A multi-modal research and judgment result credibility reuse method in unmanned aerial vehicle inspection

Abstract

The application discloses a method for credible reuse of multi-modal research and judgment results in unmanned aerial vehicle (UAV) inspection, and relates to the technical fields of UAV inspection and multi-modal intelligent research and judgment. When the same area is repeatedly inspected by the UAV, candidate historical research and judgment record packages are first screened through area coordinate keys, and then current frames or current short-time clips are collected, light check features are extracted, current area state fingerprints are generated, and consistency comparison is performed with historical area state fingerprints in the reused record packages. Subsequently, the reuse credibility is calculated in combination with area consistency, state fingerprint similarity, time attenuation, hidden danger type risk weight, model version compatibility and sensor combination consistency. Only when the reuse credibility reaches the corresponding threshold value, the static research and judgment layer or the dynamic research and judgment layer result is allowed to be reused. The application converts the ordinary candidate record hits into credible reuse judgments which are verified by the current site state, reduces the calculation amount of repeated multi-modal reasoning, and avoids the direct reuse of outdated or untrustworthy safety inspection conclusions.

Description

Technical Field

[0001] This invention relates to the fields of UAV inspection and multimodal intelligent analysis technology, and in particular to a reliable reuse method for multimodal analysis results in UAV inspection. Background Technology

[0002] Drone inspections have been widely applied in scenarios such as building construction, power maintenance, port construction, and bridge construction. Drones are typically equipped with multimodal sensors such as cameras, infrared sensors, or lidar to identify potential hazards and assess safety at construction sites, such as unprotected foundation pits, personnel not wearing safety helmets, workers not wearing safety belts at heights, vehicles driving in the wrong direction, and abnormal smoke or fire.

[0003] In actual inspections, due to factors such as inspection route planning, key area re-inspection, flight attitude adjustment, and on-site obstructions, drones may pass through the same construction area multiple times in a short period. If a complete multimodal model inference is performed each time, it will increase the computing power consumption of the drone or edge device, reducing inspection efficiency.

[0004] Existing methods for temporarily storing historical data typically rely on the existence of candidate records, whether the reused record has expired, or the popularity of the stored data as the primary criteria for data reuse. While these methods are suitable for content distribution or data response scenarios, for construction safety inspections, the reuse of historical assessment results depends not only on the existence of historical records but also on whether the current site conditions are consistent with historical conditions. If historical assessment results are directly reused simply because a candidate record is found, unreliable safety conclusions may be drawn when dynamic targets such as personnel, vehicles, machinery, and smoke / fire effects change.

[0005] Therefore, a reliable reuse method is needed for UAV safety inspection scenarios, so that historical multimodal judgment results can only be reused after passing the current on-site lightweight verification, thereby balancing the reduction of repetitive inference computation and the reliability of safety judgment. Summary of the Invention

[0006] The purpose of this invention is to provide a reliable reuse method for multimodal judgment results in UAV inspection. This method separates candidate record hits from judgment result reuse. A candidate record hit only indicates the existence of candidate historical judgment record packets; whether reuse is allowed must also be determined through lightweight consistency verification of the current frame, reuse reliability calculation, and static and dynamic hierarchical reuse judgment.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for reliable reuse of multimodal assessment results in UAV inspection, comprising the following steps: Step S1: The UAV acquires the GPS coordinates of the current inspection area during the inspection process; when the GPS coordinates cannot be acquired or the GPS signal is lost, a disallowed reuse state is generated, and the reuse of historical judgment results is skipped, and real-time multimodal judgment is performed; when the GPS coordinates can be acquired, the longitude and latitude in the GPS coordinates are truncated or rounded to generate a regional coordinate key for screening candidate judgment record packages. Step S2: Query the local trusted reuse record library using the area coordinate key to obtain the candidate historical assessment record package corresponding to the current inspection area; the candidate historical assessment record package includes at least the area coordinate key, historical multimodal assessment results, result generation timestamp, hidden danger type label, result trusted validity period, historical area status fingerprint, model version identifier, and sensor combination identifier; Step S3: Collect the current frame or current short segment of the current inspection area, extract the lightweight verification features of the current frame, and generate the current area status fingerprint based on the lightweight verification features of the current frame; the lightweight verification features of the current frame include at least one of image perception hash, edge contour features, target quantity vector, hidden danger area location summary, infrared temperature range or lidar contour summary, and fuse the lightweight verification features according to preset weights to generate a comprehensive current area status fingerprint; Step S4: Compare the current region status fingerprint with the historical region status fingerprint in the candidate historical judgment record package to obtain the status fingerprint similarity; and calculate the reuse credibility of the candidate historical judgment record package based on the consistency of region coordinate keys, status fingerprint similarity, time decay between the current system time and the result generation timestamp, risk weight of hidden danger type, model version compatibility and sensor combination consistency. Step S5: Divide the historical multimodal assessment results into a static assessment layer and a dynamic assessment layer according to the hazard type label; when the reuse credibility reaches the static assessment layer reuse threshold, the static assessment layer results are allowed to be reused; when the number of dynamic targets, the dynamic target location summary, and the dynamic target status summary corresponding to the dynamic assessment layer are consistent with the lightweight verification features of the current frame, and the reuse credibility reaches the dynamic assessment layer reuse threshold which is higher than the static assessment layer reuse threshold, the dynamic assessment layer results are allowed to be reused. Step S6: When any of the following situations occur: no candidate historical judgment record package is found, GPS signal is lost, model version is incompatible, sensor combination is inconsistent, current system time exceeds the validity period of the result, state fingerprint similarity is lower than the similarity threshold, reuse credibility is lower than the corresponding reuse threshold, or the dynamic judgment layer does not meet the reuse conditions, the reuse of the corresponding historical multimodal judgment result is prohibited, and the multimodal judgment model is called to perform real-time multimodal judgment on the currently collected image or video data; Step S7: After the real-time multimodal assessment is completed, a new multimodal assessment result, a hazard type label, and a new regional status fingerprint are generated. The reliable validity period of the result is determined based on the hazard type label. The new multimodal assessment result, result generation timestamp, hazard type label, reliable validity period of the result, new regional status fingerprint, model version identifier, and sensor combination identifier are written into the local reliable reuse record library. Step S8: Periodically scan the local trusted reuse record library and delete reuse record packages that have exceeded the validity period of the results; when writing a new multimodal judgment result, if the current region coordinate key already exists, overwrite the original reuse record package; if the current region coordinate key does not exist, add a new reuse record package.

[0008] In step S3, the image-aware hash is obtained by performing grayscale conversion, scaling, discrete cosine transform, or mean comparison on the current frame; the Hamming distance between the image-aware hash and the historical-aware hash in the historical region state fingerprint is used to determine the image similarity.

[0009] In step S3, the target quantity vector includes at least one of the following count results: number of personnel, number of vehicles, number of construction machinery, number of fireworks targets, number of targets missing safety helmets, or number of targets missing safety belts; when the target quantity vector changes relative to the historical area state fingerprint, the real-time judgment priority of the dynamic judgment layer is increased.

[0010] In step S4, the reuse credibility is determined by a weighted average of the regional consistency score, state fingerprint similarity, time decay score, and hazard type risk weight. Specifically, the hazard type risk weight corresponding to dynamic hazards is greater than that corresponding to static hazards, ensuring that the reuse threshold for dynamic hazards is higher than that for static hazards. The state fingerprint similarity is obtained by weighted fusion of at least two sub-similarity dimensions: image perceptual hash normalized distance, cosine similarity of the target quantity vector, overlap ratio of the infrared temperature range, and normalized contour distance of the LiDAR contour. The fused comprehensive similarity is then used as the state fingerprint similarity in the reuse credibility calculation.

[0011] In step S4, the time decay amount is used to represent the degree of reusability of historical multimodal judgment results as time goes by; when the difference between the current system time and the result generation timestamp increases, the reliability of reuse decreases.

[0012] In step S4, model version compatibility is determined by comparing the model version identifier in the candidate historical judgment record package with the model version identifier of the current multimodal judgment model; when the model versions are incompatible, the reuse of the candidate historical judgment record package is prohibited.

[0013] In step S4, sensor combination consistency is determined by comparing the sensor combination identifier in the candidate historical analysis record package with the sensor combination currently participating in the data collection; when the sensor combination currently participating in the data collection lacks the sensor type on which the historical analysis depends, the reuse of the candidate historical analysis record package is prohibited.

[0014] In step S5, the static assessment layer includes at least one of the following hazard assessment results: no protection for the foundation pit, cracks in the wall, damage or missing fence, damaged safety net, lack of protection for edge openings, material stacking occupying passages, blocked fire lanes, or missing fixed warning signs.

[0015] In step S5, the dynamic assessment layer includes at least one potential hazard assessment result from the following: personnel not wearing safety helmets, personnel working at height not wearing safety belts, personnel entering dangerous areas, construction machinery operating in violation of regulations, abnormal hoisting operations, open flames or smoke on site, and vehicles speeding or driving against traffic.

[0016] In step S5, when only the static analysis layer meets the reuse condition while the dynamic analysis layer does not, the static analysis layer result is reused, and real-time multimodal analysis is triggered for the dynamic analysis layer, thereby forming a processing result that combines partial reuse with local real-time analysis.

[0017] The beneficial effects of this invention are as follows: First, it transforms ordinary candidate record hits into a reliable reuse judgment after current on-site status verification, avoiding the direct application of the response logic of file-based reuse records to safety inspection conclusions; Second, through lightweight consistency verification of the current frame, it determines whether historical judgment results are still reliable without the need for complete multimodal reasoning; Third, by separating the static judgment layer and the dynamic judgment layer, it allows for the use of different reuse thresholds for stable reuse of regional hazards and frequently changing dynamic hazards such as personnel, vehicles, and smoke; Fourth, through consistency verification of model version and sensor combination, it avoids the misuse of historical results due to changes in the model or acquisition conditions. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 This is a schematic diagram of the current frame lightweight consistency verification and reuse reliability calculation process of the present invention; Figure 3 This is a schematic diagram of the process of separating and reusing the static and dynamic analysis layers in this invention. Detailed Implementation

[0019] The present invention will be further described below with reference to embodiments. These embodiments are used to explain the present invention and do not limit the scope of protection of the present invention.

[0020] Example 1

[0021] A reliable reuse method for multimodal assessment results in UAV inspection includes the following steps: Step S1: The UAV acquires the GPS coordinates of the current inspection area during the inspection process; when the GPS coordinates cannot be acquired or the GPS signal is lost, a disallowed reuse state is generated, and the reuse of historical judgment results is skipped, and real-time multimodal judgment is performed; when the GPS coordinates can be acquired, the longitude and latitude in the GPS coordinates are truncated or rounded to generate a regional coordinate key for screening candidate judgment record packages.

[0022] Step S2: Query the local trusted reuse record library using the area coordinate key to obtain the candidate historical judgment record package corresponding to the current inspection area; the candidate historical judgment record package includes at least the area coordinate key, historical multimodal judgment results, result generation timestamp, hidden danger type label, result trusted validity period, historical area status fingerprint, model version identifier, and sensor combination identifier.

[0023] Step S3: Collect the current frame or current short segment of the current inspection area, extract the lightweight verification features of the current frame, and generate the current area status fingerprint based on the lightweight verification features of the current frame; the lightweight verification features of the current frame include at least one of image perceptual hash, edge contour features, target quantity vector, hidden danger area location summary, infrared temperature range or lidar contour summary; The image-aware hash is obtained by performing grayscale conversion, scaling, discrete cosine transform, or mean comparison on the current frame; the Hamming distance between the image-aware hash and the historical-aware hash in the historical region state fingerprint is used to determine image similarity. The target quantity vector includes at least one of the following count results: number of personnel, number of vehicles, number of construction machinery, number of fireworks targets, number of targets missing safety helmets, or number of targets missing safety belts; when the target quantity vector changes relative to the historical area state fingerprint, the real-time judgment priority of the dynamic judgment layer is increased.

[0024] Step S4: Compare the current region status fingerprint with the historical region status fingerprint in the candidate historical judgment record package to obtain the status fingerprint similarity; and calculate the reuse credibility of the candidate historical judgment record package based on the consistency of region coordinate keys, status fingerprint similarity, time decay between the current system time and the result generation timestamp, risk weight of hidden danger type, model version compatibility and sensor combination consistency. The reuse credibility is determined by a weighted average of regional consistency score, state fingerprint similarity, time decay score, and risk weight of hazard type; wherein, the risk weight of hazard type corresponding to dynamic hazard is greater than that of hazard type corresponding to static hazard, so that the reuse threshold corresponding to dynamic hazard is higher than that corresponding to static hazard. The state fingerprint similarity is obtained by weighted fusion of at least two of the following sub-similarity dimensions: image perception hash normalized distance, cosine similarity of target quantity vector, overlap ratio of infrared temperature range, and normalized contour distance of lidar contour. The fused comprehensive similarity is used as the state fingerprint similarity in the reuse credibility calculation. The time decay amount is used to represent the reduced reusability of historical multimodal analysis results over time; when the difference between the current system time and the result generation timestamp increases, the reliability of reuse decreases. The model version compatibility is determined by comparing the model version identifier in the candidate historical analysis record package with the model version identifier of the current multimodal analysis model; when the model versions are incompatible, the reuse of the candidate historical analysis record package is prohibited. The consistency of the sensor combination is determined by comparing the sensor combination identifier in the candidate historical analysis record package with the sensor combination currently participating in the data collection; when the sensor combination currently participating in the data collection lacks the sensor type on which the historical analysis depends, the reuse of the candidate historical analysis record package is prohibited.

[0025] Step S5: Divide the historical multimodal assessment results into a static assessment layer and a dynamic assessment layer according to the hazard type label; when the reuse credibility reaches the static assessment layer reuse threshold, the static assessment layer results are allowed to be reused; when the number of dynamic targets, the dynamic target location summary, and the dynamic target status summary corresponding to the dynamic assessment layer are consistent with the lightweight verification features of the current frame, and the reuse credibility reaches the dynamic assessment layer reuse threshold which is higher than the static assessment layer reuse threshold, the dynamic assessment layer results are allowed to be reused. The static assessment layer includes at least one of the following hazard assessment results: no protection for the foundation pit, cracks in the wall, damage or missing fence, damaged safety net, lack of protection for edge openings, materials piled up occupying passages, blocked fire lanes, or missing fixed warning signs. The dynamic assessment layer includes at least one potential hazard assessment result from the following: personnel not wearing safety helmets, personnel working at height not wearing safety belts, personnel entering dangerous areas, construction machinery operating in violation of regulations, abnormal hoisting operations, open flames or smoke on site, and vehicles speeding or driving against the flow of traffic. When only the static analysis layer meets the reuse condition while the dynamic analysis layer does not, the static analysis layer result is reused, and real-time multimodal analysis is triggered for the dynamic analysis layer, thereby forming a processing result that combines partial reuse with local real-time analysis.

[0026] Step S6: When any of the following situations occur: no candidate historical judgment record package is found, GPS signal is lost, model version is incompatible, sensor combination is inconsistent, current system time exceeds the validity period of the result, state fingerprint similarity is lower than the similarity threshold, reuse credibility is lower than the corresponding reuse threshold, or the dynamic judgment layer does not meet the reuse conditions, the reuse of the corresponding historical multimodal judgment result is prohibited, and the multimodal judgment model is called to perform real-time multimodal judgment on the currently collected image or video data; Step S7: After the real-time multimodal assessment is completed, a new multimodal assessment result, a hazard type label, and a new regional status fingerprint are generated. The reliable validity period of the result is determined based on the hazard type label. The new multimodal assessment result, result generation timestamp, hazard type label, reliable validity period of the result, new regional status fingerprint, model version identifier, and sensor combination identifier are written into the local reliable reuse record library. Step S8: Periodically scan the local trusted reuse record library and delete reuse record packages that have exceeded the validity period of the results; when writing a new multimodal judgment result, if the current region coordinate key already exists, overwrite the original reuse record package; if the current region coordinate key does not exist, add a new reuse record package.

[0027] Example 2

[0028] This embodiment uses specific examples to illustrate the method in Example 1.

[0029] The drone obtains its current location coordinates via its GPS module and rounds the latitude and longitude to three decimal places to form a regional coordinate key. For example, coordinates (30.123456, 120.654321) are processed into the key "30.123_120.654". This key is only used to filter candidate historical analysis record packages and does not directly determine reuse.

[0030] A local trusted reusable record library can be implemented using Zvec or SQLite. The reusable record package structure includes: Key, Result_JSON, Create_Time, Expire_Time, Hazard_Type, State_Fingerprint, Model_Version, and Sensor_Set. Among them, State_Fingerprint is used to store historical region state fingerprints, and Sensor_Set is used to record the sensor combinations relied upon for historical analysis.

[0031] When the drone re-enters the same area, the system first acquires the current frame or a short segment of 1 to 3 seconds, extracting lightweight verification features from the current frame. The image channel can extract perceptual hashes and edge contour summaries; the target detection channel can extract the number of missing targets such as personnel, vehicles, machinery, smoke, and safety equipment; the infrared channel can extract the temperature range of the area; and the lidar channel can extract contour summaries. These features can be used individually or combined to form the current area status fingerprint.

[0032] State fingerprint similarity can be calculated in various ways. For example, image perceptual hash similarity can be obtained by Hamming distance normalization; target quantity vectors can be obtained by vector distance or cosine similarity; infrared temperature ranges can be obtained by the range overlap ratio; and LiDAR contour summaries can be obtained by contour distance normalization. The results of each similarity can be weighted and fused into a comprehensive state fingerprint similarity.

[0033] As an example, the reuse credibility R can be calculated as follows: R = w1 * C_region + w2 * S_state + w3 * T_decay + w4 * W_risk, where C_region is the regional consistency score, S_state is the normalized state fingerprint similarity, T_decay is the time decay function, W_risk is the risk weight of the hazard type, and w1~w4 are weight coefficients. The reuse credibility can be expressed as a weighted result of the regional consistency score, state fingerprint similarity, time decay score, and hazard type risk weight. Dynamic hazards, due to rapid changes in the field, have higher risk weights and correspondingly higher reuse thresholds; static hazards, because they are less prone to change in a short period, can correspond to lower reuse thresholds.

[0034] When the static assessment layer meets the reuse conditions while the dynamic assessment layer does not, the system can reuse the results of the static assessment layer, while performing real-time multimodal assessment only on the dynamic assessment layer. For example, results regarding unprotected foundation pits and damaged fencing can be reused, while results regarding personnel not wearing safety helmets, vehicles driving in the wrong direction, or abnormal smoke and fire require real-time assessment. This reuse mechanism reduces computational load and prevents dynamic safety hazards from being masked by historical results.

[0035] When GPS signals are lost, the current system time exceeds the validity period of the results, the state fingerprint similarity is insufficient, the reuse credibility is insufficient, the model version is incompatible, or the sensor combination is inconsistent, the system will prohibit the reuse of historical results and force the acquisition of current image or video data and perform real-time multimodal analysis. After the real-time analysis is completed, the new Result_JSON, Create_Time, Expire_Time, Hazard_Type, State_Fingerprint, Model_Version, and Sensor_Set will be written to the reuse record package, overwriting or adding the corresponding records.

[0036] The above description is merely a preferred embodiment of the present invention. For those skilled in the art, equivalent substitutions or modifications can be made without departing from the concept of the present invention, and all such substitutions or modifications should fall within the protection scope of the present invention.

Claims

1. A reliable reuse method for multimodal assessment results in UAV inspection, characterized in that, Includes the following steps: Step S1: The UAV acquires the GPS coordinates of the current inspection area during the inspection process; when the GPS coordinates cannot be acquired or the GPS signal is lost, a disallowed reuse state is generated, and the reuse of historical judgment results is skipped, and real-time multimodal judgment is performed; when the GPS coordinates can be acquired, the longitude and latitude in the GPS coordinates are truncated or rounded to generate a regional coordinate key for screening candidate judgment record packages. Step S2: Query the local trusted reuse record library using the area coordinate key to obtain the candidate historical assessment record package corresponding to the current inspection area; the candidate historical assessment record package includes at least the area coordinate key, historical multimodal assessment results, result generation timestamp, hidden danger type label, result trusted validity period, historical area status fingerprint, model version identifier, and sensor combination identifier; Step S3: Collect the current frame or current short segment of the current inspection area, extract the lightweight verification features of the current frame, and generate the current area status fingerprint based on the lightweight verification features of the current frame; the lightweight verification features of the current frame include at least one of image perception hash, edge contour features, target quantity vector, hidden danger area location summary, infrared temperature range or lidar contour summary, and fuse the lightweight verification features according to preset weights to generate a comprehensive current area status fingerprint; Step S4: Compare the current region status fingerprint with the historical region status fingerprint in the candidate historical judgment record package to obtain the status fingerprint similarity; and calculate the reuse credibility of the candidate historical judgment record package based on the consistency of region coordinate keys, status fingerprint similarity, time decay between the current system time and the result generation timestamp, risk weight of hidden danger type, model version compatibility and sensor combination consistency. Step S5: Divide the historical multimodal assessment results into a static assessment layer and a dynamic assessment layer according to the hazard type label; when the reuse credibility reaches the static assessment layer reuse threshold, the static assessment layer results are allowed to be reused; when the number of dynamic targets, the dynamic target location summary, and the dynamic target status summary corresponding to the dynamic assessment layer are consistent with the lightweight verification features of the current frame, and the reuse credibility reaches the dynamic assessment layer reuse threshold which is higher than the static assessment layer reuse threshold, the dynamic assessment layer results are allowed to be reused. Step S6: When any of the following situations occur: no candidate historical judgment record package is found, GPS signal is lost, model version is incompatible, sensor combination is inconsistent, current system time exceeds the validity period of the result, state fingerprint similarity is lower than the similarity threshold, reuse credibility is lower than the corresponding reuse threshold, or the dynamic judgment layer does not meet the reuse conditions, the reuse of the corresponding historical multimodal judgment result is prohibited, and the multimodal judgment model is called to perform real-time multimodal judgment on the currently collected image or video data; Step S7: After the real-time multimodal assessment is completed, a new multimodal assessment result, a hazard type label, and a new regional status fingerprint are generated. The reliable validity period of the result is determined based on the hazard type label. The new multimodal assessment result, result generation timestamp, hazard type label, reliable validity period of the result, new regional status fingerprint, model version identifier, and sensor combination identifier are written into the local reliable reuse record library. Step S8: Periodically scan the local trusted reuse record library and delete reuse record packages that have exceeded the validity period of the results; when writing a new multimodal judgment result, if the current region coordinate key already exists, overwrite the original reuse record package; if the current region coordinate key does not exist, add a new reuse record package.

2. The method according to claim 1, characterized in that, In step S3, the image-aware hash is obtained by performing grayscale conversion, scaling, discrete cosine transform, or mean comparison on the current frame; the Hamming distance between the image-aware hash and the historical-aware hash in the historical region state fingerprint is used to determine the image similarity.

3. The method according to claim 1, characterized in that, In step S3, the target quantity vector includes at least one of the following count results: number of personnel, number of vehicles, number of construction machinery, number of fireworks targets, number of targets missing safety helmets, or number of targets missing safety belts; when the target quantity vector changes relative to the historical area state fingerprint, the real-time judgment priority of the dynamic judgment layer is increased.

4. The method according to claim 1, characterized in that, In step S4, the reuse credibility is determined by a weighted result of regional consistency score, state fingerprint similarity, time decay score, and risk weight of hazard type. Among them, the risk weight of hazard type corresponding to dynamic hazard is greater than that of hazard type corresponding to static hazard, so that the reuse threshold corresponding to dynamic hazard is higher than that corresponding to static hazard. The state fingerprint similarity is obtained by weighted fusion of at least two sub-similarities of the following dimensions: image perception hash normalized distance, cosine similarity of target quantity vector, overlap ratio of infrared temperature range, and normalized contour distance of lidar contour. The fused comprehensive similarity is used as the state fingerprint similarity in the reuse credibility calculation.

5. The method according to claim 1, characterized in that, In step S4, the time decay amount is used to represent the degree of reusability of historical multimodal judgment results as time goes by; when the difference between the current system time and the result generation timestamp increases, the reliability of reuse decreases.

6. The method according to claim 1, characterized in that, In step S4, model version compatibility is determined by comparing the model version identifier in the candidate historical judgment record package with the model version identifier of the current multimodal judgment model; when the model versions are incompatible, the reuse of the candidate historical judgment record package is prohibited.

7. The method according to claim 1, characterized in that, In step S4, sensor combination consistency is determined by comparing the sensor combination identifier in the candidate historical analysis record package with the sensor combination currently participating in the data collection; when the sensor combination currently participating in the data collection lacks the sensor type on which the historical analysis depends, the reuse of the candidate historical analysis record package is prohibited.

8. The method according to claim 1, characterized in that, In step S5, the static assessment layer includes at least one of the following hazard assessment results: no protection for the foundation pit, cracks in the wall, damage or missing fence, damaged safety net, lack of protection for edge openings, material stacking occupying passages, blocked fire lanes, or missing fixed warning signs.

9. The method according to claim 1, characterized in that, In step S5, the dynamic assessment layer includes at least one potential hazard assessment result from the following: personnel not wearing safety helmets, personnel working at height not wearing safety belts, personnel entering dangerous areas, construction machinery operating in violation of regulations, abnormal hoisting operations, open flames or smoke on site, and vehicles speeding or driving against traffic.

10. The method according to claim 1, characterized in that, In step S5, when only the static analysis layer meets the reuse condition while the dynamic analysis layer does not, the static analysis layer result is reused, and real-time multimodal analysis is triggered for the dynamic analysis layer, thereby forming a processing result that combines partial reuse with local real-time analysis.

Images