Self-adaptive moving method and self-adaptive moving accompanying robot

Abstract

The invention discloses a self-adaptive moving method and a self-adaptive moving accompanying robot. Comprising a camera module used for collecting a color depth image of a surrounding environment; the intelligent base is used for acquiring a color depth image, inputting the color depth image into a preset simultaneous localization and mapping algorithm, outputting an identified bounding box and category, and creating an environment map; the identification module is also used for identifying a to-be-tracked moving target based on the bounding box and the category; planning a moving path of the robot based on the position of the moving target and an environment map; the universal wheel is used for tracking a target based on the planned moving path; the pain recognition module is used for evaluating the pain level of the user. According to the accompanying robot provided by the invention, the semantic technology in the SLAM system is invented, so that the environmental perception precision of the robot is improved, boundary detection is optimized, and the perception processing capability of the robot on dynamic objects in the environment is enhanced. The environment map can be constructed more accurately, and the safety and reliability of path planning are improved.

Description

Technical Field

[0001] The present application relates to the field of robotics technology, and in particular to an adaptive movement method and an adaptive movement delivery robot. Background Art

[0002] When multiple women are waiting to give birth at the same time or at night, medical staff often face the challenge of insufficient staff. As the frequency of uterine contractions increases, the pain of women gradually intensifies. Given the rigor of medical work, medical staff need to regularly monitor the pain level of women in labor in order to accurately adjust the dosage of analgesics. Therefore, introducing an efficient and intelligent pain management system embedded in a mobile robot carrier to accompany women throughout the process in a humane way can not only optimize the process of women waiting to give birth, but also is a key solution to improve the situation of "women waiting to give birth alone and their pain being ignored."

[0003] At present, the SLAM (Simultaneous Localization and Mapping) system of the delivery robot is becoming more and more important. The SLAM system enables the robot to autonomously navigate and build maps in unknown environments. However, the accuracy of the current robot's environmental perception and boundary detection needs to be improved. Summary of the invention

[0004] The embodiments of the present application provide an adaptive movement method and an adaptive movement delivery robot.

[0005] According to an embodiment of the present application, an adaptive mobile delivery robot is provided, comprising:

[0006] A camera module, used to collect color depth images of the surrounding environment;

[0007] A smart base, used to obtain the color depth image, input the color depth image into a preset simultaneous positioning and mapping algorithm, output the identified bounding box and category, and create an environmental map; further used to identify the mobile target to be tracked based on the bounding box and category; and plan the movement path of the robot based on the position of the mobile target and the environmental map;

[0008] Universal wheels, used to track the target based on the planned moving path;

[0009] Pain recognition module, used to assess the user's pain level.

[0010] In one embodiment, the color depth image is input into a preset simultaneous localization and mapping algorithm, and the identified bounding box and category are output, including:

[0011] Constructing a three-branch semantic segmentation network module in the simultaneous localization and mapping algorithm; the three-branch semantic segmentation network module includes a boundary branch, a context branch, a detail branch, a fusion layer and a fully connected layer;

[0012] The color depth image is input into the three-branch semantic segmentation network module to obtain the recognized bounding box and category.

[0013] In one embodiment, the boundary branch is composed of multiple convolutional layers, difference layers, attention mechanisms, and pooling layers, and is used to output the extracted boundary features;

[0014] The context branch is composed of multiple convolutional layers, long short-term memory networks and pooling layers, and is used to output the extracted context information;

[0015] The detail branch consists of multiple convolutional layers, attention mechanisms, and pooling layers, and is used to output the extracted detail information.

[0016] In one embodiment, it further includes:

[0017] Compute network output loss, weighted binary cross entropy loss, standard cross entropy loss, and boundary-aware cross loss;

[0018] The network output loss, weighted binary cross entropy loss, standard cross entropy loss and boundary-aware cross loss are weighted summed to obtain the total loss of the three-branch semantic segmentation network module.

[0019] In one implementation, calculating the boundary-aware intersection loss includes:

[0020] Adopting adaptive threshold graph learning algorithm to predict adaptive threshold;

[0021] Based on the adaptive threshold and differential binarization method, a predicted binarized boundary map is obtained;

[0022] Calculating the difference between the predicted binary boundary map and the true binary boundary map;

[0023] The boundary-aware intersection loss is obtained based on the sum of weighted differences at all pixel positions.

[0024] In one embodiment, identifying a moving target to be tracked based on the bounding box and the category includes:

[0025] Calculate the 3D world coordinate centroid of each object based on the center of the bounding box and the corresponding depth pixels;

[0026] Calculate the object's moving speed through the change in the center of mass position and the time interval between two consecutive frames;

[0027] Based on the object's moving speed and a preset threshold, the moving target to be tracked is identified.

[0028] In one embodiment, the pain recognition module is used to obtain the maternal facial expression data and the conductivity information collected by the skin electrical sensor;

[0029] The maternal expression data and the conductivity information collected by the skin electrical sensor are input into a preset fusion decision algorithm to obtain the maternal pain detection level.

[0030] In one embodiment, it further includes:

[0031] A touch display screen is used to display the pain detection level in the form of a time series curve on the touch display screen of the robot.

[0032] In one embodiment, it further includes:

[0033] The alarm module is used to issue an alarm message when the detected pain level remains at the highest level and continues for a preset period of time.

[0034] According to an embodiment of the present application, an adaptive movement method of a delivery robot is provided, comprising:

[0035] Collect color depth images of the surrounding environment;

[0036] Acquire the color depth image, input the color depth image into a preset simultaneous localization and mapping algorithm, output the identified bounding box and category, and create an environment map;

[0037] Based on the bounding box and the category, identifying a moving target to be tracked;

[0038] Planning a moving path of the robot based on the position of the moving target and the environment map;

[0039] The robot is controlled to track the target based on the planned moving path.

[0040] The technical solution provided by the embodiments of the present application may have the following beneficial effects:

[0041] This application provides a delivery robot that helps monitor the pain of mothers, saving the time cost of medical staff in pain assessment, especially at night. In addition, the application of this system not only optimizes the delivery system, but also makes the delivery room intelligent, improving the quality of medical services. Robot-assisted doctors can track the pain of mothers in real time, thereby significantly improving medical efficiency and the psychological and physiological comfort of mothers. The application of this innovative technology ensures that the pain of mothers during delivery is effectively managed and intervened in time, greatly improving the delivery experience of mothers.

[0042] This application adopts an improved SLAM system to improve the accuracy of the robot's environmental perception and optimize boundary detection, thereby enhancing its ability to perceive and process dynamic objects in the environment. In addition, the SLAM system can more accurately construct an environmental map to improve the safety and reliability of path planning. The improvement of the obstacle avoidance function means an improvement in the safety of personnel in the delivery room. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0044] Figure 1 is a schematic diagram of an adaptive movement method according to an embodiment of the present application;

[0045] Figure 2 is a schematic diagram of a delivery companion robot according to an embodiment of the present application;

[0046] Figure 3 is a boundary branch network diagram according to an embodiment of the present application;

[0047] Figure 4 is a context branch network diagram according to an embodiment of the present application;

[0048] Figure 5 It is a detailed branch network diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0050] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0051] The embodiment of the present application provides an adaptive mobile delivery companion robot, which can not only accurately assess the user's pain level, but also recognize the parturient and autonomously follow the parturient, bringing a good experience to the parturient.

[0052] like Figure 2 As shown, the delivery accompanying robot provided in the embodiment of the present application includes: a camera module for collecting color depth images of the surrounding environment.

[0053] In one embodiment, the robot includes a universal binocular RGB-D camera. It is set on the head of the robot. It combines the ability to capture RGB (red, green, and blue) color images and depth information, and is widely used in robot navigation, space scanning, object recognition and other fields. This camera uses binocular stereo vision technology, uses two cameras to capture images from different angles, and then calculates the depth information of each pixel through an algorithm, thereby realizing the perception of three-dimensional space.

[0054] It also includes an intelligent base set at the bottom, which is used to obtain color depth images, input the color depth images into a preset simultaneous positioning and mapping algorithm, output the recognized bounding box and category, and create an environmental map. It is also used to identify the mobile target to be tracked based on the bounding box and category; based on the position of the mobile target and the environmental map, plan the robot's moving path.

[0055] The mobility of the birthing robot is mainly achieved by its intelligent base, which integrates various sensors and processing units for sensing the environment, positioning and navigation. These include visual SLAM (Simultaneous Localization and Mapping) technology and path planning technology. To achieve the autonomous mobility function of the pain monitoring robot, the robot is required to have access to a consistent environmental model to achieve safe planning and decision-making. Prior to this, the robot needs to have the ability to create an environmental map, self-localize in the environment, and control motion.

[0056] SLAM is defined here as the process by which a delivery robot carrying multiple sensors determines its own position in the surrounding environment and creates a delivery room map. Path planning is to create the best navigation path for the robot to reach the target location based on the delivery room environment and the location of each bed. The SLAM system in the delivery room scene should be based on a dynamic environment, such as moving medical staff.

[0057] Among them, the role of the segmentation module in the SLAM system is mainly to distinguish different objects, identify static and dynamic objects / positions in the environment, and thus generate a more accurate map. This application optimizes the segmentation module to further improve its performance.

[0058] Specifically, a three-branch semantic segmentation network module in the simultaneous localization and mapping algorithm is constructed, which includes a boundary branch, a context branch, a detail branch, a fusion layer, and a fully connected layer. The color depth image is input into the three-branch semantic segmentation network module to obtain the recognized bounding box and category.

[0059] This application adopts a three-branch network architecture for semantic segmentation tasks. It includes a boundary branch, which focuses on the area with large changes (boundaries) in the image and extracts high-frequency features for boundary detection. It achieves accurate positioning of boundaries through multiple small convolution kernels and differential operations. Figure 3 As shown in the figure, the boundary branch is composed of multiple convolutional layers, differential layers, attention mechanisms, and pooling layers. Specifically, it is composed of sequentially connected convolutional layer 1, differential layer, convolutional layer 2, attention mechanism, and pooling layer to output boundary features.

[0060] It also includes a context branch, which is mainly responsible for accumulating global and contextual information in the image and capturing long-distance dependencies. It extracts global semantic information through larger convolution kernels and deeper network layers. Figure 4 As shown, it is composed of multiple convolutional layers, long short-term memory networks, and pooling layers, which are used to output the extracted context information. Specifically, it is composed of sequentially connected convolutional layer 1, accumulation layer (long short-term memory network), convolutional layer 2, and pooling layer, which outputs the context information.

[0061] It also includes a detail branch, which is responsible for extracting high-resolution detail information, mainly used to capture subtle features in the image. It maintains the integrity of the details through multi-layer convolution and small pooling windows. Figure 5 As shown, it is composed of multiple convolutional layers, attention mechanisms and pooling layers to output the extracted detail information. Specifically, it is composed of sequentially connected convolutional layer 1, convolutional layer 2, attention module, and pooling layer to output the detail information.

[0062] The three-branch network also includes a fusion layer, which is used to integrate the output features of the three branches: boundary features, context features, and detail features through a weighted fusion strategy. The fused features are then passed through a fully connected layer for the final classification or regression task.

[0063] Furthermore, the final loss function of the three-branch semantic segmentation model of the present application consists of four parts. It includes calculating the network output loss, weighted binary cross entropy loss, standard cross entropy loss and boundary-aware cross loss. The network output loss, weighted binary cross entropy loss, standard cross entropy loss and boundary-aware cross loss are weighted and summed to obtain the total loss of the three-branch semantic segmentation network module.

[0064] Among them, the network output loss l1 is the loss applied to the output of the earlier layers in the network, which is responsible for optimizing the extraction of detail information in the earlier layers. The pixel-level mean square error (MSE) is used to measure the difference between the prediction and the true value:

[0065]

[0066] Where N is the total number of pixels, D i is the i-th pixel value predicted by the detail branch, G i is the true i-th pixel value.

[0067] l2 is the weighted binary cross entropy loss (boundary detection), which is used to deal with the imbalance problem of boundary detection, that is, giving higher weights to boundary pixels and emphasizing the importance of boundary areas.

[0068]

[0069] Where N is the total number of pixels, y i is the true label (boundary = 1, non-boundary = 0), p i The probability that the i-th pixel is a boundary is predicted by the boundary branch. i is the weight of the i-th pixel:

[0070]

[0071] Among them, w boundary >w non-boundary ;

[0072] l3 is the standard cross entropy loss, which measures the difference between the true and predicted values.

[0073] The most important is the l4 boundary-aware intersection loss, which determines whether the loss of each pixel should be emphasized by comparing the boundary prediction of each pixel with the preset threshold t. If the boundary prediction value of the current pixel is greater than the preset value t, it means that this pixel is likely to be on the boundary of an item, so its loss is calculated.

[0074] In one embodiment, calculating the boundary-aware cross loss includes: predicting an adaptive threshold using an adaptive threshold map learning algorithm. In order to improve the real-time performance and accuracy of segmentation, the preset threshold t here is designed to adopt an adaptive threshold map learning mechanism. This mechanism is actually a one-branch network specifically used to predict an adaptive threshold map. This means that the network not only learns the probability of identifying the boundary area (probability map), but also learns to generate an optimal threshold for each pixel, which determines whether the corresponding pixel belongs to the edge area.

[0075] Furthermore, based on the adaptive threshold and differential binarization method, the predicted binary boundary map is obtained. The model is trained using the input image and the corresponding true label (the annotation of the boundary area). The adaptive threshold mechanism B(x, y) can be implemented by differential binarization, which can be implemented with the help of the sigmoid function σ, and its output can be approximately represented as 0 or 1:

[0076]

[0077] Here I(x,y) is the pixel value of the input image. For each pixel, this branch network predicts a threshold T(x,y) and an edge probability P(x,y). The ∈ parameter is used to control the slope of the sigmoid function, and B(x,y) represents the predicted binary boundary map.

[0078] Furthermore, the difference between the predicted binary boundary map and the true binary boundary map is calculated, and the boundary-aware intersection loss l4 is obtained based on the sum of the weighted differences at all pixel positions.

[0079] Specifically, l4=∑ (x,y) w(x,y)·|B(x,y)-G(x,y)|; where w(x,y) represents the weight of the edge region, G(x,y) is the true binary boundary map, and B(x,y) represents the predicted binary boundary map. The boundary-aware intersection loss is obtained according to the sum of the weighted differences of all pixel positions.

[0080] This application adds an adaptive threshold map generation step in the SLAM segmentation module, which enables the threshold to be dynamically adjusted in different scenarios, thereby improving the accuracy and robustness of segmentation. This adaptive mechanism allows the model to automatically adjust the threshold according to the characteristics of the local image when processing different input images to obtain the best segmentation effect. The segmentation module takes an RGB image and outputs a bounding box, category and confidence score, and a binary mask for each instance.

[0081] In one embodiment, based on the bounding box and the category, the moving target to be tracked is identified, including: calculating the 3D world coordinate center of mass of each object based on the center of the bounding box and the corresponding depth pixel; calculating the movement speed of the object through the change in the center of mass position and the time interval between two consecutive frames; and identifying the moving target to be tracked based on the movement speed of the object and a preset threshold.

[0082] Specifically, the robot also includes a tracking decision module that receives the mask and bounding box coordinates from the segmentation module. The 3D world coordinate center of mass of each dynamic object is calculated using the center of the bounding box and the corresponding depth pixels. Assume that the center of mass coordinates of the object at time t is C t =(x t ,y t ,z t ), by calculating the change in the center of mass position between two consecutive frames, the speed of the object can be estimated. The speed of the object at time t Among them C t-1 is the center of mass coordinate of the dynamic object in the previous frame (x t-1 ,y t-1 ,z t-1 ), Δt is the time interval between two frames. According to the speed change of the dynamic object in the continuous video frames and the difference in the area occupied by the entire screen, a threshold is set to determine whether the object is dynamic or static. The threshold here can also adopt the above-mentioned adaptive threshold mechanism. According to this algorithm, the robot can identify the moving target to be tracked and follow the target.

[0083] Furthermore, based on the target position and the environment map, the robot's moving path is planned, and a universal wheel is also included for tracking the target based on the planned moving path. It can ensure that it can smoothly avoid obstacles and smoothly follow the target movement. The specific path planning algorithm is not limited in the embodiments of this application, and reference can be made to the existing path planning algorithm.

[0084] In addition, infrared sensors detect human body heat to assist the robot in locating the mother. Computer vision algorithms are initially used to detect the front of the patient's face. The self-moving and flexible positioning function of this delivery robot equipped with an RGB-D camera, coupled with a front face detection system, greatly improves the problem of the existing pain assessment system that requires additional fixed cameras to aim at the patient's face at multiple angles to assess their pain level, alleviating the mother's fear of camera collection.

[0085] It also includes a pain recognition module for evaluating the user's pain level. It obtains the maternal expression data and the conductivity information collected by the skin electrical sensor; it inputs the maternal expression data and the conductivity information collected by the skin electrical sensor into a preset fusion decision algorithm to obtain the maternal pain detection level.

[0086] In one embodiment, the robot uses a high-resolution camera to capture the facial expressions of the mother. These expression data are then input into a pre-trained facial expression recognition system that can recognize and quantify the mother's pain expression. Facial expression recognition technology has developed rapidly in recent years, and can infer an individual's emotional state, such as anger, disgust, fear, happiness, neutrality, sadness, and surprise, by analyzing changes in key points on the face. In addition to facial expressions, the robot is also equipped with a skin electrode sensor to collect the conductivity information of the mother. The skin electrode signal (EDA) is a physiological indicator that reflects the emotional state, which can capture subtle changes related to emotional arousal. The analysis of EDA signals usually includes time domain features, frequency domain features, and time-frequency features such as wavelet transform. By fusing these multimodal data, the accompanying robot comprehensively analyzes the pain state of the mother through a preset fusion decision algorithm. Finally, the accompanying robot will determine the pain detection level of the mother based on the output of the fusion decision algorithm, and may automatically adjust drug delivery or take other pain management measures to ensure the comfort and safety of the mother. The specific fusion decision algorithm is not specifically limited in the present embodiment, and such an algorithm may include a machine learning model or even a more advanced deep learning network, such as a convolutional neural network (CNN) and a recurrent neural network (RNN). These models can process and analyze multi-source data to provide a more accurate pain assessment.

[0087] The advantage of this multimodal pain assessment method is that it can provide a more comprehensive and objective pain assessment, reducing the reliance on maternal self-reported pain levels, especially when it is difficult for maternal pain to clearly express their pain status. In addition, this method can also monitor pain changes in real time, providing timely feedback to medical staff, thereby optimizing pain management strategies.

[0088] The delivery robot transmits data directly to the cloud server via Wi-Fi within the local area network. This design enables the robot to connect to the hospital's internal information system to achieve efficient data exchange. This automated and direct data transmission method not only improves the efficiency of pain management, but also reduces the errors that may occur due to manual data entry, while also reducing the workload of medical staff.

[0089] In one embodiment, it also includes: a touch screen display for displaying the pain detection level in the form of a time series curve on the touch screen display of the robot. The pain detection results are displayed on the touch screen of the robot in the form of a time series curve. The x-axis of the curve represents time (one-minute intervals), and the y-axis is the pain level data of the parturient in the corresponding time dimension. The monitoring curve will be provided in a user-friendly manner for medical staff (and parturients) to view at any time and manually touch to select the pain level of the parturient at any time point, and medical staff can also choose to view the clinically significant one-minute or five-minute pain level curve change graph of the parturient.

[0090] In one embodiment, it further includes: an alarm module, which is used to issue an alarm message when the detected pain level remains at the highest level and continues for a preset period of time.

[0091] When the pain level detected by the robot remains at the highest level for a certain period of time, the machine will trigger an alarm sound, and the medical staff will adjust the subsequent dosage of analgesics according to the mother's condition. This simple notification method can quickly and agilely relieve the mother's pain. Here, all the dosage of analgesics used by the mother during the entire delivery process will be input by the medical staff on the touch screen and recorded in the robot's back-end database for easy management.

[0092] This application adds semantic technology to the SLAM system, improves the accuracy of the robot's environmental perception, optimizes boundary detection, and thus enhances its perception and processing capabilities of dynamic objects in the environment. In addition, the SLAM system can more accurately construct an environmental map to improve the safety and reliability of path planning. The improvement of obstacle avoidance function is the improvement of the safety of personnel in the delivery room. This application saves the time cost of medical staff in pain assessment, especially at night, by monitoring the pain of parturients with the assistance of robots, and realizes a scientific quantitative solution for the pain level of parturients during contractions. In addition, the application of this system not only optimizes the delivery system, but also makes the delivery room intelligent and improves the quality of medical services. Robot-assisted doctors can track the pain of parturients in real time and manage the dosage of analgesics, and take corresponding measures in time to relieve the pain of parturients, thereby significantly improving medical efficiency and the psychological and physiological comfort of parturients. It ensures that the pain of parturients during delivery is effectively managed and intervened in time, greatly improving the delivery experience of parturients.

[0093] According to another aspect of the embodiment of the present application, an adaptive movement method of a delivery robot is also provided. Figure 1 As shown, it mainly includes the following steps:

[0094] S101 collects color depth images of the surrounding environment;

[0095] S102 acquires a color depth image, inputs the color depth image into a preset simultaneous positioning and mapping algorithm, outputs a recognized bounding box and category, and creates an environment map;

[0096] S103 identifies a moving target to be tracked based on the bounding box and the category;

[0097] S104 plans the movement path of the robot based on the position of the moving target and the environment map;

[0098] S105 controls the robot to track the target based on the planned moving path.

[0099] It should be noted that when the birth companion robot provided in the above embodiment performs the adaptive movement method, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the adaptive movement method provided in the above embodiment and the adaptive mobile robot embodiment belong to the same concept. The embodiment and implementation process thereof are detailed in the adaptive mobile robot embodiment, which will not be repeated here.

[0100] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the attached claims.

Claims

1. An adaptive mobile birth companion robot, characterized in that: include: A camera module, used to collect color depth images of the surrounding environment; A smart base, used to obtain the color depth image, input the color depth image into a preset simultaneous positioning and mapping algorithm, output the identified bounding box and category, and create an environmental map; further used to identify the mobile target to be tracked based on the bounding box and category; and plan the movement path of the robot based on the position of the mobile target and the environmental map; Universal wheels, used to track the target based on the planned moving path; Pain recognition module, used to assess the user's pain level.

2. The delivery robot according to claim 1, characterized in that: The color depth image is input into a preset simultaneous localization and mapping algorithm, and the identified bounding boxes and categories are output, including: Constructing a three-branch semantic segmentation network module in the simultaneous localization and mapping algorithm; the three-branch semantic segmentation network module includes a boundary branch, a context branch, a detail branch, a fusion layer and a fully connected layer; The color depth image is input into the three-branch semantic segmentation network module to obtain the recognized bounding box and category.

3. The delivery robot according to claim 2, characterized in that: The boundary branch is composed of multiple convolutional layers, difference layers, attention mechanisms, and pooling layers, and is used to output the extracted boundary features; The context branch is composed of multiple convolutional layers, long short-term memory networks and pooling layers, and is used to output the extracted context information; The detail branch consists of multiple convolutional layers, attention mechanisms, and pooling layers, and is used to output the extracted detail information.

4. The delivery robot according to claim 2, characterized in that: Also includes: Compute network output loss, weighted binary cross entropy loss, standard cross entropy loss, and boundary-aware cross loss; The network output loss, weighted binary cross entropy loss, standard cross entropy loss and boundary-aware cross loss are weighted summed to obtain the total loss of the three-branch semantic segmentation network module.

5. The delivery robot according to claim 4, characterized in that: Calculating the boundary-aware intersection loss includes: Adopting adaptive threshold graph learning algorithm to predict adaptive threshold; Based on the adaptive threshold and differential binarization method, a predicted binarized boundary map is obtained; Calculating the difference between the predicted binary boundary map and the true binary boundary map; The boundary-aware intersection loss is obtained based on the sum of weighted differences at all pixel positions.

6. The delivery accompanying robot according to claim 1, characterized in that: Based on the bounding box and the category, identifying the moving target to be tracked includes: Calculate the 3D world coordinate centroid of each object based on the center of the bounding box and the corresponding depth pixels; Calculate the object's moving speed through the change in the center of mass position and the time interval between two consecutive frames; Based on the object's moving speed and a preset threshold, the moving target to be tracked is identified.

7. The delivery accompanying robot according to claim 1, characterized in that: The pain recognition module is used to obtain the mother's facial expression data and the conductivity information collected by the skin electrical sensor; The maternal expression data and the conductivity information collected by the skin electrical sensor are input into a preset fusion decision algorithm to obtain the maternal pain detection level.

8. The delivery robot according to claim 1 or 7, characterized in that: Also includes: A touch display screen is used to display the pain detection level in the form of a time series curve on the touch display screen of the robot.

9. The delivery robot according to claim 1 or 7, characterized in that: Also includes: The alarm module is used to issue an alarm message when the detected pain level remains at the highest level and continues for a preset period of time.

10. An adaptive movement method of a delivery robot, characterized in that: include: Collect color depth images of the surrounding environment; Acquire the color depth image, input the color depth image into a preset simultaneous localization and mapping algorithm, output the identified bounding box and category, and create an environment map; Based on the bounding box and the category, identifying a moving target to be tracked; Planning a moving path of the robot based on the position of the moving target and the environment map; The robot is controlled to track the target based on the planned moving path.

Images