Object recognition method and system based on fusion of infrared image and depth information

Abstract

The application provides an object recognition method based on infrared image and depth information fusion, which can include: collecting infrared image and depth information of an object. A single-stage target detection model is used to determine the object bounding box of the object in the infrared image and the multiple support foot bounding boxes of the object. According to the support foot bounding box and the depth information, the distance between the object and the unmanned carrier vehicle and the angle information of the object to the unmanned carrier vehicle are calculated and uploaded. The application uses a single-stage target detection model, which is beneficial to quickly training and recognizing various types of industrial pallets, and facilitates the deployment and application of the related system of the application. The infrared image has higher robustness in unstable light environments, and the detection result is more stable. In addition, the application recognizes two-dimensional images, which is faster, and avoids the large amount of calculation in the prior art when processing point cloud data of three-dimensional images, thereby increasing the time cost and operation cost.

Description

Technical Field

[0001] This application relates to the field of intelligent warehousing technology, and in particular to an object recognition method and system based on the fusion of infrared images and depth information. Background Technology

[0002] AGV (Automated Guided Vehicle) is a transport vehicle equipped with automatic guidance systems, capable of traveling along a predetermined path, and possessing safety protection and various transfer functions. In the field of smart warehousing technology, AGVs best demonstrate their automation and flexibility, enabling efficient, economical, and flexible unmanned production. However, how to quickly and accurately identify industrial pallets is a challenge that AGVs face when picking up goods.

[0003] Currently, the most common algorithms for recognizing industrial pallets are based on point cloud data processing, such as PointNet, a deep learning framework for point cloud classification / segmentation proposed by Stanford University. Point cloud data processing requires significantly more computational support, necessitating higher computing equipment costs for real-time detection of industrial pallets, inevitably leading to high computational costs. Furthermore, the cost of annotating 3D point clouds is more than ten times that of annotating ordinary images, and the annotation process is more difficult, impacting product iteration time. Additionally, depth cameras are highly susceptible to environmental factors such as ambient light and reflections, resulting in unstable depth information and making it difficult to guarantee recognition results. Summary of the Invention

[0004] This application provides an object recognition method and system based on the fusion of infrared images and depth information, in order to solve or partially solve the above-mentioned problems involved in the background art or at least one other deficiency in the prior art.

[0005] This application provides an object recognition method based on the fusion of infrared images and depth information, which may include: acquiring infrared images and depth information of an object; employing a single-stage object detection model to determine the object bounding box and multiple supporting leg bounding boxes of the object in the infrared image; and calculating and uploading the distance between the object and the automated guided vehicle (AGV), as well as the angle information of the object relative to the AGV, based on the supporting leg bounding boxes and depth information.

[0006] In some implementations, both the infrared image and depth information include all the supporting legs of the object, as well as multiple sockets on the object that are compatible with the automated guided vehicle.

[0007] In some implementations, a single-stage object detection model is used to determine the object bounding box and multiple supporting foot bounding boxes of an object in an infrared image. This may include: receiving an infrared image by the single-stage object detection model and adjusting the size of the infrared image according to the proportion of the object in the image; extracting features from the infrared image using a convolutional neural network of the single-stage object detection model to obtain a prediction result containing multiple object candidate boxes and parameter information for each object candidate box, wherein the parameter information includes the position parameters and confidence scores of the object candidate boxes; sorting the confidence scores of the multiple object candidate boxes, selecting the object candidate box with the highest confidence score, and using the object candidate box as the object bounding box.

[0008] In some implementations, a single-stage object detection model is used to determine the object bounding box and multiple supporting foot bounding boxes of the object in the infrared image. This may further include: receiving the infrared image using the single-stage object detection model and adjusting the size of the infrared image according to the proportion of the supporting feet in the image; extracting features from the infrared image using the convolutional neural network of the single-stage object detection model to obtain a prediction result containing multiple supporting foot candidate boxes and parameter information for each candidate box, wherein the parameter information includes the position parameters and confidence scores of the supporting foot candidate boxes; and selecting supporting foot candidate boxes whose confidence scores meet a preset threshold and using these as the supporting foot bounding boxes.

[0009] In some implementations, after using a single-stage target detection model to determine the object bounding box and multiple supporting leg bounding boxes of the object in the infrared image, the method may further include: using prior information about the object to filter out supporting leg bounding boxes that do not meet preset conditions from the obtained multiple supporting leg bounding boxes, wherein the preset conditions include multiple supporting leg bounding boxes being on the same straight line and supporting leg bounding boxes being located within the object bounding box; and establishing the relationship between the supporting leg bounding boxes and the object bounding boxes.

[0010] This application also provides an object recognition system based on the fusion of infrared images and depth information, which may include: a depth camera, a single-stage object detection model, and a recognition module. Specifically, the depth camera is used to acquire infrared images and depth information of the object. The single-stage object detection model is used to determine the object bounding box and multiple supporting leg bounding boxes of the object in the infrared image, respectively. The recognition module is used to calculate and upload the distance between the object and the automated guided vehicle (AGV), as well as the angle information of the object relative to the AGV, based on the supporting leg bounding boxes and the depth information.

[0011] In some implementations, both the infrared image and depth information include all the supporting legs of the object, as well as multiple sockets on the object that are compatible with the automated guided vehicle.

[0012] In some implementations, the execution steps of a single-stage object detection model may include: receiving an infrared image and adjusting the size of the infrared image according to the proportion of objects in the image; extracting features from the infrared image using a built-in convolutional neural network to obtain a prediction result containing multiple object candidate boxes and parameter information for each object candidate box, wherein the parameter information includes the position parameters and confidence scores of the object candidate boxes; sorting the confidence scores of the multiple object candidate boxes and selecting the object candidate box with the highest confidence score as the object bounding box.

[0013] In some implementations, the execution steps of the single-stage object detection model may further include: receiving an infrared image and adjusting the size of the infrared image according to the proportion of the supporting foot in the infrared image; extracting features from the infrared image using a built-in convolutional neural network to obtain a prediction result containing multiple candidate boxes for supporting feet and parameter information for each candidate box, wherein the parameter information includes the position parameters and confidence scores of the candidate boxes; and selecting candidate boxes whose confidence scores meet a preset threshold and using them as the bounding boxes for the supporting feet.

[0014] In some implementations, a filtering module and an association module may also be included. Specifically, the filtering module is used to filter out support foot bounding boxes that do not meet preset conditions from the multiple support foot bounding boxes obtained, using prior information about the identified object. The preset conditions include multiple support foot bounding boxes being on the same straight line and support foot bounding boxes being located within the object bounding box. The association module is used to establish the relationship between the support foot bounding boxes and the object bounding boxes.

[0015] The technical solution based on the above embodiments can achieve at least one of the following beneficial effects.

[0016] The object recognition method and system based on infrared image and depth information fusion according to an embodiment of this application employs a single-stage target detection model based on infrared images, which facilitates rapid training to recognize various types of industrial pallets and makes the deployment and application of the related system of this application easier. Infrared images have higher robustness in environments with unstable lighting, resulting in more stable detection results. In addition, this application recognizes two-dimensional images, resulting in faster detection speed and avoiding the large amount of computation generated when processing point cloud data of three-dimensional images in the prior art, thus avoiding increased time and computation costs. Attached Figure Description

[0017] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0018] Figure 1This is a flowchart of an object recognition method based on infrared image and depth information fusion according to an exemplary embodiment of this application; and

[0019] Figure 2 This is a schematic diagram of the structure of an object recognition system based on the fusion of infrared images and depth information according to an exemplary embodiment of this application. Detailed Implementation

[0020] To better understand this application, various aspects of this application will be described in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are merely illustrative of exemplary embodiments of this application and are not intended to limit the scope of this application in any way. Throughout the specification, the same reference numerals refer to the same elements. The expression "and / or" includes any and all combinations of one or more of the associated listed items.

[0021] In the accompanying drawings, the size, dimensions, and shapes of the elements have been slightly adjusted for ease of illustration. The drawings are for illustrative purposes only and are not strictly to scale. As used herein, the terms “approximately,” “about,” and similar terms are used to indicate approximation, not degree, and are intended to illustrate inherent deviations in measured or calculated values ​​that will be recognized by one of ordinary skill in the art. Furthermore, the order in which the steps are described in this application does not necessarily indicate the order in which these steps occur in actual operation, unless otherwise expressly defined or deduced from the context.

[0022] It should also be understood that expressions such as "comprising," "including," "having," "containing," and / or "comprising" are open-ended rather than closed-ended expressions in this specification, indicating the presence of the stated features, elements, and / or components, but not excluding the presence of one or more other features, elements, components, and / or combinations thereof. Furthermore, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire list of features, not just individual elements in the list. Additionally, when describing embodiments of this application, the word "may" is used to mean "one or more embodiments of this application." And the term "exemplary" is intended to refer to examples or illustrations.

[0023] Unless otherwise specified, all terms used herein (including engineering and technical terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that, unless expressly stated herein, terms defined in common dictionaries shall be interpreted as having the meaning consistent with their meaning in the context of the relevant art, and not as having an idealized or overly formalized meaning.

[0024] It should be noted that, where there is no conflict, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] Figure 1 This is a flowchart of an object recognition method based on the fusion of infrared images and depth information according to an exemplary embodiment of this application.

[0026] like Figure 1 As shown, this application proposes an object recognition method based on the fusion of infrared images and depth information, which may include: Step S1, acquiring infrared images and depth information of the object; Step S2, using a single-stage object detection model to determine the object bounding box and multiple supporting leg bounding boxes of the object in the infrared image; Step S3, calculating and uploading the distance between the object and the unmanned transport vehicle, and the angle information of the object relative to the unmanned transport vehicle, based on the supporting leg bounding boxes and depth information.

[0027] Step S1 involves acquiring infrared images and depth information of the object. Specifically, capturing conventional images requires a stable environment with suitable lighting conditions; in other words, conventional images are highly dependent on the environment. In environments with uncertain lighting, such as warehouses, it is obviously difficult to acquire clear and easily identifiable conventional images. Therefore, this application utilizes infrared imaging technology to acquire IR (Infrared Radiation) images, which are obtained by measuring the heat radiated outwards by an object, thereby acquiring the shape and texture characteristics of the object. When acquiring IR images using this method, it is not affected by environmental factors such as lighting conditions, can operate stably continuously, and has a certain degree of robustness. Furthermore, to facilitate the handling of the object by an AGV (Automated Guided Vehicle), it is also necessary to acquire the object's depth information.

[0028] Based on the above, this application employs a depth camera to acquire IR images and depth information of an object; more specifically, a combined light depth camera can be used. The combined light depth camera is equipped with a near-infrared laser, which projects light rays with certain structural characteristics onto the object being photographed, and then the images are captured by a dedicated infrared camera. This structured light will acquire different image phase information depending on the different depth regions of the object being photographed. Then, the built-in processing unit converts this different image phase information into depth information, thereby obtaining the object's position and depth information. In other words, the depth information includes the depth-of-field distance of the shooting space; that is, using precise data, the actual distance from the lens to each point in the shooting space is obtained.

[0029] It should be noted that this application can be applied to the scenario of AGV recognizing industrial pallets, meaning the object in this application can be considered an industrial pallet. Therefore, in this application, the depth camera is positioned slightly above the middle of the AGV's forks, tilted slightly towards the ground, to record in real-time the frontal IR image of the industrial pallet in the AGV's fork-picking direction. This also helps to clearly capture the multiple support legs of the industrial pallet and the sockets that the industrial pallet uses to mate with the AGV. The shape of the sockets can be rectangular and is not limited here. Based on this application scenario, for ease of understanding, the industrial pallet will be used as a substitute for the object in the following description.

[0030] It should be noted that the IR images and depth information acquired in this application are all acquired and recorded in real time by a depth camera, ensuring real-time performance and providing accurate and effective input data for subsequent identification steps.

[0031] Step S2: Using a single-stage target detection model, the object bounding box and multiple supporting foot bounding boxes of the object in the infrared image are determined respectively.

[0032] In some implementations, after obtaining the IR image of the industrial pallet, a single-stage target detection model is used to identify the specific location of the industrial pallet in the IR image, as well as the specific locations of the multiple support legs of the industrial pallet in the IR image, providing a basis for subsequent identification of the distance and relative angle information between the industrial pallet and the AGV. It should be noted that in this application, the IR image and depth information are transmitted to the single-stage target detection model through the AGV's built-in Robot Operating System (ROS).

[0033] In some implementations, single-stage object detection models detect targets by performing regular and dense sampling of the position, scale, and aspect ratio of objects in IR images, which has the advantage of high computational efficiency. Therefore, this application uses a single-stage object detection model to determine the bounding boxes of the industrial pallet and its support legs. More specifically, an IR image containing the industrial pallet is uploaded to the single-stage object detection model. The single-stage object detection model adjusts the size of the IR image according to the proportion of the industrial pallet in the IR image to ensure that the IR image has the most suitable clarity and size, facilitating feature extraction.

[0034] Furthermore, the convolutional neural network of the single-stage object detection model extracts features from the infrared image to obtain prediction results containing multiple industrial pallet candidate boxes and parameter information for each candidate box. The parameter information includes the positional parameters and confidence scores of the industrial pallet candidate boxes. Specifically, the single-stage object detection model has a convolutional neural network whose convolutional layers analyze and process the input IR image to extract features and output multiple industrial pallet candidate boxes, each with corresponding parameter information. This parameter information includes its specific positional coordinates, which can be represented in a Cartesian coordinate system, i.e., the location information of the industrial pallet; it also includes the width, height, and confidence score of the industrial pallet. The confidence score is the probability that the candidate box contains an industrial pallet; the higher the probability, the higher the confidence score.

[0035] Furthermore, a pre-set confidence threshold is used to filter industrial pallet candidate frames from multiple candidate frames that meet the threshold requirements. Then, the confidence scores of the selected industrial pallet candidate frames are ranked, and the industrial pallet candidate frame with the highest confidence score is selected as the bounding box of the industrial pallet. It should be noted that this application does not limit the shape of the industrial pallet; the specific shape of the industrial pallet can be determined using the above method.

[0036] Similarly, the bounding boxes of each support leg in the infrared image can be obtained according to the above method. Specifically, this includes: receiving the infrared image using a single-stage object detection model and adjusting the size of the infrared image according to the proportion of the support legs in the image; extracting features from the infrared image using the convolutional neural network of the single-stage object detection model to obtain prediction results containing multiple candidate bounding boxes for each support leg, along with parameter information for each candidate bounding box, including the positional parameters and confidence scores of the candidate bounding boxes; and selecting the candidate bounding boxes whose confidence scores meet a preset threshold and using them as the support leg bounding boxes. The principle for obtaining the support leg bounding boxes is the same as that for obtaining the industrial pallet bounding boxes, and can be referred to above; it will not be repeated here.

[0037] Of course, before using the convolutional neural network of this single-stage object detection model, it needs to be trained. This involves inputting a large number of industrial pallet bounding box samples and support leg bounding box samples into the single-stage object detection model. After analysis and calculation by the convolutional layers of the convolutional neural network, an initial industrial pallet bounding box or initial support leg bounding box is output and compared with the corresponding expected output result. If the comparison result shows an error, the weight parameters of each neuron in the convolutional layer are adjusted until the output result is within the acceptable range of the expected output result. At this point, the training of the convolutional neural network is complete.

[0038] In some implementations, the method further includes utilizing prior information from historically identified industrial pallets to filter out support foot boundary frames that do not meet preset conditions from among the obtained multiple support foot boundary frames. These preset conditions include multiple support foot boundary frames being on the same straight line and the support foot boundary frames being located within the industrial pallet boundary frame. This method ensures the accuracy of the determined support foot boundary frames.

[0039] In some implementations, the relationship between the support foot boundary frame and the industrial pallet boundary frame can also be established, providing a basis for subsequent steps to determine the angle of the industrial pallet relative to the AGV through the support foot boundary frame and its parameter information.

[0040] Step S3: Based on the support leg bounding box and depth information, calculate and upload the distance between the object and the AGV, as well as the angle information of the object relative to the AGV. Specifically, after obtaining the support leg bounding box, based on the positional parameter information corresponding to the support leg bounding box, such as position coordinates, and combined with the aforementioned collected depth information—that is, the actual distance from each point in the industrial pallet to the camera, especially the actual distance from each point on the support leg to the camera—the distance from the AGV to the industrial pallet is obtained. Furthermore, considering the relationship between the support leg and the industrial pallet, the angle information of the industrial pallet relative to the AGV can be determined based on the position coordinates of the support leg.

[0041] In some implementations, after determining the angle and position information, the AGV can be driven forward an appropriate distance to pick up the goods.

[0042] The object recognition method based on infrared image and depth information fusion according to an embodiment of this application employs a single-stage target detection model based on infrared images, which facilitates rapid training to recognize various types of industrial pallets and makes the deployment and application of the related system of this application easier. Infrared images have higher robustness in environments with unstable lighting, resulting in more stable detection results. In addition, this application recognizes two-dimensional images, resulting in faster detection speed and avoiding the large amount of computation generated when processing point cloud data of three-dimensional images in the prior art, thus avoiding increased time and computation costs.

[0043] This application also proposes an object recognition system based on the fusion of infrared images and depth information, which may include: a depth camera 1, a single-stage object detection model 2, and a recognition module 3. Specifically, the depth camera 1 is used to acquire infrared images and depth information of the object. The single-stage object detection model 2 is used to determine the object bounding box and multiple supporting leg bounding boxes of the object in the infrared image, respectively. The recognition module 3 is used to calculate and upload the distance between the object and the automated guided vehicle (AGV), as well as the angle information of the object relative to the AGV, based on the supporting leg bounding boxes and the depth information.

[0044] In some implementations, both the infrared image and depth information include all the supporting legs of the object, as well as multiple sockets on the object that are compatible with the automated guided vehicle.

[0045] In some implementations, the execution steps of the single-stage object detection model 2 may include: receiving an infrared image and adjusting the size of the infrared image according to the proportion of objects in the infrared image; extracting features from the infrared image using a built-in convolutional neural network to obtain a prediction result containing multiple object candidate boxes and parameter information for each object candidate box, wherein the parameter information includes the position parameters and confidence scores of the object candidate boxes; sorting the confidence scores of the multiple object candidate boxes and selecting the object candidate box with the highest confidence score as the object bounding box.

[0046] In some implementations, the execution steps of the single-stage object detection model 2 may further include: receiving an infrared image by the single-stage object detection model and adjusting the size of the infrared image according to the proportion of the supporting foot in the infrared image; extracting features from the infrared image using a built-in convolutional neural network to obtain a prediction result containing multiple candidate boxes for supporting feet and parameter information for each candidate box, wherein the parameter information includes the position parameters and confidence scores of the candidate boxes for supporting feet; and selecting the candidate boxes for supporting feet whose confidence scores meet a preset threshold and using them as the bounding boxes for supporting feet.

[0047] In some embodiments, a filtering module 4 and an association module 5 may also be included. Specifically, the filtering module 1 is used to filter out support foot bounding boxes that do not meet preset conditions from the multiple support foot bounding boxes obtained, using prior information about the identified object. The preset conditions include multiple support foot bounding boxes being on the same straight line and support foot bounding boxes being located within the object bounding box. The association module 5 is used to establish the relationship between the support foot bounding boxes and the object bounding boxes.

[0048] This system was proposed to implement the above method. Therefore, the specific application scenarios and principles of each module are completely consistent with the parts involved in the above method. Please refer to the above for details, and they will not be repeated here.

[0049] The object recognition system based on infrared image and depth information fusion according to an embodiment of this application employs a single-stage target detection model based on infrared images, which facilitates rapid training to recognize various types of industrial pallets and makes the deployment and application of the related system of this application easier. Infrared images have higher robustness in environments with unstable lighting, resulting in more stable detection results. In addition, this application recognizes two-dimensional images, resulting in faster detection speed and avoiding the large amount of computation generated when processing point cloud data of three-dimensional images in the prior art, thus avoiding increased time and computation costs.

[0050] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An object recognition method based on the fusion of infrared images and depth information, characterized in that, include: Acquire infrared images and depth information of objects; A single-stage target detection model is used to determine the object bounding box of the object in the infrared image and the bounding boxes of multiple supporting feet of the object. as well as Based on the support foot boundary frame and the depth information, calculate and upload the distance between the object and the unmanned transport vehicle, as well as the angle information of the object relative to the unmanned transport vehicle.

2. The method according to claim 1, characterized in that, Both the infrared image and the depth information include all the supporting legs of the object, as well as multiple sockets on the object that are compatible with the unmanned transport vehicle.

3. The method according to claim 1, wherein a single-stage target detection model is used to determine the object bounding box of the object in the infrared image and the bounding boxes of multiple supporting legs of the object, characterized in that, include: The single-stage target detection model receives the infrared image and adjusts the size of the infrared image according to the proportion of objects in the infrared image. The convolutional neural network of the single-stage target detection model extracts features from the infrared image to obtain a prediction result containing multiple object candidate boxes and parameter information of each object candidate box, wherein the parameter information includes the position parameters of the object candidate boxes and their confidence scores. The confidence scores of multiple candidate bounding boxes are sorted, and the candidate bounding box with the highest confidence score is selected as the bounding box of the object.

4. The method according to claim 1 or 3, wherein a single-stage target detection model is used to determine the object bounding box of the object in the infrared image and the bounding boxes of multiple supporting legs of the object, characterized in that, Also includes: The single-stage target detection model receives the infrared image and adjusts the size of the infrared image according to the proportion of the supporting foot in the infrared image; The convolutional neural network of the single-stage target detection model extracts features from the infrared image to obtain a prediction result containing multiple candidate boxes for supporting feet and parameter information of each candidate box for supporting feet, wherein the parameter information includes the position parameters of the candidate boxes for supporting feet and their confidence levels. The candidate support bounding boxes that meet the confidence level of the candidate support bounding boxes according to the preset threshold are selected and used as the support bounding boxes.

5. The method according to claim 4, wherein after determining the object bounding box of the object in the infrared image and the bounding boxes of the object's multiple supporting legs using a single-stage target detection model, it is characterized in that, Also includes: Using prior information about the object, among the obtained multiple support foot bounding boxes, those that do not meet preset conditions are filtered out. These preset conditions include the multiple support foot bounding boxes being on the same straight line and the support foot bounding boxes being located within the object's bounding box. Establish the relationship between the bounding box of the support foot and the bounding box of the object.

6. An object recognition system based on the fusion of infrared images and depth information, characterized in that, include: Depth cameras are used to capture infrared images and depth information of objects. A single-stage target detection model is used to determine the object bounding box of an object in the infrared image and the bounding boxes of multiple supporting feet of the object, respectively. as well as The identification module is used to calculate and upload the distance between the object and the unmanned transport vehicle, as well as the angle information of the object relative to the unmanned transport vehicle, based on the bounding box of the support foot and the depth information.

7. The system according to claim 6, characterized in that, Both the infrared image and the depth information include all the supporting legs of the object, as well as multiple sockets on the object that are compatible with the unmanned transport vehicle.

8. The system according to claim 6, characterized in that, The execution steps of the single-stage target detection model include: Receive the infrared image and adjust the size of the infrared image according to the proportion of the objects in the infrared image; The infrared image is subjected to feature extraction by a built-in convolutional neural network to obtain a prediction result containing multiple object candidate boxes and parameter information of each object candidate box, wherein the parameter information includes the position parameters of the object candidate boxes and their confidence scores. The confidence scores of multiple candidate bounding boxes are sorted, and the candidate bounding box with the highest confidence score is selected as the bounding box of the object.

9. The system according to claim 6 or 8, characterized in that, The execution steps of the single-stage target detection model also include: The single-stage target detection model receives the infrared image and adjusts the size of the infrared image according to the proportion of the supporting foot in the infrared image; The infrared image is subjected to feature extraction by the built-in convolutional neural network to obtain a prediction result containing multiple candidate boxes for supporting feet and parameter information of each candidate box for supporting feet, wherein the parameter information includes the position parameters of the candidate boxes for supporting feet and their confidence scores. The candidate support bounding boxes that meet the confidence level of the candidate support bounding boxes according to the preset threshold are selected and used as the support bounding boxes.

10. The system according to claim 6, characterized in that, Also includes: A filtering module is used to filter out support foot bounding boxes that do not meet preset conditions from a plurality of obtained support foot bounding boxes using prior information for recognizing the object. The preset conditions include that the plurality of support foot bounding boxes are on the same straight line and that the support foot bounding boxes are located within the object bounding box. as well as The association module is used to establish the relationship between the support foot bounding box and the object bounding box.

Images