Camera quick calling and visualization method based on natural language understanding

Abstract

The invention discloses a camera quick calling and visualization method based on natural language understanding, and relates to the technical field of video monitoring and artificial intelligence crossing. Comprising the steps that S1, a full-amount camera database is constructed, S2, a camera space retrieval interface is provided, and S3, a map visualization front end is constructed; s4, providing an address-coordinate conversion service; s5, local privatization deployment: deploying a camera space retrieval interface, a front-end static resource and a large language model mirror image in an intranet GPU server; s6, constructing a large model knowledge newly establishing a knowledge base table Knowledgeprompt in the model mirror image for writing data, and constructing the large model knowledge base, and S7, receiving a query demand of a user for inputting a natural language in a dialog box; s8, performing semantic understanding, and calling an interface to return data; step S9, automatically dotting the map; and S10, playing the video stream as required.

Description

Technical Field

[0001] This invention discloses a method for rapid retrieval and visualization of camera footage based on natural language understanding, which relates to the interdisciplinary fields of video surveillance and artificial intelligence. Background Technology

[0002] Traditional methods of retrieving camera data mainly rely on the following two technical approaches:

[0003] Directory tree hierarchical retrieval: In video management platforms such as traffic integrated control platforms, users first select administrative division → road → intersection → camera, which requires clicking 5-7 levels of menus on average. The operation chain is long and requires a high level of professionalism from front-line commanders.

[0004] Map selection by bounding boxes / circles: Rectangular or circular areas are manually drawn on a GIS map, and the backend returns a list of cameras via spatial SQL. While intuitive, this method still requires users to first determine the approximate extent of the target area and cannot support real-time requests in natural language; when tasks are urgent, interaction efficiency becomes a bottleneck.

[0005] In recent years, some manufacturers have attempted to introduce voice input, but these efforts have only reached the keyword matching stage: first, speech is converted to text using ASR (Automatic Speech Recognition), then regular expressions are used to extract addresses and distances, and finally, a spatial retrieval interface is called. This solution has extremely low tolerance for colloquialisms, elliptical sentences, and synonyms, resulting in a parsing failure rate >30%. Furthermore, camera databases, geocoding services, and video playback components are often deployed across different cloud vendors, and cross-public network transmission poses risks of data leakage and link latency, failing to meet the needs of confidential scenarios. Therefore, existing technologies still have at least the following drawbacks: 1. Long retrieval path, with an average time >45 seconds; 2. Unfriendly to non-professional users, requiring prior knowledge of geographical information; 3. Low accuracy of natural language parsing, leading to missed detections and false detections; 4. Cross-public network transmission poses both data security and latency risks. Summary of the Invention

[0006] This invention addresses the problems of existing technologies by providing a method for rapid retrieval and visualization of cameras based on natural language understanding. It solves the problems of "slow retrieval, high threshold, low accuracy, and insecurity" in existing technologies, and achieves a natural language-driven, second-level response, local closed-loop, and plug-and-play camera retrieval experience. It can be widely used in classified or non-classified scenarios such as emergency response, transportation, urban management, and industrial parks.

[0007] The specific solution proposed in this invention is as follows:

[0008] This invention provides a method for rapid retrieval and visualization of camera feeds based on natural language understanding, comprising:

[0009] Step S1: Build a full camera database: Create a new table named camera_info in the local PostgreSQL instance. Aggregate camera information from heterogeneous data sources into the camera_info table to form a full camera database.

[0010] Step S2: Provide a camera space retrieval interface: Deploy a Spring Boot-based RESTful service on the local server, exposing the ` / api / camera / nearby` interface to receive HTTP-GET requests.

[0011] Step S3: Build the map visualization front-end: Use Vue3 + OpenLayers framework as the front-end framework for electronic map loading and tile rendering; implement dynamic dot layers and support custom camera icons; build a video pop-up component to support automatic transcoding and playback from RTSP to WebRTC.

[0012] Step S4: Provide address-coordinate conversion service: Integrate a third-party geocoding SDK through the local server to convert the natural language address entered by the user into latitude and longitude coordinate pairs. The conversion result is used for subsequent spatial retrieval.

[0013] Step S5: Local private deployment: Deploy the camera spatial retrieval interface, front-end static resources, and large language model image together on the intranet GPU server;

[0014] Step S6: Build a large model knowledge base: Create a new knowledge base table `knowledge_prompt` in the model mirror for writing data and building the large model knowledge base.

[0015] Step S7: Receive the user's query request in natural language entered in the dialog box;

[0016] Step S8: Perform semantic understanding and then call the interface to return data: Use the large language model to determine that the intent category is spatial retrieval, call the address-coordinate conversion service, parse the address, extract the distance, assemble the HTTP request and access the camera spatial retrieval interface to obtain a JSON array, and encapsulate the array into structured text and return it to the front end;

[0017] Step S9: Perform automatic map point marking: Generate OpenLayers Feature objects one by one from the JSON array returned by the front end, and render the camera points in real time on the VectorLayer of the electronic map;

[0018] Step S10: On-demand playback of video stream: Click any camera icon on the front end, and request the RTSP transcoding address from the backend video gateway with the camera's channel_code as the parameter. After obtaining the WebRTC stream, the video will automatically play in the pop-up window, enabling click-to-play functionality.

[0019] Furthermore, in step 1 of the described method for fast retrieval and visualization of camera data based on natural language understanding, when creating the `camera_info` table, the table structure must include at least the following fields:

[0020] camera_name: The readable name of the camera;

[0021] longitude: The longitude of the camera installation location, in degrees, rounded to six decimal places;

[0022] latitude: The latitude of the camera installation location, in degrees, rounded to six decimal places;

[0023] channel_code: The unique channel code of the camera in the video management platform, used for subsequent streaming.

[0024] Furthermore, in step 2 of the aforementioned method for rapid retrieval and visualization of cameras based on natural language understanding, when the camera spatial retrieval interface receives an HTTP-GET request, the input parameters include:

[0025] poiLon—POI: Longitude of the center point;

[0026] poiLat—POI: Latitude of the center point;

[0027] distance: Search radius, in meters;

[0028] The interface internally uses PostGIS's ST_DWithin function to calculate spherical distance and returns a list of cameras within the distance range. Each element in the list must contain at least the four fields described in the camera_info table.

[0029] Furthermore, in step 6 of the aforementioned method for rapid retrieval and visualization of camera data based on natural language understanding, when creating the knowledge base table `knowledge_prompt` in the model image, the following content is written:

[0030] Camera database field descriptions;

[0031] The URL, request method, meaning of input parameters, and examples of the camera spatial retrieval interface;

[0032] And examples of typical user questions.

[0033] This invention also provides a device for rapid retrieval and visualization of camera data based on natural language understanding, including a full-database camera management module, an interface management module, a front-end management module, an address service module, a deployment module, a large model knowledge base management module, a receiving module, and a semantic understanding module.

[0034] The full camera database management module constructs a full camera database: A new table named `camera_info` is created in the local PostgreSQL instance, and camera information from heterogeneous data sources is aggregated into the `camera_info` table to form the full camera database.

[0035] The interface management module provides a camera space retrieval interface: a Spring Boot-based RESTful service is deployed on the local server, exposing the ` / api / camera / nearby` interface to receive HTTP-GET requests.

[0036] The front-end management module builds a map visualization front-end: it uses the Vue3 + OpenLayers framework as the front-end framework for loading electronic maps and rendering tiles; it implements dynamic dot layers, supports custom camera icons; and it builds a video pop-up component that supports automatic transcoding and playback from RTSP to WebRTC.

[0037] The address service module provides address-coordinate conversion service: by integrating a third-party geocoding SDK on the local server, it converts the natural language address entered by the user into latitude and longitude coordinate pairs, and the conversion result is used for subsequent spatial retrieval;

[0038] Deployment module local private deployment: Deploy the camera spatial retrieval interface, front-end static resources, and large language model image together on the internal network GPU server;

[0039] The large model knowledge base management module constructs the large model knowledge base: A new knowledge base table, `knowledge_prompt`, is created in the model image for writing data and building the large model knowledge base.

[0040] The receiving module receives the user's query request in natural language entered in the dialog box;

[0041] The semantic understanding module performs semantic understanding and then calls the interface to return data: using the large language model to determine that the intent category is spatial retrieval, calling the address-coordinate conversion service, parsing the address, extracting the distance, assembling the HTTP request and accessing the camera spatial retrieval interface to obtain a JSON array, and encapsulating the array into structured text to return to the front end;

[0042] The front-end management module automatically marks points on the map: it generates OpenLayers Feature objects one by one by parsing the JSON array returned by the front-end, and renders the camera points in real time on the VectorLayer of the electronic map;

[0043] The front-end management module allows for on-demand playback of video streams: clicking any camera icon on the front end sends an RTSP transcoding address to the back-end video gateway, using the camera's channel_code as a parameter. After obtaining the WebRTC stream, the video is automatically played in a pop-up window, enabling click-to-play functionality.

[0044] Furthermore, when creating a new `camera_info` table in the full-database database management module for a natural language understanding-based camera fast retrieval and visualization device, the table structure must include at least the following fields:

[0045] camera_name: The readable name of the camera;

[0046] longitude: The longitude of the camera installation location, in degrees, rounded to six decimal places;

[0047] latitude: The latitude of the camera installation location, in degrees, rounded to six decimal places;

[0048] channel_code: The unique channel code of the camera in the video management platform, used for subsequent streaming.

[0049] Furthermore, when the interface management module of the camera fast retrieval and visualization device based on natural language understanding receives an HTTP-GET request at the camera spatial retrieval interface, the input parameters include:

[0050] poiLon—POI: Longitude of the center point;

[0051] poiLat—POI: Latitude of the center point;

[0052] distance: Search radius, in meters;

[0053] The interface internally uses PostGIS's ST_DWithin function to calculate spherical distance and returns a list of cameras within the distance range. Each element in the list must contain at least the four fields described in the camera_info table.

[0054] Furthermore, when creating a new knowledge base table `knowledge_prompt` in the model image, the large model knowledge base management module of the natural language understanding-based camera fast retrieval and visualization device writes the following content:

[0055] Camera database field descriptions;

[0056] The URL, request method, meaning of input parameters, and examples of the camera spatial retrieval interface;

[0057] And examples of typical user questions.

[0058] The advantages of the method of the present invention are:

[0059] (1) Interaction efficiency is improved by an order of magnitude. By combining a large language model with a local knowledge base, users only need to input a colloquial description to complete the entire chain of operations from semantic understanding to address parsing, spatial retrieval, map point marking, and video playback within 3 seconds. The average time taken is reduced from 45 seconds to 3 seconds, meeting the "second-level response" requirement for emergencies.

[0060] (2) The parsing accuracy has been greatly improved. By leveraging the capabilities of the large model in semantic understanding, reference resolution, and synonym expansion, end-to-end understanding of colloquial, inverted, and elliptical sentences is achieved. On the self-built test set, the address extraction accuracy reaches 98.7% and the distance extraction accuracy reaches 99.2%, which is significantly higher than the 70% of the traditional regularization scheme.

[0061] (3) Zero learning cost user experience. Users do not need to memorize professional information such as administrative divisions, road names, and camera numbers, nor do they need to master map drawing tools. They can retrieve target videos by describing them in everyday language, truly achieving "barrier-free" operation, which can cover multiple roles such as commanders, community grid workers, and field police officers.

[0062] (4) Dual protection of data security and link latency. The camera database, geocoding service, large language model and video gateway are all deployed on the local intranet to achieve "data not leaving the domain and model not going offline", which meets the mandatory requirement that video streams cannot be transmitted across the public network, and avoids the additional latency of 300-500ms caused by public network RTT. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0064] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0065] Example 1

[0066] This invention provides a method for rapid retrieval and visualization of camera feeds based on natural language understanding, comprising:

[0067] Step S1: Build a full camera database: Create a new table named camera_info in the local PostgreSQL instance, and aggregate camera information from heterogeneous data sources into the camera_info table to form a full camera database.

[0068] The newly created table is named camera_info, and its structure must include at least the following fields:

[0069] camera_name — The readable name of the camera;

[0070] longitude — the longitude of the camera installation location, in degrees, rounded to six decimal places;

[0071] latitude—the latitude of the camera installation location, in degrees, rounded to six decimal places;

[0072] channel_code — The unique channel code of the camera in the video management platform, used for subsequent streaming.

[0073] Camera information from heterogeneous data sources such as Skynet, traffic, and social networks is aggregated into the camera_info table to form a full camera database.

[0074] Step S2: Provide a camera space retrieval interface: Deploy a Spring Boot-based RESTful service on the local server, exposing the / api / camera / nearby interface to receive HTTP-GET requests.

[0075] This interface accepts HTTP-GET requests, and the input parameters include:

[0076] poiLon—Longitude of the center point of POI;

[0077] poiLat—Latitude of the center point of POI;

[0078] distance—search radius, in meters;

[0079] The interface internally uses PostGIS's ST_DWithin function to calculate spherical distance and returns a list of cameras within the distance range. Each element in the list must contain at least the four fields described in S1.

[0080] Step S3: Build the map visualization front-end: Use Vue3 + OpenLayers framework as the front-end framework for electronic map loading and tile rendering; implement dynamic dot layers and support custom camera icons; build a video pop-up component to support automatic transcoding and playback from RTSP to WebRTC.

[0081] Step S4: Provide address-coordinate conversion service: Integrate a third-party geocoding SDK through the local server to convert the natural language address (structured string) entered by the user into latitude and longitude coordinate pairs. The conversion result is used for subsequent spatial retrieval.

[0082] Step S5: Local private deployment: Deploy the camera spatial retrieval interface, front-end static resources, and large language model image together on the internal network GPU server.

[0083] Step S6: Build a large model knowledge base: Create a new knowledge base table `knowledge_prompt` in the model mirror to write data and build a large model knowledge base.

[0084] In the model mirror, create a new knowledge base table named knowledge_prompt and write the following content:

[0085] Camera database field descriptions;

[0086] S2 interface URL, request method, input parameter meaning and example;

[0087] Typical user question example: "Retrieve camera feeds within 200 meters of XX Square". Through the Prompt-Engineering approach, the model gains the ability to infer from "address → latitude and longitude → interface call → result return".

[0088] Step S7: Receive the user's query request in natural language in the dialog box, such as "Retrieve the cameras within 300 meters of the XX intersection";

[0089] Step S8: Perform semantic understanding, then call the interface to return data:

[0090] The intent category was determined to be spatial retrieval using a large language model.

[0091] Call the address-coordinate conversion service to parse the address, resolving "XX intersection" into (poiLon, poiLat), and extracting "300 meters" as the distance.

[0092] Construct an HTTP request and access the camera space retrieval interface / api / camera / nearby to obtain a JSON array.

[0093] The array is encapsulated as structured text and returned to the front end;

[0094] Step S9: Perform automatic map point marking: Generate OpenLayers Feature objects one by one from the JSON array returned by the front end, and render the camera points in real time on the VectorLayer of the electronic map;

[0095] Step S10: On-demand playback of video stream: Click any camera icon on the front end, and request the RTSP transcoding address from the backend video gateway with the camera's channel_code as the parameter. After obtaining the WebRTC stream, the video will automatically play in the pop-up window, enabling click-to-play functionality.

[0096] Example 2

[0097] This invention also provides a device for rapid retrieval and visualization of camera data based on natural language understanding, including a full-database camera management module, an interface management module, a front-end management module, an address service module, a deployment module, a large model knowledge base management module, a receiving module, and a semantic understanding module.

[0098] The full camera database management module constructs a full camera database: A new table named `camera_info` is created in the local PostgreSQL instance, and camera information from heterogeneous data sources is aggregated into the `camera_info` table to form the full camera database.

[0099] The interface management module provides a camera space retrieval interface: a Spring Boot-based RESTful service is deployed on the local server, exposing the ` / api / camera / nearby` interface to receive HTTP-GET requests.

[0100] The front-end management module builds a map visualization front-end: it uses the Vue3 + OpenLayers framework as the front-end framework for loading electronic maps and rendering tiles; it implements dynamic dot layers, supports custom camera icons; and it builds a video pop-up component that supports automatic transcoding and playback from RTSP to WebRTC.

[0101] The address service module provides address-coordinate conversion service: by integrating a third-party geocoding SDK on the local server, it converts the natural language address entered by the user into latitude and longitude coordinate pairs, and the conversion result is used for subsequent spatial retrieval;

[0102] Deployment module local private deployment: Deploy the camera spatial retrieval interface, front-end static resources, and large language model image together on the internal network GPU server;

[0103] The large model knowledge base management module constructs the large model knowledge base: A new knowledge base table, `knowledge_prompt`, is created in the model image for writing data and building the large model knowledge base.

[0104] The receiving module receives the user's query request in natural language entered in the dialog box;

[0105] The semantic understanding module performs semantic understanding and then calls the interface to return data: using the large language model to determine that the intent category is spatial retrieval, calling the address-coordinate conversion service, parsing the address, extracting the distance, assembling the HTTP request and accessing the camera spatial retrieval interface to obtain a JSON array, and encapsulating the array into structured text to return to the front end;

[0106] The front-end management module automatically marks points on the map: it generates OpenLayers Feature objects one by one by parsing the JSON array returned by the front-end, and renders the camera points in real time on the VectorLayer of the electronic map;

[0107] The front-end management module allows for on-demand playback of video streams: clicking any camera icon on the front end sends an RTSP transcoding address to the back-end video gateway, using the camera's channel_code as a parameter. After obtaining the WebRTC stream, the video is automatically played in a pop-up window, enabling click-to-play functionality.

[0108] The information interaction and execution of readable programs between the modules in the above-mentioned device are based on the same concept as the method embodiments of the present invention, and the specific details can be found in the descriptions in the method embodiments of the present invention, and will not be repeated here.

[0109] Similarly, the advantages of the device of the present invention are:

[0110] (1) Interaction efficiency is improved by an order of magnitude. By combining a large language model with a local knowledge base, users only need to input a colloquial description to complete the entire chain of operations from semantic understanding to address parsing, spatial retrieval, map point marking, and video playback within 3 seconds. The average time taken is reduced from 45 seconds to 3 seconds, meeting the "second-level response" requirement for emergencies.

[0111] (2) The parsing accuracy has been greatly improved. By leveraging the capabilities of the large model in semantic understanding, reference resolution, and synonym expansion, end-to-end understanding of colloquial, inverted, and elliptical sentences is achieved. On the self-built test set, the address extraction accuracy reaches 98.7% and the distance extraction accuracy reaches 99.2%, which is significantly higher than the 70% of the traditional regularization scheme.

[0112] (3) Zero learning cost user experience. Users do not need to memorize professional information such as administrative divisions, road names, and camera numbers, nor do they need to master map drawing tools. They can retrieve target videos by describing them in everyday language, truly achieving "barrier-free" operation, which can cover multiple roles such as commanders, community grid workers, and field police officers.

[0113] (4) Dual protection of data security and link latency. The camera database, geocoding service, large language model and video gateway are all deployed on the local intranet to achieve "data not leaving the domain and model not going offline", which meets the mandatory requirement that video streams cannot be transmitted across the public network, and avoids the additional latency of 300-500ms caused by public network RTT.

[0114] It should be noted that not all steps and modules in the above processes and device structures are mandatory; some steps or modules can be omitted as needed. The execution order of each step is not fixed and can be adjusted as required. The system structure described in the above embodiments can be a physical structure or a logical structure. That is, some modules may be implemented by the same physical entity, or some modules may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.

[0115] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for fast retrieval and visualization of camera feeds based on natural language understanding, characterized by: include: Step S1: Build a full camera database: Create a new table named camera_info in the local PostgreSQL instance. Aggregate camera information from heterogeneous data sources into the camera_info table to form a full camera database. Step S2: Provide a camera space retrieval interface: Deploy a Spring Boot-based RESTful service on the local server, exposing the ` / api / camera / nearby` interface to receive HTTP-GET requests. Step S3: Build the map visualization front-end: Use Vue3 + OpenLayers framework as the front-end framework for electronic map loading and tile rendering; implement dynamic dot layers and support custom camera icons; build a video pop-up component to support automatic transcoding and playback from RTSP to WebRTC. Step S4: Provide address-coordinate conversion service: Integrate a third-party geocoding SDK through the local server to convert the natural language address entered by the user into latitude and longitude coordinate pairs. The conversion result is used for subsequent spatial retrieval. Step S5: Local private deployment: Deploy the camera spatial retrieval interface, front-end static resources, and large language model image together on the intranet GPU server; Step S6: Build a large model knowledge base: Create a new knowledge base table `knowledge_prompt` in the model mirror for writing data and building the large model knowledge base. Step S7: Receive the user's query request in natural language entered in the dialog box; Step S8: Perform semantic understanding and then call the interface to return data: Use the large language model to determine that the intent category is spatial retrieval, call the address-coordinate conversion service, parse the address, extract the distance, assemble the HTTP request and access the camera spatial retrieval interface to obtain a JSON array, and encapsulate the array into structured text and return it to the front end; Step S9: Perform automatic map point marking: Generate OpenLayers Feature objects one by one from the JSON array returned by the front end, and render the camera points in real time on the VectorLayer of the electronic map; Step S10: On-demand playback of video stream: Click any camera icon on the front end, and request the RTSP transcoding address from the backend video gateway with the camera's channel_code as the parameter. After obtaining the WebRTC stream, the video will automatically play in the pop-up window, enabling click-to-play functionality.

2. The method for fast retrieval and visualization of cameras based on natural language understanding according to claim 1, characterized in that when creating the camera_info table in step 1, the table structure includes at least the following fields: camera_name: The readable name of the camera; longitude: The longitude of the camera installation location, in degrees, rounded to six decimal places; latitude: The latitude of the camera installation location, in degrees, rounded to six decimal places; channel_code: The unique channel code of the camera in the video management platform, used for subsequent streaming.

3. The method for rapid retrieval and visualization of camera data based on natural language understanding as described in claim 2. Its characteristic is that when the camera space retrieval interface in step 2 receives an HTTP-GET request, the input parameters include: poiLon—POI: Longitude of the center point; poiLat—POI: Latitude of the center point; distance: Search radius, in meters; The interface internally uses PostGIS's ST_DWithin function to calculate spherical distance and returns a list of cameras within the distance range. Each element in the list must contain at least the four fields described in the camera_info table.

4. The method for rapid retrieval and visualization of camera data based on natural language understanding as described in claim 1, characterized in that: In step 6, when creating the knowledge base table `knowledge_prompt` in the model mirror, write the following content: Camera database field descriptions; The URL, request method, meaning of input parameters, and examples of the camera spatial retrieval interface; And examples of typical user questions.

5. A device for rapid retrieval and visualization of camera footage based on natural language understanding, characterized in that: It includes a full-database camera management module, an interface management module, a front-end management module, an address service module, a deployment module, a large-scale model knowledge base management module, a receiving module, and a semantic understanding module. The full camera database management module constructs a full camera database: A new table named `camera_info` is created in the local PostgreSQL instance, and camera information from heterogeneous data sources is aggregated into the `camera_info` table to form the full camera database. The interface management module provides a camera space retrieval interface: a Spring Boot-based RESTful service is deployed on the local server, exposing the ` / api / camera / nearby` interface to receive HTTP-GET requests. The front-end management module builds a map visualization front-end: it uses the Vue3 + OpenLayers framework as the front-end framework for loading electronic maps and rendering tiles; it implements dynamic dot layers, supports custom camera icons; and it builds a video pop-up component that supports automatic transcoding and playback from RTSP to WebRTC. The address service module provides address-coordinate conversion service: by integrating a third-party geocoding SDK on the local server, it converts the natural language address entered by the user into latitude and longitude coordinate pairs, and the conversion result is used for subsequent spatial retrieval; Deployment module local private deployment: Deploy the camera spatial retrieval interface, front-end static resources, and large language model image together on the internal network GPU server; The large model knowledge base management module constructs the large model knowledge base: A new knowledge base table, `knowledge_prompt`, is created in the model image for writing data and building the large model knowledge base. The receiving module receives the user's query request in natural language entered in the dialog box; The semantic understanding module performs semantic understanding and then calls the interface to return data: using the large language model to determine that the intent category is spatial retrieval, calling the address-coordinate conversion service, parsing the address, extracting the distance, assembling the HTTP request and accessing the camera spatial retrieval interface to obtain a JSON array, and encapsulating the array into structured text to return to the front end; The front-end management module automatically marks points on the map: it generates OpenLayers Feature objects one by one by parsing the JSON array returned by the front-end, and renders the camera points in real time on the VectorLayer of the electronic map; The front-end management module allows for on-demand playback of video streams: clicking any camera icon on the front end sends an RTSP transcoding address to the back-end video gateway, using the camera's channel_code as a parameter. After obtaining the WebRTC stream, the video is automatically played in a pop-up window, enabling click-to-play functionality.

6. The camera fast retrieval and visualization device based on natural language understanding according to claim 5, characterized in that: When creating a new camera_info table in the full camera database management module, the table structure must include at least the following fields: camera_name: The readable name of the camera; longitude: The longitude of the camera installation location, in degrees, rounded to six decimal places; latitude: The latitude of the camera installation location, in degrees, rounded to six decimal places; channel_code: The unique channel code of the camera in the video management platform, used for subsequent streaming.

7. A camera fast retrieval and visualization device based on natural language understanding according to claim 6, characterized in that the interface... When the management module receives an HTTP-GET request at the camera space retrieval interface, the input parameters include: poiLon—POI: Longitude of the center point; poiLat—POI: Latitude of the center point; distance: Search radius, in meters; The interface internally uses PostGIS's ST_DWithin function to calculate spherical distance and returns a list of cameras within the distance range. Each element in the list must contain at least the four fields described in the camera_info table.

8. The camera fast retrieval and visualization device based on natural language understanding according to claim 5, characterized in that when the large model knowledge base management module creates a new knowledge base table knowledge_prompt in the model image, it writes the following content: Camera database field descriptions; The URL, request method, meaning of input parameters, and examples of the camera spatial retrieval interface; And examples of typical user questions.

Images