquired, the fluorescent material is irradiated with light of different frequency bands, so that the phase
  • the fluorescent material of the adjacent goods is displayed under the light source of different frequency bands.
  • some boxes are painted with red fluorescent material under infrared light, and some boxes are coated with fluorescent materials that are colored red or other colors under ultraviolet light.
  • the shooting angle of the shooting device is fixed, and the infrared light can be respectively irradiated to obtain the standard image A, and the ultraviolet light can be used to obtain the standard image B.
  • the ultraviolet light can be irradiated to obtain the standard image B, and then the infrared light can be irradiated to obtain the standard image A.
  • standard image A and the standard image B standard data of different boxes can be separately obtained. It can be understood that different boxes can be smeared by fluorescent materials having a plurality of different frequency bands to obtain accurate standard data of the box. Further, when the boxes are stacked and fixed, the shooting angle of the photographing device is fixed, and the step of acquiring the sampled image may be performed before, during or after the standard image corresponding to the color light source of the plurality of frequency bands is respectively acquired.
  • the photographing device sets a viewing angle to obtain the sampled image and the standard image together. For example, after capturing a sampled image under a non-specific light source, a specific light source can be turned on immediately to obtain a standard image. It can be understood that the order of obtaining the two is not limited. Thereby, the time for acquiring the standard image is effectively reduced. Furthermore, the stacking mode or the viewing angle can be changed to acquire the next set of sampled images and standard images, which improves the overall training efficiency.
  • the recognition model uses the feature data as a recognition condition for outputting the segmentation data. Therefore, the recognition model trained in this manner can more accurately identify the goods to be identified that do not have fluorescent materials in practical applications. It can be understood that, in other embodiments, if the object to be identified is set as an object having a fluorescent material in an actual application, the training recognition model may be used to identify the feature data of the fluorescent material as the segmentation condition.
  • an embodiment of the object recognition method of the present application includes:
  • S22 Input the image into the target recognition model of the training completion, and output the segmentation data of the identified goods to be identified.
  • S23 Obtain data of the to-be-identified goods by using the segmentation data, where the data of the goods to be identified is used to plan the position and/or posture of the goods to be identified.
  • the target recognition model is trained by the method provided by the first or second embodiment of the target recognition model training method of the present application.
  • the robot before the robot acquires the goods, it is necessary to acquire data of the goods, such as spatial data and/or the holding surface, in order to plan the posture of the robot for obtaining the goods. Therefore, before the goods are held, the goods need to be identified first.
  • an image of the goods to be identified can be acquired by using a photographing device (for example, a camera), the image is input into the trained target recognition model, and the goods to be identified are output through the target recognition model.
  • the segmentation data, and the robot can obtain spatial data of each item to be identified according to the segmentation data, such as the position and/or posture of each face, the position of the edge contour, and the size information of the length, width, and height, etc., according to the spatial data. Plan the acquisition surface of the cargo and the robot to obtain the position and/or posture of the cargo.
  • the recognition model used for the target recognition is obtained during the training process.
  • the standard image is acquired, the target edge is coated with a fluorescent material, and the standard data of the target to be identified is obtained according to the color development of the fluorescent material in the standard image. Therefore, the standard data can be directly obtained according to the collected standard image, and after the image is acquired, the standard image is obtained by manual sketching, and then the standard data is obtained. Therefore, the method of the embodiment saves the time of artificial sketching, thereby improving the speed and efficiency of the model training, and finally is beneficial to planning the posture of the obtained goods and improving the holding efficiency.
  • the first embodiment of the object recognition device 60 of the present application includes: a communication circuit 601 and a processor 602 connected to each other;
  • the communication circuit 601 is configured to acquire a sample image of the target to be identified; the processor 602 is configured to input the sample image into the recognition model, output the identified segmentation data, and compare the identified segmentation data with standard data of the standard image. Obtaining an identification error, feeding the recognition error to the recognition model, and correcting the recognition model;
  • the edge of the target to be identified is coated with a fluorescent material, and standard data of the target to be identified is obtained according to the color development of the fluorescent material in the standard image.
  • the processor 602 controls the operation of the target recognition device 60, and the processor 602 may also be referred to as a CPU (Central Processing). Unit, central processing unit).
  • Processor 602 may be an integrated circuit chip with signal processing capabilities.
  • the processor 602 can also be a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component.
  • the general purpose processor may be a microprocessor or the processor or any conventional processor or the like.
  • the processor 602 is further configured to: input the sampled image of the target to be identified into the neural network recognition model, identify the target to be identified by using the neural network recognition model, and output the segmentation data of the target to be identified.
  • the segmentation data includes: a region coordinate sequence of the target to be identified, or an edge coordinate sequence of the target to be identified, or a sequence of orthographic projection edge coordinates of the object to be identified, or an image on the sample image after the target to be identified is marked, or segmented The area image of the target to be identified.
  • the processor 602 is further configured to: return to perform the step of acquiring the sampled image of the target to be identified, to obtain the transformed image or the sampled image of the target to be identified after the acquired captured angle of view, until the accuracy of the model is recognized When the preset threshold is exceeded, the recognition model training is completed.
  • the communication circuit 601 is further configured to acquire an image of the to-be-identified item; the processor 602 is further configured to input the image into the training-completed recognition model, and output the segmented data of the identified item to be identified; and then use the segmentation data.
  • the communication circuit 601 is further configured to acquire an image of the to-be-identified item; the processor 602 is further configured to input the image into the training-completed recognition model, and output the segmented data of the identified item to be identified; and then use the segmentation data.
  • Obtain the data of the goods to be identified, and the data of the goods to be identified is used to plan the posture of the goods to be identified.
  • the data of the goods to be identified includes the holding surface of the goods to be identified, and/or the spatial data of the goods to be identified, for example, the coordinate information of the goods to be identified in space, the length, width and height information of the space, etc. are used to describe the space. Information, etc.
  • the processor 602 is further configured to mark an edge of the to-be-identified item in the image, and output the marked image.
  • the target recognition device 60 can acquire the sample image and the standard image of the object to be identified from other external devices or systems through the communication circuit 601.
  • the specific process of the target recognition device 60 using the processor 602 to train the target recognition model may refer to the method provided by the first or second embodiment of the target recognition model training method of the present application and the content of an embodiment of the target recognition method of the present application. It will not be repeated.
  • the target recognition device when the target recognition device performs the target recognition model training process, when the standard image is acquired, the target edge is coated with a fluorescent material, and the standard data of the target to be identified is obtained according to the color development of the fluorescent material in the standard image. Therefore, the standard data can be directly obtained according to the collected standard image, and after the image is acquired, the standard image is obtained by manual sketching, and then the standard data is obtained. Therefore, the method of the embodiment saves the time of manual sketching, thereby improving the speed and efficiency of the model training.
  • the target recognition device may also acquire the standard image and the sampled image using a photographing device connected to the communication circuit.
  • the structure of the second embodiment of the object recognition device of the present application is similar to the structure of the first embodiment of the object recognition device of the present application, and is not described here again, except that the target identification device of the embodiment is different.
  • the 70 further includes: a photographing device 603 connected to the communication circuit 601 for photographing an image of the object to be identified whose edge is coated with the fluorescent material to acquire the standard image.
  • the photographing device 603 may be a general camera, a video camera, a 3D camera, or the like, or may be a camera having a special function, such as a camera with a light source that causes the fluorescent material to display a color.
  • the fluorescent material may directly be a material that exhibits a certain color (for example, black) under natural light, and the fluorescent material may also be a colorless fluorescent material, which is colorless under natural light, but displays specificity under light of certain frequency bands. Color, such as red.
  • the number of targets to be identified in the standard image may be one or two or more.
  • different fluorescent materials are applied to the edges of the adjacent target to be identified, including different fluorescent materials respectively displaying different colors under illumination of a specific frequency band; or under different light sources Different fluorescent materials develop color separately.
  • the edges of the adjacent objects to be identified may be displayed in different colors, so that when the edge contours of the adjacent objects to be identified are compared, the adjacent objects to be identified may be easily distinguished, thereby further improving the recognition accuracy of the target recognition model. .
  • the photographing device 603 is further configured to capture a sample image of the target to be identified, and transmit the sample image to the communication circuit 601.
  • step S131 of the second embodiment of the target recognition model training method of the present application reference may be made to the content of step S131 of the second embodiment of the target recognition model training method of the present application, which is not repeated here.
  • different imaging devices may also be used to separately acquire the sampled image and the standard image of the target to be identified.
  • the target recognition device may first use the light source to emit light of a preset frequency band such that the fluorescent material at the edge of the target to be recognized displays a preset color to obtain a standard image.
  • the structure of the third embodiment of the object recognition device of the present application is similar to the structure of the second embodiment of the object recognition device of the present application, and details are not described herein, except that the target identification device of the embodiment is different.
  • the 80 further includes a light source 604 connected to the photographing device 603 for generating light of a preset frequency band to illuminate the fluorescent material, so that when the photographing device 603 captures an image of the object to be recognized, the fluorescent material displays a preset color.
  • the preset frequency band is a frequency band that causes the fluorescent material to display a predetermined color of light
  • the same fluorescent material may display different colors under different light beams, and different kinds of fluorescent materials are illuminated by light of the same frequency band. Different colors can also be displayed, so the specific value of the preset frequency band can be selected according to the type and characteristics of the fluorescent material.
  • the specific process of the target recognition device 80 illuminating the target to be recognized by the light source 604 and acquiring the standard image by using the photographing device 603 can refer to the content of step S130 of the second embodiment of the target recognition model training method of the present application. Repeat again.
  • an embodiment of the robot 90 of the present application includes: a robot arm 901 and a target recognition device 902 that are connected to each other.
  • the structure and function of the target identification device 902 can refer to the content of any one of the first to third embodiments of the target identification device of the present application, and is not repeated here.
  • the robot arm 901 is configured to plan a pose of the target to acquire the target object according to the data of the target recognized by the target recognition device 902.
  • the robot arm 901 can be provided with an end effector (not shown). After the robot 90 recognizes the spatial data of the target according to the target recognition device 902, the robot can plan to obtain the optimal pose of the target, thereby controlling the robot. The end effector of 901 acquires the target.
  • the robot uses the target recognition device to train the target recognition model
  • the standard image is acquired, the target edge is coated with a fluorescent material, and the standard data of the target to be identified is obtained according to the color development of the fluorescent material in the standard image. Therefore, the standard data can be directly obtained according to the collected standard image, and after the image is acquired, the standard image is obtained by manual sketching, and then the standard data is obtained. Therefore, the method of the embodiment saves the time of manual sketching, thereby improving the speed and efficiency of model training.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

La présente invention concerne un procédé d'apprentissage d'un modèle de reconnaissance de cible, un procédé de reconnaissance de cible, un appareil et un robot. Le procédé d'apprentissage d'un modèle de reconnaissance de cible consiste : à obtenir une image d'échantillon d'une cible à reconnaître ; à entrer l'image d'échantillon dans un modèle de reconnaissance, et à émettre en sortie des données de segmentation post-reconnaissance ; à comparer les données de segmentation à des données standard d'une image standard afin d'obtenir une erreur de reconnaissance ; à renvoyer l'erreur de reconnaissance au modèle de reconnaissance, et à corriger le modèle de reconnaissance, les bords de la cible étant marqués, lorsqu'une image standard est obtenue, avec un matériau fluorescent, et les données standard de la cible à reconnaître étant obtenues en fonction de la couleur résultante du matériau fluorescent dans l'image standard. La présente invention améliore la vitesse et l'efficacité d'apprentissage de modèle.
PCT/CN2018/075134 2018-02-02 2018-02-02 Procédé d'apprentissage de modèle de reconnaissance de cible, procédé de reconnaissance de cible, appareil et robot Ceased WO2019148453A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2018/075134 WO2019148453A1 (fr) 2018-02-02 2018-02-02 Procédé d'apprentissage de modèle de reconnaissance de cible, procédé de reconnaissance de cible, appareil et robot
CN201880002216.8A CN109313710A (zh) 2018-02-02 2018-02-02 目标识别模型训练方法、目标识别方法、设备及机器人

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/075134 WO2019148453A1 (fr) 2018-02-02 2018-02-02 Procédé d'apprentissage de modèle de reconnaissance de cible, procédé de reconnaissance de cible, appareil et robot

Publications (1)

Publication Number Publication Date
WO2019148453A1 true WO2019148453A1 (fr) 2019-08-08

Family

ID=65221748

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/075134 Ceased WO2019148453A1 (fr) 2018-02-02 2018-02-02 Procédé d'apprentissage de modèle de reconnaissance de cible, procédé de reconnaissance de cible, appareil et robot

Country Status (2)

Country Link
CN (1) CN109313710A (fr)
WO (1) WO2019148453A1 (fr)

Cited By (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112308003A (zh) * 2020-11-06 2021-02-02 中冶赛迪重庆信息技术有限公司 一种废钢货车装载状态识别方法、系统、设备及介质
CN113561181A (zh) * 2021-08-04 2021-10-29 北京京东乾石科技有限公司 目标检测模型的更新方法、装置及系统
CN114241450A (zh) * 2021-10-27 2022-03-25 上海艾豚科技有限公司 识别汽车内饰件上表皮和缝线颜色的方法
CN114264607A (zh) * 2021-12-29 2022-04-01 佛山市帆思科材料技术有限公司 基于机器视觉的瓷砖色差在线检测系统与方法
CN114310954A (zh) * 2021-12-31 2022-04-12 北京理工大学 一种护理机器人自适应升降控制方法和系统
CN114627288A (zh) * 2022-02-23 2022-06-14 厦门聚视智创科技有限公司 一种复杂3d背景下的背景分割目标识别方法
CN114693921A (zh) * 2020-12-25 2022-07-01 东声(苏州)智能科技有限公司 神经网络模型优化的评估方法、存储介质和计算机
CN115019300A (zh) * 2022-08-09 2022-09-06 成都运荔枝科技有限公司 用于自动化仓库货物识别的方法
CN116075864A (zh) * 2020-08-28 2023-05-05 先讯美资电子有限责任公司 视觉媒体中的人类型的分类
CN121582602A (zh) * 2025-11-28 2026-02-27 北京华软恒信科技发展有限公司 基于计算机视觉的物资盘点机器自动识别计数方法

Families Citing this family (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110046545A (zh) * 2019-03-05 2019-07-23 深兰科技（上海）有限公司 一种物品摆放系统、方法、装置、电子设备及存储介质
CN110222235A (zh) * 2019-06-11 2019-09-10 百度在线网络技术（北京）有限公司 三维立体内容显示方法、装置、设备及存储介质
CN113361511B (zh) * 2020-03-05 2024-10-01 顺丰科技有限公司 修正模型的建立方法、装置、设备及计算机可读存储介质
CN111746169A (zh) * 2020-05-26 2020-10-09 湖南天琪智慧印刷有限公司 定位标记热敏纸及其定位方法与其制作方法
CN111708794B (zh) * 2020-06-22 2024-05-03 中国平安财产保险股份有限公司 基于大数据平台的数据比对方法、装置和计算机设备
CN114596236B (zh) * 2020-12-04 2024-09-10 国网智能科技股份有限公司 一种密闭腔体低照度图像增强方法及系统
CN112388655B (zh) * 2020-12-04 2021-06-04 齐鲁工业大学 一种基于触觉振动信号与视觉图像融合的抓取物识别方法
CN112200172B (zh) * 2020-12-07 2021-02-19 天津天瞳威势电子科技有限公司 一种可行驶区域的检测方法及装置
CN114911221B (zh) * 2021-02-09 2023-11-28 北京小米机器人技术有限公司 机器人的控制方法、装置及机器人
CN116645413B (zh) * 2023-06-02 2024-06-07 湖州丽天智能科技有限公司 一种光伏电池板位置识别方法、系统及光伏机器人
CN120976859B (zh) * 2025-08-07 2026-03-27 国网山东省电力公司菏泽供电公司 一种基于图像处理技术的变电站内多目标识别和追踪方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1738426A (zh) * 2005-09-09 2006-02-22 南京大学 一种视频运动目标分割与跟踪方法
CN101354359A (zh) * 2008-09-04 2009-01-28 湖南大学 药液内运动可见异物的检测跟踪与识别方法
CN101573959A (zh) * 2006-11-01 2009-11-04 索尼株式会社 运动图片中的片段跟踪
CN105678332A (zh) * 2016-01-08 2016-06-15 昆明理工大学 火焰图像cnn识别建模的转炉炼钢终点判断方法及系统
CN106023220A (zh) * 2016-05-26 2016-10-12 史方 一种基于深度学习的车辆外观部件图像分割方法
CN106874914A (zh) * 2017-01-12 2017-06-20 华南理工大学 一种基于深度卷积神经网络的工业机械臂视觉控制方法

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9213920B2 (en) * 2010-05-28 2015-12-15 Zazzle.Com, Inc. Using infrared imaging to create digital images for use in product customization
MX340723B (es) * 2011-07-28 2016-07-22 Medetect Ab Método para proporcionar imágenes de una sección de tejido.
CN103186901A (zh) * 2013-03-29 2013-07-03 中国人民解放军第三军医大学 全自动图像分割方法
US9870613B2 (en) * 2014-11-05 2018-01-16 Carestream Health, Inc. Detection of tooth condition using reflectance images with red and green fluorescence
US9953217B2 (en) * 2015-11-30 2018-04-24 International Business Machines Corporation System and method for pose-aware feature learning
CN105654045B (zh) * 2015-12-29 2019-04-05 大连楼兰科技股份有限公司 应用在主动驾驶技术中识别交通管制人员的方法
CN105787482A (zh) * 2016-02-26 2016-07-20 华北电力大学 一种基于深度卷积神经网络的特定目标轮廓图像分割方法
CN107392895A (zh) * 2017-07-14 2017-11-24 深圳市唯特视科技有限公司 一种基于卷积循环网络的3d血管结构提取方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1738426A (zh) * 2005-09-09 2006-02-22 南京大学 一种视频运动目标分割与跟踪方法
CN101573959A (zh) * 2006-11-01 2009-11-04 索尼株式会社 运动图片中的片段跟踪
CN101354359A (zh) * 2008-09-04 2009-01-28 湖南大学 药液内运动可见异物的检测跟踪与识别方法
CN105678332A (zh) * 2016-01-08 2016-06-15 昆明理工大学 火焰图像cnn识别建模的转炉炼钢终点判断方法及系统
CN106023220A (zh) * 2016-05-26 2016-10-12 史方 一种基于深度学习的车辆外观部件图像分割方法
CN106874914A (zh) * 2017-01-12 2017-06-20 华南理工大学 一种基于深度卷积神经网络的工业机械臂视觉控制方法

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116075864A (zh) * 2020-08-28 2023-05-05 先讯美资电子有限责任公司 视觉媒体中的人类型的分类
CN112308003A (zh) * 2020-11-06 2021-02-02 中冶赛迪重庆信息技术有限公司 一种废钢货车装载状态识别方法、系统、设备及介质
CN114693921A (zh) * 2020-12-25 2022-07-01 东声(苏州)智能科技有限公司 神经网络模型优化的评估方法、存储介质和计算机
CN113561181A (zh) * 2021-08-04 2021-10-29 北京京东乾石科技有限公司 目标检测模型的更新方法、装置及系统
CN113561181B (zh) * 2021-08-04 2023-01-31 北京京东乾石科技有限公司 目标检测模型的更新方法、装置及系统
CN114241450A (zh) * 2021-10-27 2022-03-25 上海艾豚科技有限公司 识别汽车内饰件上表皮和缝线颜色的方法
CN114264607A (zh) * 2021-12-29 2022-04-01 佛山市帆思科材料技术有限公司 基于机器视觉的瓷砖色差在线检测系统与方法
CN114264607B (zh) * 2021-12-29 2022-06-28 佛山市帆思科材料技术有限公司 基于机器视觉的瓷砖色差在线检测系统与方法
CN114310954A (zh) * 2021-12-31 2022-04-12 北京理工大学 一种护理机器人自适应升降控制方法和系统
CN114310954B (zh) * 2021-12-31 2024-04-16 北京理工大学 一种护理机器人自适应升降控制方法和系统
CN114627288A (zh) * 2022-02-23 2022-06-14 厦门聚视智创科技有限公司 一种复杂3d背景下的背景分割目标识别方法
CN115019300A (zh) * 2022-08-09 2022-09-06 成都运荔枝科技有限公司 用于自动化仓库货物识别的方法
CN121582602A (zh) * 2025-11-28 2026-02-27 北京华软恒信科技发展有限公司 基于计算机视觉的物资盘点机器自动识别计数方法

Also Published As

Publication number Publication date
CN109313710A (zh) 2019-02-05

Similar Documents

Publication Publication Date Title
WO2019148453A1 (fr) Procédé d'apprentissage de modèle de reconnaissance de cible, procédé de reconnaissance de cible, appareil et robot
CN112977974B (zh) 烟包外观质量检测装置及其方法、烟包包装机
CN103679916A (zh) 印码在线检测方法及系统
WO2020253048A1 (fr) Procédé, appareil et dispositif de reconnaissance d'image basés sur un apprentissage profond, et support de stockage
CN109923856A (zh) 补光控制装置、系统、方法以及移动设备
WO2019000118A1 (fr) Procédé de détection visuelle pour processus industriel, dispositif de commande, système et dispositif de stockage
WO2021095991A1 (fr) Dispositif et procédé de génération d'une image de défaut
WO2025211482A1 (fr) Procédé et système de compensation de données satellitaires manquantes
CN108628267A (zh) 一种物方扫描成像系统的分离式、分布式操控方法
CN110766636A (zh) 一种用于ai芯片的智能图像处理方法
CN108333186A (zh) 基于烟垛表面的霉变检测系统及其检测方法
WO2021261905A1 (fr) Appareil et procédé de reconnaissance d'opérations de travail et de mesure de production à base d'analyse d'image
CN114833038A (zh) 一种涂胶路径规划方法及其系统
CN112784796A (zh) 一种自学习的无感人脸识别系统
CN116740199A (zh) 应用于薄膜印刷的颜色智能识别系统
CN110675386B (zh) 一种b族链球菌的检测系统
WO2022108249A1 (fr) Procédé, appareil et programme de génération de données d'apprentissage, et procédé de détection de substances étrangères l'utilisant
CN107421457B (zh) 基于模拟图像采集的非接触式细丝直径测量装置及方法
CN118869899B (zh) 一种基于双目事件相机和深度相机的第一视角数据采集方法及系统
WO2019061128A1 (fr) Procédé de traitement d'image et dispositif électronique associé, et support de stockage
CN109426221A (zh) 一种基于机器视觉技术的港口闸口控制系统
WO2024097777A3 (fr) Systèmes de génération, de stockage et d'utilisation de conception de banques de dents numériques augmentées
CN106846302A (zh) 一种工具正确取件的检测方法及基于此方法的考核台
CN110062217A (zh) 一种三色激光器的对准方法及对准装置
CN114708252B (zh) 设备面板的巡检方法及系统

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18903130

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18903130

Country of ref document: EP

Kind code of ref document: A1