A large model and recorder-based automatic identification method for inspection videos

Abstract

The application relates to the field of inspection, in particular to a property inspection video automatic identification method based on a large model and a recorder, which comprises the following steps: one, providing a property inspection video automatic identification system based on a large model and a recorder; two, data acquisition and preprocessing; three, content analysis; four, data standardization processing; five, intelligent judgment of inspection items; six, standardized report generation and storage; through a double-large-model collaborative architecture, the video multi-modal understanding advantage of qwen3-vl-plus and the logical reasoning capability of qwen-plus are exerted, end-to-end automatic processing from video to report is realized, the inspection efficiency is greatly improved, and the traditional inspection process of several hours is shortened to seconds. Structured data output solves the problems that traditional inspection records are unreliable and difficult to trace, the standardized report is convenient for manual checking and supports machine analysis, and the large model flexible reasoning capability can adapt to the differentiated needs of different scenes such as parks and commercial complexes.

Description

Technical Field

[0001] This invention relates to the field of inspection, specifically to an automatic identification method for inspection videos based on a large model and a recorder. Background Technology

[0002] In the current property inspection field, the traditional model mainly relies on manual, timed, and fixed-point recording, depending on on-site observation by staff and the completion of paper forms. With the development of AI technology, some companies have introduced single AI video analysis systems, but these solutions still have significant shortcomings. The main deficiencies are as follows: Unable to objectively verify "person's arrival": Currently, there is a lack of objective, traceable, and non-repudiable technical verification methods to determine whether security personnel have truly completed their patrol tasks according to the prescribed time, route, and frequency. Mainstream methods, such as manual sign-in, paper records, or patrol baton contact sensors, are easily forged, proxy clocking in, or retroactive recording.

[0003] There is no effective way to prove that "the task has been completed": Even if personnel arrive on time, it cannot be guaranteed that all pre-set inspection items have been actually and completely performed. For example, the inspection checklist may only indicate "inspect fire hydrants," but there is no evidence to show whether the patrolman opened the cabinet door, checked the hoses, or confirmed that the pressure gauge pointer was within the normal range.

[0004] Unstructured data format: The analysis results are mostly unstructured text or simple tags, lacking a unified and standardized structured output, making it difficult to support subsequent data traceability, statistical analysis and machine secondary processing.

[0005] Poor scenario adaptability: Inspection rules are mostly based on preset fixed logic and lack intelligent reasoning capabilities, which cannot meet the differentiated inspection needs of different parks and different facility types.

[0006] Therefore, it is essential to propose an automatic identification method for inspection videos based on large models and recorders. Summary of the Invention

[0007] The purpose of this invention is to provide an automatic identification method for inspection videos based on a large model and a recorder, which makes inspections more efficient, inspection records traceable, and has better adaptability, in order to solve the existing technical defects and unmet technical requirements.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an automatic identification method for inspection videos based on a large model and a recorder, comprising the following steps: I. A property inspection video automatic recognition system based on a large model and a recorder is provided, including a law enforcement recorder, a six-axis gyroscope, a Bluetooth beacon, a video content analysis model, a data standardization processing module, an inference and judgment model, and a storage module (database). The law enforcement recorder includes at least a high-definition camera, a six-axis gyroscope, and a Bluetooth communication module. The high-definition camera is electrically connected to the six-axis gyroscope. The law enforcement recorder establishes a signal connection with the Bluetooth beacon through the Bluetooth communication module. The Bluetooth beacon is deployed in each building floor (within the inspection area). The law enforcement recorder, video content analysis model, data standardization processing module, inference and judgment model, and storage module (database) are electrically connected in sequence. II. Data Acquisition and Preprocessing 2.1) Acquire inspection videos using a high-definition camera assisted by Bluetooth beacons and a six-axis gyroscope; 2.2) Send the preprocessed video to the video content analysis model; This application achieves dual spatial and temporal positioning through the linkage of law enforcement recorders and Bluetooth beacons, and generates an immutable digital evidence chain by combining high-fidelity video analysis, fundamentally solving the problems of the patrol process being invisible and the responsibility being untraceable, and objectively verifying that "the person has arrived" and "the matter has been completed".

[0009] Furthermore, the overall content of step two can be simply understood as obtaining patrol videos through the patrol recorder. In terms of specific logic, it can be understood as follows: Bluetooth beacons are deployed in the patrol area. When the law enforcement recorder approaches the Bluetooth beacon (e.g., within 10 meters), it will trigger a video writing operation, synchronously recording the current geographical location of the Bluetooth beacon into the video file.

[0010] When the law enforcement recorder is collecting video, it will simultaneously read the status of the six-axis gyroscope to determine whether the person is walking or stationary, thereby controlling the frame rate of the video recording.

[0011] III. Content Analysis 3.1) Receive pre-processed inspection videos through a video content analysis model; 3.2) Then, the inspection video is fused and analyzed to generate a structured intermediate report; 3.3) Send the structured intermediate report to the data standardization processing module; IV. Data Standardization Processing 4.1) Receive structured intermediate reports through the data standardization processing module; 4.2) Then convert the structured intermediate report into a structured dataset; 4.3) Send the structured dataset to the inference and judgment model; V. Intelligent Judgment of Inspection Items 5.1) Obtain the inspection list matching the current building-floor in the storage module through the reasoning and judgment model, and receive the structured dataset; 5.2) The structured dataset is compared item by item to intelligently determine whether each inspection item meets the specifications and identify abnormal items; 5.3) Output the intelligent judgment results of the inspection items and send them to the report generation module; In step 5.1), it needs to be explained that the inspection list is preset in the database. The reasoning and judgment model dynamically loads the corresponding inspection list from the database based on the geographical location of the current video recording (obtained from Bluetooth beacons).

[0012] VI. Standardized Report Generation and Storage 6.1) The report generation module receives the intelligent judgment results of the inspection items; 6.2) Then, based on the preset template, the intelligent judgment results of the inspection items are integrated into a standardized inspection report; 6.3) The standardized inspection report is sent to the storage module (database) to form a complete chain of evidence that is traceable, verifiable, and machine-analyzable.

[0013] This application presents a property inspection video analysis method based on the collaboration of two major models, qwen3-vl-plus and qwen-plus. Through the organic integration of multimodal perception, structured processing and intelligent reasoning, it achieves dual objective verification of "whether the person has arrived" and "whether the task has been completed" during the inspection process.

[0014] Preferably, the specific content of step 2.1) includes: 2.1.1) By deploying Bluetooth beacons on each floor of the building, automatic spatial segmentation of the video stream in the inspection video is achieved; 2.1.2) The user's state is determined by a six-axis gyroscope, and the video recording parameters are adjusted accordingly; 2.1.3) Generate an inspection video with embedded timestamp watermark, corresponding location tag and user status; that is, the inspection video contains timestamp, motion status and spatial location information; Preferably, step 3.2) specifically includes: The video content analysis model uses Interleaved-MRoPE technology to fuse and analyze the visual, temporal, and spatial three-dimensional information in the preprocessed video. At the same time, DeepStack technology is introduced to integrate ViT multi-level features, which enhances the ability to capture fine-grained visual details and align text with images. It supports OCR recognition of text information such as device identification and parameter scales in the video, and generates a structured intermediate report containing a device list, status description (summary), timestamp, and building-floor identification. The structured intermediate report is in JSON format.

[0015] This application outputs standardized JSON format content reports and standardized inspection reports to ensure that the data is traceable, parsable, and statistically analyzeable. Preferably, the specific content of step 4.2) is as follows: The data standardization processing module cleans, standardizes terminology, and extracts features from the structured intermediate reports. By establishing a standardized terminology mapping table, it accurately matches non-standard expressions such as device status descriptions and location information with preset standardized terms, transforming the structured intermediate reports into a structured dataset with consistent format and semantic standardization to meet subsequent logical judgment requirements.

[0016] Preferably, in step 5.1), the inspection items included in the inspection checklist include: Fire protection facilities: fire extinguisher pressure value, fire hydrant box integrity, emergency lighting status; Safety facilities: status of access control system, working status of surveillance cameras, and unobstructed access to evacuation routes; Electrical equipment: operating status of distribution boxes, appearance of wires and cables, and working condition of lighting systems; Other facilities: elevator operation status, water supply and drainage system, building exterior.

[0017] Preferably, the specific content of step 5.2) is as follows: 5.2.1) Compare the structured dataset (i.e. the equipment description in the inspection video) with the corresponding inspection items in the inspection list one by one to determine whether each inspection item in the inspection video meets the preset standard of the corresponding inspection item in the inspection list. This application allows for customized inspection item configurations to adapt to the differentiated inspection needs of various scenarios such as industrial parks and commercial complexes.

[0018] 5.2.2) If all comparison results meet the corresponding preset standards in the inspection checklist, the inspection process is deemed to meet the requirements. 5.2.3) If it is determined that there are discrepancies between the structured dataset and the preset standards in the inspection checklist, then the missing, incorrect, or non-compliant operations must be pointed out. 5.2.4) Finally, give the overall conclusion: whether the inspection process meets the requirements ("meets" / "does not meet"). If it does not meet the requirements, the relevant reasons for the non-compliance should be output in text form, and the text content should not exceed 100 words.

[0019] In this application, it is necessary to further explain that: in step 5.2.1), the inspection checklist can be flexibly configured through the back-end system, and inspection items for personnel operations can be added, such as: opening the fire hydrant cabinet door for inspection, and filling out and updating the fire hydrant cabinet equipment maintenance card.

[0020] In step 5.2.2), the inspection process includes three levels of checks: first, determining whether the inspection personnel have reached each inspection point (which can be determined by Bluetooth beacons); second, determining whether the inspection personnel have performed the inspection actions as required; and third, assisting the inspection personnel in discovering equipment abnormalities during the inspection process.

[0021] Step 5.2.3) includes both equipment inspection and assessment of whether the inspection personnel are performing the inspection in a standardized manner. This can be achieved simply by configuring the operational requirements for personnel in the inspection checklist.

[0022] Preferably, in step 6.2), the standardized inspection report includes basic inspection information, a detailed list of problems, statistics of qualified items, overall conclusions, and explanations of abnormal situations. The report will also include a conclusion on whether the inspection process meets the requirements.

[0023] Preferably, the video content analysis model is qwen3-vl-plus, the inference and judgment model is qwen-plus, and the inference and judgment model supports a context length of 128k and an output length of 8k.

[0024] This application leverages the synergistic advantages of two major models, employing a video understanding model and a logical reasoning model to work together. The former is responsible for multimodal video content parsing, while the latter performs compliance judgments on inspection items. This achieves a precise combination of video visual detail extraction and efficient logical judgment of inspection items, reducing the rate of missed detections and false judgments, improving the accuracy of analysis, and constructing an end-to-end workflow from video input to report output. The entire process of content analysis, item judgment, and report generation can be completed without manual intervention, achieving full automation.

[0025] Preferably, in step 2.1.2), when a person is detected to be stationary (indicating that a detailed equipment inspection is being carried out), recording is performed at a high frame rate (greater than or equal to 15 fps) to ensure that key images are clear, thereby optimizing data efficiency while ensuring the clarity of key details. When personnel are detected walking (non-inspection state), automatically switch to low frame rate (greater than 5 fps) recording to reduce redundant data; In step 2.1.3), the inspection video is in MP4 format with a resolution of 1920×1080.

[0026] Preferably, in step 6.2): Basic inspection information includes: time, building, floor, and inspection personnel identification; The overall conclusion includes: qualified, unqualified, or partially requiring review; A detailed list of issues includes: device ID, issue description, time and location of occurrence; The statistics for qualified items include: quantity and percentage; Abnormal situations include: insufficient image quality and lack of device coverage. In this application, the description of abnormal situations does not conflict with the problem description. The problem description refers to the abnormalities found during the inspection, such as the fire extinguisher pressure being too low. The abnormal situations refer to unexpected situations that occur during the inspection, such as the law enforcement recorder having too low a battery, resulting in blurry video and inaccurate final identification results.

[0027] Compared with the prior art, the beneficial effects of the present invention are: 1. This application, through a dual-model collaborative architecture, fully leverages the video multimodal understanding advantages of qwen3-vl-plus and the logical reasoning capabilities of qwen-plus, achieving end-to-end automated processing from video to report. This significantly improves inspection efficiency, reducing the traditional inspection process of several hours to seconds. Structured data output solves the problems of unreliable and difficult-to-trace traditional inspection records. Standardized reports are easy for manual viewing and also support machine parsing, providing support for digital management of property inspections. The flexible reasoning capabilities of the large models enable the system to adapt to the differentiated needs of different scenarios such as industrial parks and commercial complexes, enhancing the scalability and practicality of the solution.

[0028] 2. This application uses a law enforcement recorder as the data acquisition terminal, combined with Bluetooth beacons to achieve automatic building-floor positioning, and uses a gyroscope to dynamically adjust the frame rate. While ensuring the accuracy of key checkpoint analysis, it significantly reduces storage and computing overhead, further improving the system's practicality and deployment economy. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall process of the present invention; Detailed Implementation The following will refer to the appendices in the embodiments of the present invention. Figure 1The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0030] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.

[0031] Please see Figure 1 Embodiments of the present invention: Example: like Figure 1 As shown: An automatic identification method for inspection videos based on a large model and a recorder, including the following steps: I. A property inspection video automatic recognition system based on a large model and a recorder is provided, including a law enforcement recorder, a six-axis gyroscope, a Bluetooth beacon, a video content analysis model, a data standardization processing module, an inference and judgment model, and a storage module (database). The law enforcement recorder includes at least a high-definition camera, a six-axis gyroscope, and a Bluetooth communication module. The high-definition camera is electrically connected to the six-axis gyroscope. The law enforcement recorder establishes a signal connection with the Bluetooth beacon through the Bluetooth communication module. The Bluetooth beacon is deployed in each building floor (within the inspection area). The law enforcement recorder, video content analysis model, data standardization processing module, inference and judgment model, and storage module (database) are electrically connected in sequence. The video content analysis model is specifically qwen3-vl-plus, and the reasoning and judgment model is specifically qwen-plus. The reasoning and judgment model supports a 128k context length and an 8k output length, and has powerful capabilities for understanding complex instructions and logical reasoning.

[0032] II. Data Acquisition and Preprocessing 2.1) Acquire inspection videos using a high-definition camera assisted by Bluetooth beacons and a six-axis gyroscope; 2.1.1) Automatic spatial segmentation of video streams in inspection videos is performed by deploying Bluetooth beacons on each floor of the building; In this embodiment, property management staff use law enforcement recorders to collect patrol videos, which can simultaneously record video, movement status, and spatial location information during the patrol process.

[0033] Bluetooth beacons are pre-deployed in key areas on different floors of each building (such as elevator lobbies, stairwells, and equipment room entrances). When the body camera passes within the beacon's coverage area, it automatically receives the broadcast signal, parses the corresponding building number and floor identifier, and associates the location metadata with the currently recorded video stream in real time. Based on this, the system automatically segments the original video along the "building-floor" dimension, generating sub-video segments with clear spatial attribution. This mechanism ensures that each frame of video can be traced back to a specific physical location and precise time, providing objective and irrefutable technical evidence of whether a person has arrived at a designated patrol point. Simultaneously, the body camera uses a gyroscope to determine the wearer's status in real time. 2.1.2) The user's state is determined by a six-axis gyroscope, and the video recording parameters are adjusted accordingly; When a person is detected to be stationary (indicating that a detailed equipment inspection is in progress), record at a high frame rate (15 fps) to ensure that key shots are clear; When personnel are detected walking (not in inspection mode), automatically switch to low frame rate (1 fps) recording to reduce redundant data; 2.1.3) Generate inspection videos with embedded timestamp watermarks and corresponding location tags. Each video segment has an embedded timestamp watermark and corresponding location tag (e.g., building_id="B", floor="5").

[0034] The inspection video is in MP4 format with a resolution of 1920×1080.

[0035] 2.2) Send the preprocessed video to the video content analysis model; III. Content Analysis 3.1) The preprocessed inspection video is received through the video content analysis model, and dense frame sampling is performed according to the configured parameters (1 frame is sampled every 15 frames, and the model is called once every 60 sampled frames); at the same time, the system prompts given to the model are as follows: You are a professional video content structured analysis engine. Based on the video content description provided by the user (which may include time segments, scene details, character behavior, etc.), please generate a **highly structured video content report that retains the original details**, and output it in **strict JSON format**, **without containing any explanations, comments, Markdown, code block identifiers, or additional text**.

[0036] 3.2) Then, the inspection video is fused and analyzed to generate a structured intermediate report; The video content analysis model uses Interleaved-MRoPE technology to fuse and analyze the visual, temporal, and spatial three-dimensional information in the preprocessed video. At the same time, DeepStack technology is introduced to integrate ViT multi-level features, which enhances the ability to capture fine-grained visual details and align text with images. It supports OCR recognition of text information such as device identification and parameter scales in the video, and generates a structured intermediate report containing a device list, status description (summary), timestamps (segments), and building-floor identification. The time information in the structured intermediate report is extracted from the video timestamp watermark in JSON format.

[0037] The output JSON object must contain the following fields: - `summary` (string): Use a sentence to summarize the overall theme, purpose, and flow of the video.

[0038] - `segments` (array): Video segments divided in chronological order. Each segment is an object containing: - `start_time`(string, format `"yyyy-mm-dd hh:mm:ss"`) - `end_time`(string, format `"yyyy-mm-dd hh:mm:ss"`) - `title` (string): The subtitle of this section (e.g., "Fire Safety Facilities Inspection on the 7th Floor") - `visual_description` (string): Describes the scene content in detail, including character features (clothing, actions), environment (floors, signs, equipment), items (documents, tools, signs), etc., preserving as much original detail as possible.

[0039] **Require:** - All timestamps are extracted from the video timestamp watermark and formatted as `"yyyy-mm-dd hh:mm:ss"`.

[0040] - `visual_description` should reproduce the details in the original description as closely as possible, with particular attention to the following items: curtains, and should not be simplified or generalized.

[0041] - If audio is not mentioned in the input, `audio_or_action_summary` should focus on observable behavior.

[0042] - **The output must be valid, directly parsable JSON, and contain only JSON content.** The final generated structured JSON report expands upon the original structure with additional fields, as shown in the example below: json { Summary: This video documents a routine safety inspection conducted in a modern office building. The inspection team started from the 4th floor and checked the elevator operation, fire-fighting facilities on the 7th and 8th floors, office areas, and the conference center in sequence, finally completing the inspection process. "segments": [ { "start_time": "yyyy-mm-dd hh:mm:ss", "end_time": "yyyy-mm-dd hh:mm:ss", "title": "Patrol Preparation and Departure", "visual_description": "The video begins in the elevator lobby on the 4th floor. A stainless steel elevator door is visible in the frame, and the external display shows the current floor as '4' with a downward arrow. The inspector is holding a mobile phone and a paper document (later confirmed to be a 'Patrol Sign-in Sheet'). The elevator descends to the 1st floor, and the inspector enters and ascends to the 2nd floor." "location_tag": "Building 1 - 7th Floor" }, ... ] } ... It should be noted that the above content is only a basic template for the model's runtime; the "*" and "..." indicate different elements. In this context, "*" is a Markdown (Lightweight Markup Language) symbol used to emphasize content in the middle, and "..." has the same meaning as the Chinese ellipsis. Here, it indicates that there will be multiple video parsing segments with different content but the same format.

[0043] 3.3) Send the structured intermediate report to the data standardization processing module; IV. Data Standardization Processing 4.1) Receive structured intermediate reports through the data standardization processing module; 4.2) Then convert the structured intermediate report into a structured dataset; The data standardization processing module cleans, standardizes terminology, and extracts features from the structured intermediate reports. By establishing a standardized terminology mapping table, it accurately matches non-standard expressions such as device status descriptions and location information with preset standardized terms, transforming the structured intermediate reports into a structured dataset with consistent format and semantic standardization.

[0044] This involves establishing a standardized terminology mapping table to accurately match non-standard expressions such as equipment status descriptions and location information with preset standardized terms (for example, mapping "door closed" and "door is closed" to "box door is closed normally"), thereby achieving consistency and accuracy in information expression and providing a standardized foundation for subsequent data processing, analysis, and management.

[0045] 4.3) Send the structured dataset to the inference and judgment model; V. Intelligent Judgment of Inspection Items 5.1) Obtain the inspection list matching the current building-floor in the storage module through the reasoning and judgment model, and receive the structured dataset; The inspection checklist includes the following items: Fire protection facilities: fire extinguisher pressure value, fire hydrant box integrity, emergency lighting status; Safety facilities: status of access control system, working status of surveillance cameras, and unobstructed access to evacuation routes; Electrical equipment: operating status of distribution boxes, appearance of wires and cables, and working condition of lighting systems; Other facilities: elevator operation status, water supply and drainage system, building exterior.

[0046] Different buildings or floors can be configured with differentiated lists according to actual needs (e.g., drainage pumps need to be checked in underground garages, and fire doors need to be checked in high-rise office areas). The system automatically loads the applicable rules for the corresponding area based on the building_id and floor in the video clip.

[0047] 5.2) The structured dataset is compared item by item to intelligently determine whether each inspection item meets the specifications and identify abnormal items; 5.2.1) Compare the structured dataset with the corresponding inspection items in the inspection checklist one by one to determine whether each inspection item in the inspection video meets the preset standard of the corresponding inspection item in the inspection checklist; 5.2.2) If all comparison results meet the corresponding preset standards in the inspection checklist, the inspection process is deemed to meet the requirements. 5.2.3) If it is determined that there are discrepancies between the structured dataset and the preset standards in the inspection checklist, then the missing, incorrect, or non-compliant operations must be pointed out. 5.2.4) Finally, give the overall conclusion: whether the inspection process meets the requirements ("meets" / "does not meet"). If it does not meet the requirements, the relevant reasons for the non-compliance should be output in text form, and the text content should not exceed 100 words.

[0048] 5.3) Output the intelligent judgment results of the inspection items and send them to the report generation module; The report template is designed to clearly require the model to compare each item with the structured video data based on the pre-set inspection checklist, determine whether it meets the requirements, mark the location of the problem, describe the content of the problem, and give an overall conclusion.

[0049] VI. Standardized Report Generation and Storage 6.1) The report generation module receives the intelligent judgment results of the inspection items; 6.2) Then, based on the preset template, the intelligent judgment results of the inspection items are integrated into a standardized inspection report; The standardized inspection report includes basic inspection information, a detailed list of problems, statistics on qualified items, overall conclusions, and explanations of any abnormal situations.

[0050] Basic inspection information includes: time, building, floor, and inspection personnel identification; The overall conclusion includes: qualified, unqualified, or partially requiring review; A detailed list of issues includes: device ID, issue description, time and location of occurrence; The statistics for qualified items include: quantity and percentage; Abnormal situations include: insufficient image quality and lack of device coverage.

[0051] 6.3) Standardized inspection reports are transmitted to the storage module (database), supporting querying, statistics, and visualization analysis by time, region, facility type, and other dimensions. This report can not only be reviewed by management personnel but also serve as a digital credential of performance, providing objective evidence in safety incident investigations or liability determinations, forming a complete "personnel present, task completed" closed loop.

[0052] In this embodiment, specifically, after receiving the intelligent judgment result of the inspection item, the report generation module simultaneously provides the model system with the following prompts: You are a professional inspection and auditing officer. Please analyze and judge based on the following two parts: 1. **Inspection Checklist** (i.e., Standard Operating Procedures): {{checklist}} 2. **Textual description of the inspection video** (i.e., a transcription or summary of the actual execution process): {{answer}} Your task is: - Compare the inspection checklist with the video and text descriptions item by item; - Determine whether each item has been executed correctly and completely; - Identify missing, incorrect, or non-compliant operations; - The final overall conclusion is: whether the inspection process meets the requirements ("Meets" / "Does not meet"); - If not, please briefly explain the main reasons.

[0053] **Output format requirements:** json { "item_checks": [ { "check_item": "Specific items in the inspection checklist", "executed": "Complies" or "Does not comply", "evidence": "Key statements in the video description that support or do not support this entry", "datetime": "yyyy-mm-dd hh:mm:ss - yyyy-mm-dd hh:mm:ss, corresponding time information" }, / / ... other entries ], "overall_compliance": "Satisfied" or "Not satisfied", "summary": A brief summary of the overall assessment, not exceeding 100 words. } ``` Please make rigorous judgments based on facts and do not speculate on content not mentioned.

[0054] After receiving standardized video description data and inspection checklists, the model first precisely analyzes each requirement in the checklist to clarify the inspection standards and key judgment points. Then, it matches the equipment status, environmental characteristics, and personnel operations information in the video description data with each checklist item. For example, for the inspection item "fire extinguisher pressure value normal," the model searches the video description for explicit statements such as "fire extinguisher pressure gauge pointer is in the green zone" or "pressure is normal," or uses indirect information such as "fire extinguisher appearance is intact, no abnormalities found" for auxiliary inference. If the video description does not mention the pressure status of the fire extinguisher, or contains contradictory descriptions such as "pressure gauge pointer points to the red zone," it is judged as "non-compliant," and the corresponding video description statements and timestamps are recorded as evidence. For inspection items involving operational procedures, such as "evacuation route unobstructed," the model focuses on analyzing whether there are compliant descriptions in the video such as "the route is free of debris" or "fire doors are closed but not locked," or non-compliant situations such as "the route is blocked by goods" or "fire doors open at too large an angle," combining this with time information to form a complete judgment basis. Through this step-by-step, meticulous comparison and reasoning, the model can accurately identify compliance and anomalies during the inspection process, laying a solid foundation for generating the final report.

[0055] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0056] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for automatic identification of inspection videos based on a large model and a recorder, characterized in that, Includes the following steps: I. A property inspection video automatic recognition system based on a large model and a recorder is provided, including a law enforcement recorder, a six-axis gyroscope, a Bluetooth beacon, a video content analysis model, a data standardization processing module, an inference and judgment model, and a storage module. The law enforcement recorder includes at least a high-definition camera and a Bluetooth communication module. The high-definition camera and the six-axis gyroscope are electrically connected. The law enforcement recorder establishes a signal connection with the Bluetooth beacon through the Bluetooth communication module. The Bluetooth beacon is deployed in each floor of the building. The law enforcement recorder, the video content analysis model, the data standardization processing module, the inference and judgment model, and the storage module are electrically connected in sequence. II. Data Acquisition and Preprocessing 2.1) Acquire inspection videos using a high-definition camera assisted by Bluetooth beacons and a six-axis gyroscope; 2.2) Send the preprocessed video to the video content analysis model; III. Content Analysis 3.1) Receive pre-processed inspection videos through a video content analysis model; 3.2) Then, the inspection video is fused and analyzed to generate a structured intermediate report; 3.3) Send the structured intermediate report to the data standardization processing module; IV. Data Standardization Processing 4.1) Receive structured intermediate reports through the data standardization processing module; 4.2) Then convert the structured intermediate report into a structured dataset; 4.3) Send the structured dataset to the inference and judgment model; V. Intelligent Judgment of Inspection Items 5.1) Obtain the inspection list matching the current building-floor in the storage module through the reasoning and judgment model, and receive the structured dataset; 5.2) The structured dataset is compared item by item to intelligently determine whether each inspection item meets the specifications and identify abnormal items; 5.3) Output the intelligent judgment results of the inspection items and send them to the report generation module; VI. Standardized Report Generation and Storage 6.1) The report generation module receives the intelligent judgment results of the inspection items; 6.2) Then, based on the preset template, the intelligent judgment results of the inspection items are integrated into a standardized inspection report; 6.3) Send the standardized inspection report to the storage module.

2. The automatic identification method for inspection videos based on a large model and a recorder according to claim 1, characterized in that the specific content of step 2.1) includes: 2.1.1) By deploying Bluetooth beacons on each floor of the building, automatic spatial segmentation of the video stream in the inspection video is achieved; 2.1.2) The user's state is determined by a six-axis gyroscope, and the video recording parameters are adjusted accordingly; 2.1.3) Generate inspection videos with embedded timestamp watermarks, corresponding location tags, and user status.

3. The automatic identification method for inspection videos based on a large model and a recorder according to claim 1, characterized in that, In step 3.2), the specific content is as follows: The video content analysis model uses Interleaved-MRoPE technology to fuse and analyze the visual, temporal, and spatial three-dimensional information in the preprocessed video. At the same time, DeepStack technology is introduced to integrate ViT multi-level features, which enhances the ability to capture fine-grained visual details and align text with images. It supports OCR recognition of text information such as device identification and parameter scales in the video, and generates a structured intermediate report containing a device list, status description, timestamp, and building-floor identification. The structured intermediate report is in JSON format.

4. The automatic identification method for inspection videos based on a large model and a recorder according to claim 1, 2, or 3, characterized in that, The specific content of step 4.2) is as follows: The data standardization processing module cleans, standardizes terminology, and extracts features from the structured intermediate reports. By establishing a standardized terminology mapping table, it accurately matches non-standard expressions such as device status descriptions and location information with preset standardized terms, transforming the structured intermediate reports into a structured dataset with consistent format and semantic standardization.

5. The automatic identification method for inspection videos based on a large model and a recorder according to claim 4, characterized in that, In step 5.1), the inspection items included in the inspection checklist are: Fire protection facilities: fire extinguisher pressure value, fire hydrant box integrity, emergency lighting status; Safety facilities: status of access control system, working status of surveillance cameras, and unobstructed access to evacuation routes; Electrical equipment: operating status of distribution boxes, appearance of wires and cables, and working condition of lighting systems; Other facilities: elevator operation status, water supply and drainage system, building exterior.

6. The automatic identification method for inspection videos based on a large model and a recorder according to claim 5, characterized in that, The specific content of step 5.2) is as follows: 5.2.1) Compare the structured dataset with the corresponding inspection items in the inspection checklist one by one to determine whether each inspection item in the inspection video meets the preset standard of the corresponding inspection item in the inspection checklist; 5.2.2) If all comparison results meet the corresponding preset standards in the inspection checklist, the inspection process is deemed to meet the requirements. 5.2.3) If it is determined that there are discrepancies between the structured dataset and the preset standards in the inspection checklist, then the missing, incorrect, or non-compliant operations must be pointed out. 5.2.4) Finally, give the overall conclusion: whether the inspection process meets the requirements. If not, the relevant reasons for the non-compliance should be output in text form, and the text content should not exceed 100 words.

7. A method for automatic identification of inspection videos based on a large model and a recorder, as described in claim 1, 2, 3, 5, or 6, characterized in that, In step 6.2), the standardized inspection report includes basic inspection information, a detailed list of problems, statistics of qualified items, overall conclusions, and explanations of abnormal situations.

8. The automatic identification method for inspection videos based on a large model and a recorder according to claim 7, characterized in that, The video content analysis model is specifically qwen3-vl-plus, and the inference and judgment model is specifically qwen-plus. The inference and judgment model supports a context length of 128k and an output length of 8k.

9. A method for automatic identification of inspection videos based on a large model and a recorder, as described in claim 2, 3, 5, 6, or 8, characterized in that, In step 2.1.2), when a person is detected to be stationary, recording is performed at a high frame rate of 15fps or higher. When people are detected walking, the recording automatically switches to a low frame rate of less than 5fps. In step 2.1.3), the inspection video is in MP4 format with a resolution of 1920×1080.

10. The automatic identification method for inspection videos based on a large model and a recorder according to claim 7, characterized in that, In step 6.2): Basic inspection information includes: time, building, floor, and inspection personnel identification; The overall conclusion includes: qualified, unqualified, or partially requiring review; A detailed list of issues includes: device ID, issue description, time and location of occurrence; The statistics for qualified items include: quantity and percentage; Abnormal situations include: insufficient image quality and lack of device coverage.

Images