Data acquisition intelligent interaction robot

Abstract

The invention discloses a data acquisition intelligent interaction robot, relates to the technical field of intelligent interaction, is suitable for a museum, realizes autonomous navigation by integrating a laser radar and a visual SLAM (Simultaneous Localization and Mapping) technology, and performs real-time perception and early warning on audience behaviors by fusing a depth camera and a posture recognition model. The system has a multi-modal interaction capability, supports voice recognition, semantic understanding, touch interaction and knowledge retrieval, automatically triggers an exhibit explanation event according to audience behaviors, realizes aligned broadcast in combination with a navigation fine tuning function, and improves immersion and interactivity of explanation; the robot can dynamically generate behavior data such as interaction logs, audience staying data and explanation content calling frequency, structured processing is conducted through the behavior data caching module, the behavior data are periodically synchronized to the background of the museum through the communication module, and data closed loop and system optimization are achieved; according to the robot, the convenience, safety and personalized service level of exhibit information acquisition are improved, and the robot has good practical value and popularization prospect.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of intelligent interaction, and particularly relates to a data acquisition intelligent interaction robot. BACKGROUND

[0002] The existing museum service robot mainly relies on laser radar or visual SLAM technology to realize autonomous navigation, and combines voice recognition and broadcast functions to complete basic route touring and exhibit explanation tasks. In terms of human-computer interaction, the audience and the robot generally realize information exchange through voice dialogue or touch screens, and part of the system can query and display multimedia of the exhibit information. At the same time, part of the robots are equipped with cameras and depth sensors for environment perception to support obstacle avoidance and simple crowd detection functions, which play a positive role in improving the exhibition experience and information service;

[0003] However, the existing museum service robot mainly focuses on path navigation and simple voice broadcast, lacks fine perception and intelligent response ability to the audience behavior, and is difficult to realize active identification of the audience state and personalized push of the interaction content. At the same time, the existing system mainly adopts a single sensor or a fixed route touring mode, lacks flexible task scheduling and adaptive path planning ability, and is prone to problems such as untimely obstacle avoidance and inaccurate interaction in a complex crowd environment, thereby limiting the intelligent and practical development of the robot in a public cultural scene. SUMMARY

[0004] In view of the above problems, the application provides a data acquisition intelligent interaction robot.

[0005] The application can be realized by the following technical scheme: a data acquisition intelligent interaction robot, comprising a path planning module, a navigation control module, a behavior perception module, a visual recognition module, a voice broadcast module, a content retrieval module, a voice recognition module and a behavior data cache module.

[0006] The navigation control module controls the robot to move according to the work map data and the task path, and synchronizes the real-time acquisition of the robot position information to the behavior perception module and the voice interaction module in real time.

[0007] The behavior perception module generates an explanation trigger event instruction by analyzing the behavior perception of the current position information of the robot, and sends the explanation trigger event instruction and the corresponding exhibit code to the voice broadcast module. After receiving the explanation trigger event instruction, the voice broadcast module calls the exhibit knowledge base, extracts the corresponding explanation content according to the exhibit code, and then calculates the horizontal plane distance and the orientation angle according to the three-dimensional position of the audience provided by the behavior perception module to determine the optimal interaction distance.

[0008] The voice recognition module continuously monitors the microphone input, captures a valid voice segment, sends the valid voice segment to the ASR engine to obtain a transcription text, calls the named entity recognition, and identifies the exhibit name; the query request is delivered to the content retrieval module through an internal message bus for further processing;

[0009] The content retrieval module receives and parses the query request, extracts the intent label and target exhibit code list therein for retrieval, and pushes to the voice broadcast module;

[0010] The behavior data cache module collects the output data of other modules;

[0011] The communication module periodically synchronizes data with the museum back-end through the network, transmits the cached data, and clears the uploaded part.

[0012] As a preferred embodiment of the application, the specific process of the navigation control module controlling the robot to move according to the working map data and the task path is as follows:

[0013] The navigation control module sends a start pre-scanning instruction to the laser radar and visual SLAM sensor, instructs the sensor to return point cloud and image data, the map management unit fuses with the local Topo map to generate working map data; then obtains the tour task data issued by the museum back-end management system from the task scheduling module; and the path planning module receives the tour task data, calculates the tour task path according to the tour task data and the SLAM composition result and the working map data, and sends the obtained path points to the navigation control module;

[0014] The navigation control module controls the robot chassis to move and drives the robot to move according to the path points; in the moving process, the robot position state is acquired in real time, and obstacle avoidance is realized in cooperation with the laser radar and visual SLAM sensor; and the real-time acquired robot position information is synchronously sent to the behavior perception module and the voice interaction module.

[0015] As a preferred embodiment of the application, the specific process of generating the explanation triggering event instruction by the behavior perception analysis is as follows:

[0016] The image frame and depth map at the current time are extracted from the depth camera and RGB camera synchronous processing; the depth and RGB are aligned to generate a multi-channel fusion matrix; the fusion matrix is scaled to a fixed size according to the model input requirement, and the pixel value is normalized to [0, 1];

[0017] According to the sliding window strategy, a plurality of overlapping subgraphs are cropped, key point detection is performed on the plurality of overlapping subgraphs one by one, the normalized multi-channel image is input into a pre-trained OpenPose-style CNN, and an N x (x, y, confidence) format skeleton key point list is output; and the points with a key point confidence lower than a threshold value 0.3 are discarded; by using the depth channel, the 2D key points are mapped back to a 3D point set in the robot coordinate system;

[0018] In the behavior perception module, a key point sequence cache 3.2 speed and acceleration calculation in the past T=2 seconds is maintained:

[0019] Difference is made in the time dimension for each key point, and a speed vector and an acceleration vector are calculated; the 3D coordinates of the skeleton of each frame, the speed, the acceleration, the joint angle and the hand-exhibit distance are spliced into a d-dimensional feature vector; and all frame features in the sliding window are collected into a Txd time sequence feature matrix; the time sequence feature matrix is input into a lightweight time sequence model to output three classification probabilities, including an ordinary state, a close state and a dangerous posture; the output three classification probabilities are compared with corresponding posture determination preset threshold values to obtain a warning posture or a close pre-warning or an ordinary state; the classification results and the 3D position of the audience at the corresponding time are transmitted through an internal message bus, and a warning event is pushed to the warning processing module;

[0020] The behavior perception module receives the view data, and judges whether the audience staying time exceeds a set threshold value in combination with the time sequence data; if a certain audience long gazes at a certain exhibit, a explanation triggering event instruction is generated.

[0021] As a preferred embodiment of the application, the specific process for determining the optimal interaction distance is:

[0022] According to the three-dimensional position (x a , y a , z a ) of the audience provided by the behavior perception module, relative to the robot coordinate system; the navigation module reads the current pose (x r , y r , theta r ) of the robot, wherein theta r is the orientation of the robot;

[0023] The horizontal plane distance and the orientation angle are calculated, specifically: The optimal interaction distance d opt is set in the system parameters; the distance that needs to be advanced or retreated is calculated: delta d = D 当前 -d opt ;

[0024] Regarding fine-tuning of the plan: Calculate the yaw error, Δθ = normalizeAngle(m 目标 -θ r ),

[0025] The normalizeAngle function normalizes the angle to [-π, π].

[0026] Moving along the current direction, v lin =K d ×Δd, v ang =K θ ×Δθ,K d K θ This is the velocity gain coefficient; (v lin v ang The command is converted into a chassis motion command and gradually adjusted to the target point; during the fine-tuning motion execution, the command is continuously updated (x). r y r θ r Based on the audience's pose, recalculate Δd and Δθ in real time; and determine whether |Δd| < ∈ d and |Δθ|<∈ θ Then, stop fine-tuning and switch to static interactive state.

[0027] In a preferred embodiment of the present invention, the specific process of calling named entity recognition to identify the exhibit name is as follows:

[0028] Once a valid speech segment is captured, it is sent to the ASR engine to obtain the transcribed text. The text is then denoised, segmented, and filtered for stop words, and input into the intent recognition model to obtain intent tags. At the same time, named entity recognition is used to identify the exhibit name. Screen click events are monitored. If the user selects the corresponding exhibit via touch screen, the exhibit code is read and prioritized in case of entity conflict. The intent tag, entity information, user language preference, and current timestamp are encapsulated into a unified query request object. For the case of "query the next exhibit" without entity information, the empty target field is retained for automatic backend location.

[0029] In a preferred embodiment of the present invention, the specific process of extracting the intent tags and target exhibit code list for retrieval is as follows:

[0030] The retrieval strategy is determined based on the intent tag. If the intent is "to query the next exhibit", the next exhibit code is located in the exhibit sequence table; otherwise, the exhibit code provided in the request is used directly. Next, the corresponding SQL query is constructed, and the exhibit code is used as a condition to retrieve data fields including exhibit title, detailed description, image URL, video URL, and QR code from the exhibit table. The returned multimedia links are validated and pre-signed, and the text content is localized according to the user's language preference. Then, all fields are organized into a standardized response object.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] 1. This invention constructs a multi-channel fusion visual perception mechanism by integrating depth cameras, RGB images and visual SLAM technology, and combines OpenPose pose estimation and temporal depth feature analysis model to identify the approach behavior and dangerous postures of the audience in real time, trigger warning events and respond proactively, realizing proactive protection of exhibit safety and intelligent early warning of behavior, and significantly improving the robot's safe interaction capabilities in complex environments such as museums.

[0033] 2. This invention supports a fusion of voice recognition, intent understanding, touch selection, and content retrieval. It can generate explanation trigger events based on audience behavior, automatically call the exhibit knowledge base for voice, text, and video explanations, and achieve directional broadcasting through navigation fine-tuning. This improves the convenience of obtaining exhibit information and the interactive experience, and enhances the immersive experience and personalized service level of the visitor. Attached Figure Description

[0034] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0035] Figure 1 This is a schematic diagram showing the connections of the various modules of the present invention. Detailed Implementation

[0036] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figure 1 As shown, a data acquisition intelligent interactive robot includes a task scheduling module, a path planning module, a navigation control module, a behavior perception module, a visual recognition module, a voice broadcasting module, a touch control module, a content retrieval module, a voice recognition module, a behavior data caching module, and a communication module.

[0038] The following describes how the aforementioned intelligent interactive robot is applied in a museum for intelligent interactive work:

[0039] The navigation control module sends a "start pre-scan" command to the LiDAR and visual SLAM sensors. The sensors return point cloud and image data, which the map management unit merges with the local Topo map to generate working map data. Then, it obtains the tour task data issued by the museum's back-end management system from the task scheduling module, including time, area list, and inspection frequency. The path planning module receives the tour task data, calculates the tour task path based on the SLAM mapping results and the working map data, and then sends the obtained path points to the navigation control module.

[0040] The navigation control module controls the movement of the robot chassis (such as hub motors, servo electronic controls, etc.) and drives the robot to move according to the path points. During the movement, the robot's position status is acquired in real time, and obstacle avoidance is achieved in conjunction with LiDAR / visual SLAM sensors. The real-time robot position information is also synchronized to the behavior perception module and the voice interaction module in real time.

[0041] The behavior perception module performs behavior perception analysis on the robot's current position information, specifically as follows:

[0042] The visual recognition module calls the depth camera to continuously collect surrounding images and obtain real-time visual data; the convolutional neural network model is used to detect the number, position, posture and gaze direction of the audience, which are recorded as visual data; the output visual data is transmitted to the behavior analysis module.

[0043] Further processing is performed based on the obtained audience posture to detect and obtain audience warning postures. Warnings are then issued based on these postures. Specifically:

[0044] Simultaneous processing from the depth camera and RGB camera extracts the image frame and depth map of the current moment; and aligns the depth and RGB to generate a multi-channel fusion matrix (H×W×4: R / G / B / D); the fusion matrix is ​​scaled to a fixed size (e.g., 256×256) according to the model input requirements, and the pixel values ​​are normalized to [0, 1].

[0045] Next, the image is cropped into several overlapping sub-images using a sliding window strategy to ensure no blind spots in large scenes;

[0046] Keypoint detection (pose estimation) is performed on each of the obtained overlapping sub-images. The normalized multi-channel image is then input into a pre-trained OpenPose-style CNN, which outputs a list of skeleton keypoints in N×(x, y, confidence) format. Keypoints with confidence scores below a threshold of 0.3 are discarded. Finally, the 2D keypoints are mapped back to a 3D point set in the robot coordinate system using the depth channel.

[0047] In the behavior awareness module, a keypoint sequence cache is maintained for the past T = 2 seconds (e.g., 60 frames). 3.2 Velocity and acceleration calculation:

[0048] For each keypoint, the time dimension is subtracted to calculate the velocity and acceleration vectors; the relative angles of joints such as shoulder-wrist and torso-head are calculated, as well as the distance from the hand to the nearest exhibit plane; the 3D coordinates, velocity, acceleration, joint angles, and hand-exhibit distance of each frame are concatenated into a d-dimensional feature vector; finally, the features of all frames within the sliding window are aggregated into a T×d temporal feature matrix.

[0049] The temporal feature matrix is ​​input into a lightweight temporal model (such as 1D-Conv+BiLSTM+fully connected layer) and outputs three classification probabilities, including: normal state P0, proximity state P1, and dangerous posture P2. The output three classification probabilities are then compared with the corresponding preset posture determination thresholds (such as 0.7). If P2 is greater than 0.7, it is determined as a "warning posture"; otherwise, if P1 is greater than 0.7, it is determined as a "proximity warning"; otherwise, the normal state is maintained.

[0050] Finally, the classification results, the 3D positions of the audience at the corresponding time, the coordinates of the main key points, and the model confidence are encapsulated into a warning event data structure; and the warning event is pushed to the warning processing module through the ROS service or internal message bus.

[0051] The behavior perception module receives visual data and combines it with time-series data to determine whether the audience's dwell time exceeds the set threshold for voice broadcasting. If an audience member stares at an exhibit for a long time, an "explanation trigger event" is generated. At the same time, the event and the corresponding exhibit code are sent to the voice broadcasting module.

[0052] The broadcast module receives the "explanation trigger event," calls the exhibit knowledge base, and extracts the corresponding explanation content based on the exhibit code; it uses a TTS speech synthesis chip to convert the text into speech output; and it simultaneously displays exhibit information, including images, videos, and QR codes, on the touch screen.

[0053] After the event is triggered, the behavior perception module provides the audience's three-dimensional position (x, y). a y a , z a(x) relative to the robot's coordinate system; the navigation module reads the robot's current pose (x) r y r θ r ), where θ r Orientation of the robot;

[0054] The optimal interaction distance is determined by calculating the horizontal plane distance and the orientation angle, specifically as follows: By setting the optimal interaction distance d in the system parameters opt (For example, 1.2m); the distance to move forward / backward is calculated as: Δd = D 当前 -d opt ;

[0055] Regarding fine-tuning of the plan: Calculate the yaw error, Δθ = normalizeAngle(m 目标 -θ r ),

[0056] The normalizeAngle function normalizes the angle to [-π, π].

[0057] Moving along the current direction, v lin =K d ×Δd, v ang =K θ ×Δθ, where K d K θ This is the velocity gain coefficient; (v lin v ang The command is converted into a chassis motion command and gradually adjusted to the target point; during the fine-tuning motion execution, the command is continuously updated (x). r y r θ r Based on the audience's pose, recalculate Δd and Δθ in real time; and determine whether |Δd| < ∈ d and |Δθ|<∈ θ After setting the (preset convergence threshold), stop fine-tuning and switch to static interactive state;

[0058] The above is the complete calculation process for the "navigation fine-tuning + audio pointing" stage, ensuring that the robot can quickly and smoothly approach the audience and accurately align with their position for voice interaction.

[0059] The system uses a voice recognition module to listen to audience voice input, such as "Introduce the next exhibit," "Whose work is this?", as well as inappropriate language and questions about seating arrangements. Simultaneously, a touchscreen module allows audiences to select exhibits of interest via screen taps. Based on the monitored audience voice data and selected exhibits, a query request is generated and sent to the content retrieval module. Specifically:

[0060] The speech recognition module continuously monitors microphone input. Once a valid speech segment is captured, it is sent to the ASR engine to obtain the transcribed text. The text is then denoised, segmented, and filtered for stop words, and input into the intent recognition model to obtain intent tags (such as "query the next exhibit" or "query the author"). Simultaneously, named entity recognition is used to identify the exhibit name or number. The touch module listens for screen click events in parallel. If the user selects an exhibit via touch, the exhibit's code is read, and this code is prioritized in case of entity conflict. Next, the system encapsulates the intent tag, entity information (or touch code), user language preference, and current timestamp into a unified query request object. For the case of "query the next exhibit" without entity information, an empty target field is retained for automatic backend location. Finally, the query request is delivered to the content retrieval module for further processing via the internal message bus.

[0061] The content retrieval module first receives and parses the query request, extracting the intent tags (such as "query the next exhibit" or "query the author") and the target exhibit code list. Then, it determines the retrieval strategy based on the intent tags: if it's "query the next exhibit," it locates the next exhibit code in the exhibit sequence table; otherwise, it directly uses the exhibit code provided in the request. Next, the system constructs the corresponding SQL query or calls the retrieval engine API, retrieving fields including exhibit title, detailed description, image URL, video URL, and QR code data from the exhibit table or index, using the exhibit code as the condition. The returned multimedia links are validated and pre-signed, and the text content is localized according to the user's language preferences. All fields are then organized into a standardized response object. Finally, this response object is pushed to the voice broadcast module, where TTS generates and plays the audio narration; it is also passed to the display module, rendering images, videos, and QR codes on the touchscreen. Simultaneously, the retrieval log is recorded and asynchronously uploaded to the backend for subsequent monitoring and optimization.

[0062] The behavior data caching module collects and summarizes the output data from other modules, including the current location of the exhibition area, the number of visitors and their dwell time, the exhibit number and number of interactions, and the frequency of the explanation content being called; it then processes the data into structured JSON format and stores it in a temporary cache.

[0063] The communication module periodically synchronizes data with the museum's backend via WLAN / 5G, transmits cached data, and clears uploaded data.

[0064] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A data acquisition intelligent interactive robot, comprising a path planning module, a navigation control module, a behavior perception module, a visual recognition module, a voice broadcasting module, a content retrieval module, a voice recognition module, and a behavior data caching module; characterized in that: The navigation control module controls the robot to move based on the work map data and task path, and synchronizes the real-time robot position information to the behavior perception module and voice interaction module in real time. The behavior perception module performs behavior perception analysis on the robot's current position information to generate explanation trigger event instructions, and at the same time sends the explanation trigger event instructions and the corresponding exhibit codes to the voice broadcast module; After receiving the instruction to trigger the explanation, the voice broadcast module calls the exhibit knowledge base, extracts the corresponding explanation content according to the exhibit code, and then calculates the horizontal plane distance and orientation angle based on the three-dimensional position of the audience provided by the behavior perception module, and determines the optimal interaction distance. The speech recognition module continuously monitors microphone input, captures valid speech segments, sends the valid speech segments to the ASR engine to obtain transcribed text, calls named entity recognition, and identifies the exhibit names; query requests are sent to the content retrieval module for further processing via the internal message bus; The content retrieval module receives and parses the query request, extracts the intent tags and target exhibit code list for retrieval, and pushes them to the voice broadcast module; The behavior data caching module retrieves and aggregates the output data from other modules. The communication module periodically synchronizes data with the museum's backend via the network, transmits cached data, and clears uploaded data.

2. The intelligent interactive data acquisition robot according to claim 1, characterized in that, The specific process by which the navigation control module controls the robot's movement based on the working map data and the task path is as follows: The navigation control module sends a pre-scan start command to the lidar and visual SLAM sensor, instructing the sensor to return point cloud and image data. The map management unit merges the data with the local Topo map to generate working map data. Then, the tour task data issued by the museum's back-end management system is obtained from the task scheduling module; The path planning module receives the cruise mission data, calculates the cruise mission path based on the SLAM mapping results and working map data, and then sends the obtained path points to the navigation control module. The navigation control module controls the movement of the robot chassis and drives the robot to move according to the path points. During the movement, it acquires the robot's position status in real time and uses LiDAR and visual SLAM sensors to avoid obstacles. The real-time robot position information is also synchronized to the behavior perception module and the voice interaction module.

3. The intelligent interactive robot for data acquisition according to claim 2, characterized in that, The specific process of performing the behavior perception analysis to generate and explain the trigger event instructions is as follows: Simultaneous processing from depth camera and RGB camera extracts the image frame and depth map of the current moment; and aligns the depth and RGB to generate a multi-channel fusion matrix; the fusion matrix is ​​scaled to a fixed size according to the model input requirements, and the pixel values ​​are normalized to [0, 1]; The image is cropped into several overlapping sub-images using a sliding window strategy. Keypoint detection is performed on each of the overlapping sub-images. The normalized multi-channel image is then input into a pre-trained OpenPose-style CNN, which outputs a list of skeleton keypoints in N×(x, y, confidence) format. Points with a confidence level below the threshold of 0.3 are discarded; the 2D keypoints are mapped back to a 3D point set in the robot coordinate system using the depth channel. In the behavior perception module, a cache of key point sequences within the past T=2 seconds is maintained. 3.2 Velocity and acceleration calculation: For each keypoint, the velocity and acceleration vectors are calculated by subtracting them over time. The 3D skeleton coordinates, velocity, acceleration, joint angles, and hand-exhibit distance of each frame are concatenated into a d-dimensional feature vector. The features of all frames within the sliding window are aggregated into a T×d temporal feature matrix. The temporal feature matrix is ​​input into a lightweight temporal model to output three-class probabilities, including normal state, proximity state, and dangerous posture. The output three-class probabilities are then compared with the corresponding preset posture judgment thresholds to obtain a warning posture, proximity warning, or normal state. The classification results and the 3D position of the audience at the corresponding time are pushed to the warning processing module via the internal message bus. The behavior perception module receives visual data and combines it with time-series data to determine whether the audience's dwell time exceeds a set threshold. If a visitor stares at an exhibit for an extended period of time, a prompt to provide an explanation will be generated.

4. The intelligent interactive data acquisition robot according to claim 3, characterized in that, The specific process for determining the optimal interaction distance is as follows: Based on the audience's three-dimensional position (x) provided by the behavior perception module a y a , z a (x) relative to the robot's coordinate system; the navigation module reads the robot's current pose (x) r y r θ r ), where θ r Orientation of the robot; By calculating the horizontal plane distance and the orientation angle, specifically: Set the optimal interaction distance d in the system parameters. opt The distance to move forward or backward is calculated as follows: Δd = D 当前 -d opt ; Regarding fine-tuning of the plan: Calculate the yaw error, Δθ = normalizeAngle(m 目标 -θ r ), The normalizeAngle function normalizes the angle to [-π, π]. Moving along the current direction, v lin =K d ×Δd, v ang =K θ ×Δθ,K d K θ This is the velocity gain coefficient; (v lin v ang Convert the command to chassis movement and gradually adjust to the target point; During the fine-tuning process, continuously update (x) r y r θ r Based on the audience's pose, recalculate Δd and Δθ in real time; and determine whether |Δd| < ∈ d and |Δθ|<∈ θ Then, stop fine-tuning and switch to static interactive state.

5. The intelligent interactive data acquisition robot according to claim 4, characterized in that, The specific process of calling named entity recognition to identify the exhibit name is as follows: Once a valid speech segment is captured, it is sent to the ASR engine to obtain the transcribed text. The text is then denoised, segmented, and filtered for stop words, and input into the intent recognition model to obtain intent tags. At the same time, named entity recognition is used to identify the exhibit name. Screen click events are monitored. If the user selects the corresponding exhibit via touch screen, the exhibit code is read and prioritized in case of entity conflict. The intent tag, entity information, user language preference, and current timestamp are encapsulated into a unified query request object. For the case of "query the next exhibit" without entity information, the empty target field is retained for automatic backend location.

6. The intelligent interactive data acquisition robot according to claim 5, characterized in that, The specific process of extracting the intent tags and target exhibit codes for retrieval is as follows: The retrieval strategy is determined based on the intent tag. If the intent is "to query the next exhibit", the next exhibit code is located in the exhibit sequence table; otherwise, the exhibit code provided in the request is used directly. Next, the corresponding SQL query is constructed, and the exhibit code is used as a condition to retrieve data fields including exhibit title, detailed description, image URL, video URL, and QR code from the exhibit table. The returned multimedia links are validated and pre-signed, and the text content is localized according to the user's language preference. Then, all fields are organized into a standardized response object.

Images