Low-power intelligent camera recognition system for construction site hazard detection

Abstract

The application relates to the technical field of intelligent security monitoring and computer vision recognition, in particular to a low-power intelligent camera recognition system for construction site hazard source detection; the system comprises a pyroelectric infrared sensor, an image sensor, an identification processor, a rechecking unit and a control management unit; the sensor end performs pixel difference, grid energy statistics and connected domain aggregation on continuous frames, and outputs a first wake-up flag and an active area; the identification processor preliminarily screens a low-resolution global image, when the condition is met, only reads a corresponding high-resolution area to judge the hazard source category and outputs a position result; when the background interference is too large, subsequent processing is terminated; the application reduces invalid data movement and identification calculation, reduces power consumption and improves the construction site hazard source detection accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent security monitoring and computer vision recognition technology, specifically a low-power intelligent camera recognition system for detecting hazardous sources at construction sites. Background Technology

[0002] The existing construction site hazard detection system mainly uses a pyroelectric infrared sensor to trigger and wake up an image sensor to collect on-site data. Then, the image signal processor, memory and recognition processor perform unified transmission, caching and recognition processing on the entire high-resolution image.

[0003] However, existing solutions lack mechanisms for continuous frame change filtering and active region extraction at the image sensor end, as well as for initial screening of low-resolution regions at the recognition processor end. Because the entire high-resolution image is directly read and calculated uniformly, environmental disturbances such as swaying shade nets, exhaust heat flow, and strong light sweeping can easily enter the subsequent recognition link. In complex construction site scenarios, the above defects cause a large amount of invalid data transfer and invalid recognition calculations, resulting in increased overall system power consumption and reduced standby time. The single triggering mechanism leads to frequent false triggering of the system, poor recognition stability, and inaccurate grasp of the location of local hazards, reducing the effectiveness of practical applications. Summary of the Invention

[0004] The purpose of this invention is to provide a low-power intelligent camera recognition system for detecting hazards at construction sites. This system avoids the need to move large amounts of invalid data and cause a sudden increase in overall system power consumption. It also more easily filters out environmental disturbances and accurately pinpoints the location of local hazards, thereby improving the system's recognition stability. Specifically, the technical solution of this invention includes:

[0005] The system includes a pyroelectric infrared sensor, an image sensor, an image signal processor, memory, and a recognition processor; a data bus connects the image sensor and the recognition processor; and...

[0006] The sensor-side preprocessing unit, integrated into the image sensor, is used to perform motion energy statistics based on the continuous frame image data output by the image sensor to generate an active area after receiving the micro wake-up signal output by the pyroelectric infrared sensor, and output a first-level wake-up flag according to the preset feature judgment conditions.

[0007] The initial screening unit, deployed on the recognition processor, is used to acquire the low-resolution global image and active area output by the image sensor after receiving the first-level wake-up flag, and extract the regional image features to evaluate the background interference value and the target determination value; when the background interference value does not exceed the preset background interference threshold and the target determination value reaches the target existence threshold, it outputs the second-level wake-up flag and the initial detection area set.

[0008] The re-inspection unit is activated upon receiving the secondary wake-up flag, retrieves high-resolution image data of the corresponding area of ​​the initial inspection area set, performs hazard source category analysis, and outputs hazard source category, hazard source judgment value, and location result.

[0009] The control and management unit is used to uniformly manage the state switching of various system components, terminate subsequent recognition operations when the background interference value exceeds the preset background interference threshold or the target judgment value does not reach the target existence threshold; and is used to adjust the historical sample records and system judgment thresholds based on the recognition results output by the re-inspection unit.

[0010] Preferably, the sensor-side preprocessing unit is configured to: perform pixel-level difference based on continuous frame image data to generate a motion mask, mesh the current global image corresponding to the continuous frame image data to statistically analyze the mesh motion energy, calculate the discrete motion value characterizing the motion energy distribution law, and perform connected component aggregation operation based on the motion mask to obtain the active region.

[0011] Preferably, the specific logic for the sensor-side preprocessing unit to output a first-level wake-up flag based on preset feature determination conditions is as follows:

[0012] When the motion discrete value is less than the preset preprocessing blocking threshold, and the maximum area of ​​the active region and the number of active regions both meet the preset minimum effective area threshold and the upper limit of the number of connected regions, it is determined that there is suspected entity motion, and a first-level wake-up flag and active region are output; otherwise, the current frame data is discarded at the image sensor and the sleep state is restored.

[0013] Preferably, the primary screening unit is configured as follows:

[0014] Generate a region filtering mask based on the circumscribed coordinates of the activity region;

[0015] Background filtering is performed on a low-resolution global image based on a region filtering mask to obtain the target region image data.

[0016] Extract pixel distribution features or edge features from the target region image data to calculate background interference value and target determination value, and generate a set of initial detection regions containing low-resolution coordinates when the background interference value does not exceed the preset background interference threshold and the target determination value reaches the target existence threshold.

[0017] Preferably, the re-inspection unit is configured as follows:

[0018] Obtain preset scale mapping coefficients and preset region expansion parameters, map and expand the low-resolution coordinates in the initial detection region set to the original high-resolution coordinates to obtain a high-definition reading region set; and perform hazard source category determination calculation based on the high-resolution image data corresponding to the high-definition reading region set.

[0019] Preferably, the feedback adjustment function of the control management unit includes:

[0020] The system determines whether an alarm is a real alarm or a false alarm by comparing the hazard source determination value with the preset confirmation threshold, thereby updating the system's historical false alarm rate.

[0021] Obtain the ambient illuminance parameters of the system currently in operation;

[0022] Based on historical false alarm rates and ambient illuminance parameters, the background interference threshold of the primary screening unit is dynamically adjusted.

[0023] Preferably, the logic for the control management unit to dynamically adjust the background interference threshold is as follows: based on the preset basic exit threshold, compensation adjustment is performed based on the historical false alarm rate and ambient illuminance parameters.

[0024] Preferably, the logic for the control management unit to dynamically adjust the preprocessing blocking threshold is as follows: based on the preset basic discrete threshold, compensation adjustment is performed by comprehensively considering the historical false alarm rate data and continuous trigger rate data of the system.

[0025] Preferably, the control and management unit is also used to compare the coordinate differences between the initial inspection area set and the high-definition reading area set to calculate the target positioning deviation rate, and dynamically update the area expansion parameters used by the re-inspection unit according to the target positioning deviation rate.

[0026] Preferably, the system further includes an early warning output unit, which generates an alarm signal and triggers an external alarm prompt when the hazard source category output by the re-inspection unit matches a preset category and the hazard source judgment value is greater than or equal to a preset alarm threshold; it does not trigger an alarm when the preset alarm threshold is not reached.

[0027] Compared with the prior art, the present invention has the following beneficial effects:

[0028] 1. This invention achieves hierarchical wake-up and hierarchical processing from the sensor side to the processor side by using a pyroelectric infrared sensor for coarse triggering, continuous frame pixel difference and grid energy statistics at the image sensor end, active area extraction and first-level wake-up output, combined with the recognition processor end for calculating background interference values ​​and target judgment values ​​based on low-resolution global images and active areas, and second-level wake-up control. It can intercept background disturbances such as swaying sunshade nets, exhaust heat flow, and strong light sweeping at the front end, reduce invalid data transfer and invalid recognition calculations, reduce overall system power consumption and improve standby battery life.

[0029] 2. This invention generates an alarm signal containing the hazard category and location results when the hazard source category matches and the hazard source judgment value reaches the alarm threshold through the early warning output unit, and controls the external alarm device to issue an alarm prompt, thereby forming a complete detection link from low-power screening, local precise re-inspection, parameter closed-loop update to on-site early warning output, and improving the practical application effect of the construction site hazard source detection system. Attached Figure Description

[0030] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application:

[0031] Figure 1 This is a schematic diagram of a low-power intelligent camera recognition system for detecting hazardous sources at construction sites, provided in an embodiment of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0033] A low-power intelligent camera recognition system for detecting hazardous sources at construction sites includes: a pyroelectric infrared sensor, an image sensor, an image signal processor, memory, and a recognition processor; a data bus connects the image sensor and the recognition processor; and...

[0034] The sensor-side preprocessing unit, integrated into the image sensor, is used to perform motion energy statistics based on the continuous frame image data output by the image sensor to generate an active area after receiving the micro wake-up signal output by the pyroelectric infrared sensor, and output a first-level wake-up flag according to the preset feature judgment conditions.

[0035] The initial screening unit, deployed on the recognition processor, is used to acquire the low-resolution global image and active area output by the image sensor after receiving the first-level wake-up flag, and extract the regional image features to evaluate the background interference value and the target determination value; when the background interference value does not exceed the preset background interference threshold and the target determination value reaches the target existence threshold, it outputs the second-level wake-up flag and the initial detection area set.

[0036] The re-inspection unit is activated upon receiving the secondary wake-up flag, retrieves high-resolution image data of the corresponding area of ​​the initial inspection area set, performs hazard source category analysis, and outputs hazard source category, hazard source judgment value, and location result.

[0037] The control and management unit is used to uniformly manage the state switching of various system components, terminate the subsequent recognition operation when the background interference value exceeds the preset background interference threshold or the target judgment value does not reach the target existence threshold; and is used to adjust the historical sample records and system judgment thresholds based on the recognition results output by the re-inspection unit.

[0038] The sensor-side preprocessing unit is configured to: perform pixel-level difference based on continuous frame image data to generate a motion mask, mesh the current global image corresponding to the continuous frame image data to statistically analyze the mesh motion energy, calculate the discrete motion value characterizing the distribution law of motion energy, and perform connected component aggregation operation based on the motion mask to obtain the active region;

[0039] The specific logic of the sensor-side preprocessing unit outputting the first-level wake-up flag based on preset feature determination conditions is as follows:

[0040] When the motion discrete value is less than the preset preprocessing blocking threshold, and the maximum area and the number of active regions both meet the preset minimum effective area threshold and the upper limit of the number of connected regions, it is determined that there is suspected entity motion, and a first-level wake-up flag and active region are output; otherwise, the current frame data is discarded at the image sensor and the sleep state is restored.

[0041] The primary screening unit is configured as follows:

[0042] Generate a region filtering mask based on the circumscribed coordinates of the activity region;

[0043] Background filtering is performed on a low-resolution global image based on a region filtering mask to obtain the target region image data.

[0044] Extract pixel distribution features or edge features from the target region image data to calculate background interference value and target determination value, and generate a set of initial detection regions containing low-resolution coordinates when the background interference value does not exceed the preset background interference threshold and the target determination value reaches the target existence threshold.

[0045] The re-inspection unit is configured as follows:

[0046] Obtain preset scale mapping coefficients and preset region expansion parameters, map and expand the low-resolution coordinates in the initial detection region set to the original high-resolution coordinates to obtain a high-definition reading region set; and perform hazard source category determination calculation based on the high-resolution image data corresponding to the high-definition reading region set.

[0047] In this embodiment, the system operates using a tiered wake-up and tiered processing approach, such as... Figure 1 As shown; the pyroelectric infrared sensor is only responsible for providing a coarse trigger signal indicating whether there is a thermal change. The image sensor first completes the continuous frame change filtering and only outputs a first-level wake-up flag when the preset conditions are met.

[0048] The recognition processor first performs an initial screening of the low-resolution global image, and only outputs a secondary wake-up flag when the background interference value meets the preset stop processing condition and the target judgment value meets the target existence threshold. Then, it only performs hazard source category judgment on the high-resolution image data corresponding to the initial detection area. Thus, the system output results include hazard source category, hazard source judgment value and location result. At the same time, the control and management unit decides whether to terminate the subsequent processing in advance based on the background interference value to avoid invalid data transfer and invalid recognition calculation.

[0049] Furthermore, based on the above implementation method, each step will be explained in detail;

[0050] The micro-wake-up signal output by the pyroelectric infrared sensor is acquired, and the preprocessing mode of the image sensor is started according to the hardware interrupt triggering relationship to generate a preprocessing task of continuous frame image data.

[0051] Among them, the micro wake-up signal refers to the low-power trigger signal output by the pyroelectric infrared sensor after detecting a change in thermal radiation. This signal can be sent to the control management unit and the image sensor through level changes of general-purpose input / output pins or hardware interrupt lines, such as thermal changes caused by thermal flow disturbances, personnel entry, or mechanical passage. In terms of electrical performance, the micro wake-up signal is specifically manifested as a high-level wake-up pulse or rising edge trigger pulse triggered by the above-mentioned physical thermal changes.

[0052] Continuous frame image data refers to the raw image data of two or more adjacent frames acquired by the image sensor within the current system cycle. The data first enters the line buffer or on-chip cache inside the image sensor, rather than being directly sent to the image signal processor or memory.

[0053] This implementation performs local judgment on the image sensor side and moves early screening to the pixel array layer; this configuration is used to intercept false triggering events such as sunshade net swing, exhaust heat flow, and strong light disturbance, changing the originally indiscriminate full-link wake-up to conditional step-by-step wake-up.

[0054] This implementation method can achieve pre-stage screening based on hardware interrupts and image sensor pre-processing, avoiding direct high-power processing after the pyroelectric infrared sensor is falsely triggered.

[0055] The original pixel values ​​of two consecutive frames in the continuous frame image data are obtained. According to the preset pixel-level difference threshold, pixel-level difference operation is performed on the two consecutive frames to generate a binary motion mask. The pixel-level difference operation is used to extract the position of inter-frame change in order to distinguish between static background and dynamic area.

[0056] By calculating the absolute difference between the original pixel value at the current moment and the original pixel value in the previous frame, the inter-frame difference value at the corresponding coordinate position is extracted. This inter-frame difference value represents the intensity of brightness change at the current position in adjacent sampling periods in this scene, and is used to determine whether effective motion has occurred at this position rather than static noise.

[0057] Furthermore, the inter-frame difference value is compared with a preset pixel-level difference threshold: when the difference value is greater than the threshold, the corresponding position is marked as a valid change point, and other cases are marked as no change points, thereby directly generating a binary motion mask;

[0058] Among them, the pixel-level differential threshold is used as a threshold for filtering out background noise and weak flicker at the image sensor end. The value can be set to a calibrated value within a preset range in combination with the sensor noise level; marking as a valid change point indicates that there is a suspicious change at that position; the generated binary motion mask is preferably stored in the on-chip storage array at the image sensor end, and subsequent grid statistics and connected component aggregation are directly performed on the mask;

[0059] This implementation uses pixel-by-pixel differential combined with threshold compression, which can more accurately preserve local motion traces and avoid global brightness fluctuations from interfering with or covering the local features of the real target. This implementation can directly extract the changed position based on pixel-level differential and binarization filtering, avoiding invalid background changes from entering subsequent links.

[0060] A low-resolution global image is acquired. Based on the division rules of multiple non-overlapping grids, grid energy statistics are performed on the binary motion mask to generate the motion energy and distribution probability of each grid region. The low-resolution global image refers to the large field-of-view low-resolution image used by the image sensor for rapid judgment, such as the compressed global preview image, whose resolution is lower than the original high-resolution image used in subsequent re-examination.

[0061] Specifically, the width and height of the low-resolution global image are preferably 1 / 4 to 1 / 8 of the original high-resolution image; for example, when the original high-resolution image is 1920×1080 pixels, the low-resolution global image can be set to 480×270 pixels, so as to reduce the amount of bus transmission data to one-sixteenth of the original image while ensuring that the basic connected shape of the target can be distinguished.

[0062] Non-overlapping grids refer to several regions on the low-resolution image divided into a preset number of rows and columns, such as eight rows and eight columns of grids, with each grid covering a continuous block of pixels. For each grid region, the total number of pixels whose mask values ​​are effective change points within the grid is directly used as the motion energy of that grid region.

[0063] The motion energy characterizes the level of motion activity within the corresponding grid. The larger the value, the more concentrated the pixel changes within that grid. The motion energy of each grid region is then divided by the sum of the motion energy of all grids globally to normalize the motion energy and obtain the distribution probability of each grid. This distribution probability characterizes the proportion of the motion energy of that grid to the current global motion energy, and is used to determine whether the change is locally concentrated or globally scattered.

[0064] This implementation method preserves spatial distribution information through grid energy statistics, enabling the separation of locally continuous human entry from globally scattered heat flow fluctuations and light sweeping. This implementation method can obtain the basis for changing spatial distribution based on grid energy statistics, avoiding misjudgment by only looking at the total change.

[0065] The distribution probability of each grid region is obtained. Based on the distribution probability, the motion discrete value of the current frame is calculated to generate an active region set. The motion discrete value describes the dispersion of motion energy in each grid and can be represented by information entropy or variance. This implementation prefers to use information entropy.

[0066]

[0067] in, and These represent the row and column numbers of the grid, respectively; For the first Line 1 The probability distribution of the corresponding grid, the range of the double summation covering all globally divided grid regions, and the minimum constant. This is used to prevent logarithmic operations from becoming meaningless, ensuring stable calculations even in low-motion scenarios;

[0068] Calculated information entropy It characterizes the degree of dispersion of motion distribution in the current frame and is directly used as the motion discrete value. When the motion discrete value is greater than the preset high-order threshold, it usually indicates that the degree of change dispersion is higher than the preset high-order threshold, which is common in background disturbances such as heat flow and strong light sweeping of searchlights.

[0069] When the discrete value of the motion is less than or equal to the preset low-bit threshold, it usually indicates that the changes are concentrated and the motion characteristics are closer to the edge of personnel, equipment or local hazard source entities. After obtaining the binary motion mask, the connected component aggregation is further performed on the mask to obtain the set of active regions. Each active region contains the bounding box coordinates, area and mesh energy percentage.

[0070] The bounding box coordinates are used to give the location of the candidate region, the area is used to measure whether the region has reached the minimum entity scale, and the grid energy ratio is used to reflect the contribution of the region to global changes. The connected component aggregation results are preferably written to the output buffer of the image sensor and given to the subsequent stage for reading in the form of structured region parameters. In complex construction site scenes, the pre-stage screening needs to maintain separability when large-scale global disturbances and local entity entry coexist.

[0071] By introducing motion discrete values ​​in this implementation, dispersed noise can be reduced and concentrated activity can be preserved under the same triggering conditions. It is especially suitable for extreme working conditions such as strong light sweeping and wind blowing the covering. It can give the result of whether it is possible to be an entity and the entity's position coordinates based on motion discrete values ​​and connected component aggregation, avoiding the need for subsequent processing units to perform indiscriminate traversal search calculations on the global image data.

[0072] Obtain the set of active areas and determine the first-level wake-up flag according to the preset judgment conditions; among them, according to the system judgment rules, the preset judgment conditions include the joint verification of three indicators: the first indicator is whether the motion discrete value is less than the preprocessing blocking threshold.

[0073] The initial value of the preprocessing blocking threshold is calibrated by combining the baseline information entropy of the image sensor in a standard background environment with no wind and no strong light interference. It is used to delineate the critical line of the degree of dispersion between the motion of concentrated entity targets and the divergent background noise.

[0074] The second metric is to determine whether the area of ​​the largest active region among all connected active regions is greater than the minimum effective area threshold, which is used to reflect whether the active region has reached the preset minimum effective entity size standard; the third metric is to determine whether the number of active regions extracted in the current frame is less than or equal to the upper limit of the number of connected regions, which is used to identify and remove overly fragmented disturbance scenes.

[0075] The upper limit of the number of connected regions is set based on the estimated maximum number of effective discrete entity targets that may appear simultaneously within the preset monitoring field of view. For example, it can be set to a fixed integer between 5 and 15 to prevent dense fragmented connected regions formed by large-area leaf shaking or snowflakes falling from being misjudged.

[0076] The system performs strict logical AND operations on this: only when the above three indicators are met simultaneously is it determined that there is suspected entity movement, thus writing a level one wake-up flag into the image sensor status register and writing the set of active areas into the output buffer; once any condition is not met, the current frame data is directly released at the image sensor end, and the control management unit restores the image sensor to sleep state.

[0077] This implementation uses a combination of discreteness, region area, and region number to simultaneously suppress large-scale scattered artifacts and excessive fragmentation noise, reducing false blocking caused by relying on a single feature; it can output a first-level wake-up flag and active region based on joint determination to avoid false release or false blocking in the previous stage.

[0078] The first-level wake-up flag is obtained. Based on the bounding box coordinates of the active area, the low-resolution global image is filtered to generate background interference value, target judgment value and initial detection area set. In the specific implementation, the first-level wake-up flag can be written to the interrupt status bit by the image sensor. After the recognition processor receives the interrupt, it only opens the low-speed sensor bus to read the low-resolution global image and active area parameters, instead of immediately reading the high-resolution full frame data.

[0079] The active region serves as location reference information. A binary mask image of the same size as the low-resolution global image is generated in the local memory of the recognition processor. The mask image is zero outside the active region and has valid values ​​within the active region. The mask image is then multiplied pixel by pixel with the low-resolution global image to obtain the region image data.

[0080] Background interference value is used to measure whether the current candidate region is more like background disturbance, and target determination value is used to measure whether the current candidate region has real target features; background interference value can be calculated by non-zero pixel ratio or edge gradient features; target determination value can also be obtained by combining non-zero pixel ratio or edge gradient features.

[0081] In this embodiment, the specific calculation rules for the regional image data are set as follows: the system performs global pixel statistics and edge detection on the regional image data, i.e., the candidate blocks selected by the mask;

[0082] The specific calculation rules are set as follows: the total number of non-zero pixels with brightness values ​​greater than the preset noise baseline in the statistical region image is divided by the total coverage area of ​​the bounding box to obtain the non-zero distribution density; strong edge pixels with gradient magnitudes greater than the preset edge threshold in the region are extracted, and the connectivity distribution rate of these edge pixels is calculated.

[0083] Since the typical characteristics of background disturbances such as the swaying of shade nets are large change areas but lack obvious edge gradient features, while real entities have continuous edge contours with strong connectivity, the system directly maps the non-zero distribution density multiplied by a preset gain base of, for example, 100 as the background disturbance value; at the same time, it maps the connectivity distribution rate of strong edge pixels multiplied by a full-scale base of, for example, 100 as the target determination value.

[0084] For example: Suppose we extract image data of a region of 40×40 pixels, or 1600 pixels in total. Its non-zero distribution density is calculated to be 0.75. However, due to the lack of continuous gradients inside, the connectivity distribution rate of strong edges is lower than the preset benchmark, and the calculated value is 0.05. According to this rule, the background interference value is 75 and the target determination value is 5.

[0085] Among them, the proportion of non-zero pixels reflects the effective structural coverage within the candidate area, and the edge gradient features reflect the intensity of contour changes in the candidate area, such as the contour edges of people and the edges of equipment components. When the background interference value is less than or equal to, for example, the background interference threshold currently set to 30 by the system, and the target judgment value is greater than or equal to, for example, the target threshold currently set to 20 by the system, since 75 is much greater than 30 and the judgment value is less than the threshold in the above example, the condition is not met, and the control management unit directly shuts down subsequent processing.

[0086] If the conditions are met, the recognition processor generates a preliminary detection region set in local memory, which contains low-resolution coordinates; if the conditions are not met, the control management unit directly shuts down the image signal processor, memory, and re-detection unit.

[0087] This implementation uses the region prior provided by the sensor to perform mask screening and makes judgments around the candidate region. It can concentrate the subsequent calculations on the local region where the probability of the target existence is higher than a preset threshold. In the field of video scene understanding, the early stage usually focuses on two indicators: artifact filtering rate and suspicious target retention rate.

[0088] This implementation uses a region filtering map to directly set large areas of background to zero. When wind blows and obstructions sway frequently or strong light sweeps across at night, the background interference value will rise to exceed the stopping threshold, thus terminating the processing at an earlier stage. When the candidate region has clear edges and concentrated changes, the target judgment value will rise to meet the target judgment threshold, thus retaining it to enter the next stage. Preliminary target verification can be completed based on region mask and low-resolution filtering, avoiding sending invalid background frames into subsequent high-resolution judgment.

[0089] A preliminary detection area set is obtained. Based on the preset scale mapping coefficient and area expansion parameter, the low-resolution coordinates are mapped to generate a high-resolution reading area set. The corresponding high-resolution image data is then used to determine the hazard source category. The scale mapping coefficient represents the proportional relationship between the low-resolution coordinates and the original high-resolution coordinates. The area expansion parameter is used to expand the corresponding boundary range outward after the low-resolution coordinates are mapped to compensate for positioning errors and missing context.

[0090] The system combines the scale mapping coefficient with the region expansion parameter through business logic as a combined calculation rule of coordinate transformation and boundary expansion. It directly performs mapping and expansion operations on the low-resolution coordinates in the initial detection region set. The calculation result is the corresponding high-definition reading region set, which is used to guide the image sensor, image signal processor and memory to perform high-resolution data reading in local areas instead of reading the full frame data.

[0091] In practice, the low-resolution coordinates in the initial detection area set are first written into the area buffer on the recognition processor side, and the control management unit opens the corresponding high-resolution area reading process accordingly; the read high-resolution image data is written into memory for the re-detection unit to call.

[0092] The re-inspection unit performs category judgment calculations on the high-resolution image data to obtain the hazard source category, hazard source judgment value and location results. The recognition processor then sends the above results back to the control and management unit through the internal data bus.

[0093] The category determination calculation is as follows: high-resolution image data is input into a pre-trained target detection neural network model for feature extraction and classification regression. The output predicted category is the hazard source category, the output predicted classification confidence score is the hazard source determination value, and the output predicted bounding box coordinates are the location result.

[0094] This implementation uses low-resolution to high-resolution region mapping and on-demand reading, which can limit bandwidth and computation to the candidate range. It is especially suitable for construction site monitoring scenarios where the hazard source usually only occupies a local area of ​​the field of view. High-resolution local re-inspection can be achieved based on scale mapping and region expansion, avoiding the extra power consumption caused by full-frame high-resolution reading and full-frame depth judgment.

[0095] In actual deployment, the system can be installed at fixed monitoring points in the construction site area; the pyroelectric infrared sensor is used to generate coarse triggers, the image sensor enters continuous frame differential and grid statistical processing according to the interrupt, the recognition processor reads the low-resolution global image according to the first-level wake-up flag and forms the initial inspection area, and the re-inspection unit only performs hazard source category judgment on the high-resolution image data corresponding to the high-resolution reading area set.

[0096] Finally, the control and management unit decides whether to stop processing based on the background interference value, and writes the hazard source category, hazard source judgment value and location result into the event log area to provide input data for subsequent parameter adjustment; in this way, new on-site image data can continuously enter the process according to the system cycle, and trigger subsequent identification and result output when the conditions are met;

[0097] Furthermore, the feedback adjustment function of the control management unit includes:

[0098] The system determines whether an alarm is a real alarm or a false alarm by comparing the hazard source determination value with the preset confirmation threshold, thereby updating the system's historical false alarm rate.

[0099] Obtain the ambient illuminance parameters of the system currently in operation;

[0100] Based on historical false alarm rates and ambient illuminance parameters, the background interference threshold of the primary screening unit is dynamically adjusted.

[0101] The logic of the control management unit to dynamically adjust the background interference threshold is as follows: based on the preset basic exit threshold, compensation adjustment is performed based on the historical false alarm rate and ambient illuminance parameters.

[0102] The logic of the control management unit to dynamically adjust the preprocessing blocking threshold is as follows: based on the preset basic discrete threshold, compensation adjustment is performed by comprehensively considering the historical false alarm rate data and continuous trigger rate data of the system.

[0103] The control and management unit is also used to compare the coordinate differences between the initial inspection area set and the high-definition reading area set to calculate the target positioning deviation rate, and dynamically update the area expansion parameters used by the re-inspection unit based on the target positioning deviation rate.

[0104] In this embodiment, the control management unit is not only responsible for state switching, but also for writing back the re-inspection results as subsequent processing parameters;

[0105] Specifically, after the re-inspection unit outputs the hazard source category, hazard source judgment value and location result, the control and management unit first determines whether the event is a real hazard source or a false trigger. Then, it combines the ambient illuminance output by the image signal processor, the sensor wake-up frequency, the coordinate difference between the initial inspection area set and the high-definition reading area set, and the statistical results obtained from periodic full-frame scanning to update the stop processing threshold, preprocessing blocking threshold and area expansion parameters.

[0106] With the above configuration, the intensity of the pre-stage screening and the range of the subsequent reading in the system cycle can be adjusted according to changes in the on-site working conditions;

[0107] Furthermore, based on the above implementation method, the feedback processing steps are explained in detail;

[0108] The system obtains the hazard determination value output by the re-inspection unit, and determines whether the current event is a real hazard or a false trigger based on the preset confirmation threshold, and generates a corresponding sample record. The confirmation threshold is a preset value used to distinguish between a hazard that has been confirmed as a hazard and a false trigger.

[0109] When the hazard determination value reaches or exceeds the confirmation threshold, the control management unit maintains the alarm link and writes the corresponding image area, hazard category, and location results into the hazard confirmation sample cache area; when the hazard determination value is lower than the confirmation threshold, the control management unit marks the event as a false trigger and writes the corresponding image area and determination result into the false trigger sample cache area, while updating the historical false alarm rate; the sample cache area here can be a designated storage area in memory, or it can be written to a non-volatile storage medium before the end of the system cycle;

[0110] This implementation method continues to use the re-inspection results for subsequent statistics, providing a direct basis for threshold adjustment; this can quantify the problem of unfiltered artifacts in the previous stage, rather than maintaining a fixed threshold in the long term; it can complete event classification and sample preservation based on the hazard source judgment value, avoiding the lack of result feedback path;

[0111] The exposure parameters of the image signal processor are obtained, and the current ambient illuminance is converted based on the exposure parameters to generate an illuminance input for threshold adjustment. The exposure parameters are preferably obtained from the automatic exposure metering results inside the image signal processor. The current ambient illuminance represents the brightness level of the monitored area within the current system cycle and is used to determine whether it is in a low-light environment, such as low light at night or insufficient brightness due to partial occlusion.

[0112] In a specific implementation, the control management unit reads the exposure parameters from the status register or shared memory area of ​​the image signal processor. After reading, the parameters are used for the current parameter calculation. After the calculation is completed, the temporary buffer corresponding to the current reading is released.

[0113] This implementation method directly incorporates ambient illuminance into the adjustment criteria for the stop processing threshold, which can avoid excessive blocking of the real target by the pre-stage in low light conditions; it can extract ambient illuminance input based on exposure parameters, avoiding the use of the same blocking intensity during the day in low light scenes;

[0114] The system obtains the preset basic exit threshold, historical false alarm rate adjustment coefficient, low light critical illuminance threshold, and environmental illuminance compensation coefficient. Based on the historical false alarm rate and the current environmental illuminance, it adjusts the stop processing threshold to generate an updated stop processing threshold. The basic exit threshold represents the starting threshold for the system to stop processing in the initial screening stage under standard scenarios.

[0115] The historical false alarm rate adjustment coefficient is used to control the proportion of the influence of the historical false alarm rate on the stop processing threshold; the low light critical illuminance threshold serves as the boundary condition for distinguishing between normal illuminance and low light illuminance.

[0116] The ambient illuminance compensation coefficient is used to control the proportional parameter of the threshold reduction under low light conditions; and this compensation coefficient has a dimension conversion function, which is used to convert the physical dimension of the illuminance difference into a numerical dimension consistent with the basic exit threshold; the product of the historical false alarm rate adjustment coefficient and the historical false alarm rate is calculated to obtain the first adjustment term; wherein, the historical false alarm rate represents the proportion of events that are judged as false triggers by the subsequent level within the preset statistical window;

[0117] When the historical false alarm rate increases, it indicates that more background artifacts were missed in the initial screening stage. In order to implement more stringent blocking in the future, the dynamic stop processing threshold should be narrowed towards a more stringent lower limit. For this reason, in terms of the technical motivation for configuring algorithm parameters, this embodiment explicitly sets the historical false alarm rate adjustment coefficient to a negative parameter such as -50.

[0118] Therefore, when the basic exit threshold is added as the first adjustment term, it essentially makes the stop processing threshold decrease as the false alarm rate increases, thus raising the blocking threshold; then the difference between the weak light critical illuminance threshold and the current ambient illuminance is calculated; when the difference is greater than zero, the product of the difference and the ambient illuminance compensation coefficient is used as the second adjustment term; when the difference is less than or equal to zero, the second adjustment term is set to zero.

[0119] The setting mechanism of the second adjustment item is as follows: when the illuminance is lower than the weak light critical illuminance threshold, the image signal-to-noise ratio decreases and the overall interference assessment value is prone to a positive deviation higher than the actual interference distribution. In order to avoid the interference target being blocked and exited prematurely due to the attenuation of image details, it is necessary to increase the stop processing threshold through bias compensation. For this reason, the ambient illuminance compensation coefficient configured by the system is specially set to a negative parameter, for example, -0.5.

[0120] Since the embodiment logically performs the subtraction of the second adjustment term, that is, subtracting a negative product, this forms a positive compensation bias in mathematical calculation, successfully achieving the upward expansion of the threshold in low light environment; the base exit threshold is added to the first adjustment term and the second adjustment term is subtracted, and the calculated numerical result is directly used as the updated stop processing threshold applied to the primary screening unit in the current system cycle.

[0121] Assuming the system's base exit threshold is 30, the historical false alarm rate adjustment coefficient is -50, and the ambient illuminance compensation coefficient is -0.5; in a low-light nighttime scene, the current ambient illuminance is extracted to be 10 Lux, and the preset low-light critical illuminance is 30 Lux, the difference is 20; the calculated second adjustment term is... Meanwhile, if the recent historical false alarm rate climbs to 10%, or 0.1%, then the first adjustment term will be... The updated stop processing threshold according to the calculation logic = ;

[0122] This achieves the dynamic superposition and balance of the system's false alarm compensation mechanism in low-light environments and the judgment threshold tightening mechanism under frequent false alarms;

[0123] In construction sites with strong interference at night, common evaluation criteria are whether false alarms continue to increase and whether low light causes significant missed alarms. If only a fixed threshold is used, the system will continue to generate invalid calculations in subsequent stages once the historical false alarms increase. If only a single static method is used to raise the judgment threshold, it is easy to intercept the real target in the front-end processing when the light is low.

[0124] This implementation takes into account both false alarm rate and illuminance, achieving a better balance between the two; the stop processing threshold can be updated based on historical false alarm rate and current ambient illuminance, avoiding the long-term use of a fixed single threshold in complex lighting scenarios;

[0125] The preset basic discrete threshold, false negative compensation coefficient, and burst suppression coefficient are obtained, and the false negative rate and continuous trigger rate are retrieved. Based on the false negative rate and continuous trigger rate, the preprocessing blocking threshold is adjusted synchronously to generate the updated preprocessing blocking threshold. Among them, the basic discrete threshold represents the starting threshold for the image sensor to determine whether the motion is sufficiently concentrated in a standard scene.

[0126] The false alarm compensation coefficient is used to control the adjustment ratio of the false alarm rate to the relaxation of the previous stage; the burst suppression coefficient is used to control the adjustment ratio of the continuous trigger rate to the tightening of the previous stage; the false alarm rate represents the proportion of undetected results obtained by comparing the preset periodic full-frame scan detection results with the actual output results of the system; the continuous trigger rate represents the degree of high-frequency wake-up at the sensor end within the preset statistical window, and is used to determine whether a burst of false triggers occurs.

[0127] The specific calculation rules for the false alarm rate and the continuous trigger rate are as follows: For the calculation of the false alarm rate, the control management unit counts the total number of real hazard sources identified by periodically forcibly starting full-frame scanning within a preset statistical period, such as the past hour, as the baseline denominator, and counts the number of targets that were not triggered by the normal output alarm due to the blockage in the initial screening or pre-processing stage, but were found in the baseline of the full-frame scanning as the numerator. The ratio of the two is the false alarm rate of the current statistical period.

[0128] For example, if a full-frame scan detects 20 real targets within one hour, but the system correctly identifies only 18 and the remaining 2 were falsely blocked by the upstream system, then the false negative rate is [missing information]. For the calculation of the continuous trigger rate, the system counts the total number of times the sensor generates a level 1 wake-up flag within a preset sliding time window of, for example, the past 5 minutes. This total number is divided by, for example, a preset normal wake-up baseline of 50 times, and the ratio is truncated to 1 when it exceeds 1. The resulting ratio is the continuous trigger rate.

[0129] If it is actually triggered 40 times within 5 minutes, then the consecutive trigger rate is... This indicates that the system is currently in a high-frequency wake-up state and the front-end blocking needs to be upgraded; multiply the missed detection compensation coefficient by the missed detection rate to obtain the missed detection compensation term;

[0130] The burst suppression coefficient is multiplied by the continuous trigger rate to obtain the burst suppression term; the basic discrete threshold is added to the missed detection compensation term and the burst suppression term is subtracted. The updated preprocessing blocking threshold is directly calculated by arithmetic combination. This threshold is used as the dynamic judgment benchmark for joint judgment at the image sensor end in the current system cycle.

[0131] When the false alarm rate increases, it indicates that the pre-stage blocking ratio is higher than the preset benchmark, and negative compensation by lowering the threshold should be performed; when the continuous trigger rate increases, it indicates that the pre-stage false trigger frequency is higher than the preset wake-up frequency, and positive compensation by raising the threshold should be performed; assuming the system's basic discrete threshold is set to 2.0, the false alarm compensation coefficient is 5.0, and the burst suppression coefficient is 1.0;

[0132] If the false negative rate for the current statistical period is 0.1, then the calculated false negative compensation term is 0.5; if the recent continuous trigger rate is 0.8, then the burst suppression term is 0.8; according to the calculation logic, the updated preprocessing blocking threshold = 2.0 + 0.5 - 0.8 = 1.7; through this dynamic calculation, the system successfully tightens the judgment threshold when facing high-frequency burst false triggers.

[0133] This implementation also incorporates the blocking threshold at the front-end image sensor for adjustment, enabling synchronous changes in the processing intensity between the front and back stages, reducing mismatches caused by overly strict front-end or overly lenient back-end; the preprocessing blocking threshold can be synchronously adjusted based on the false negative rate and continuous trigger rate, preventing the system from only correcting the back-end while ignoring the front-end;

[0134] The initial detection area set and the high-definition reading area set are obtained. Based on the quantification results of the coordinate difference, the area expansion parameters are dynamically adjusted to generate updated area expansion parameters. Among them, the target positioning deviation rate is used to characterize the degree of offset between the low-resolution initial detection area and the high-resolution actual reading area.

[0135] The deviation rate can be calculated using the cross-union ratio or the center point distance. If the cross-union ratio is used, the lower the ratio, the lower the overlap rate of the regions. If the center point distance is used, the greater the distance, the greater the positional offset.

[0136] The system uses the intersection-union ratio (IUU) to calculate the target positioning deviation rate as follows: the control management unit extracts the area of ​​the low-resolution initial inspection area and maps it to the high-resolution coordinate frame area, and the final output of the re-inspection unit is the area of ​​the precise outer frame of the hazard source. The IUU is obtained by calculating the ratio of the overlapping intersection area of ​​the two to the union area of ​​the two.

[0137] The system directly subtracts the intersection-union ratio (IU) from the constant 1, and the difference is taken as the target positioning deviation rate. For example, if the IU of the mapped initial detection box and the final accurate re-detection box is calculated to be 0.6, then the target positioning deviation rate is... This indicates a 40% offset deviation; the basic outward expansion boundary represents the default outward expansion boundary width after the region is mapped.

[0138] The positioning compensation coefficient is used to control the proportion of the influence of positioning deviation on the expansion boundary increment; the positioning compensation coefficient is multiplied by the target positioning deviation rate to obtain the expansion increment; then the increment is added to the basic expansion boundary, and the result obtained by addition is directly used as the updated regional expansion parameter;

[0139] This expansion parameter is used as a dynamic boundary value for expanding the high-resolution read area in the current system cycle; when the positioning deviation of the low-resolution area increases, this value increases accordingly, in order to retain more context and reduce target truncation during subsequent reads;

[0140] Hazards on construction sites are often located in local areas. If the size of the low-resolution initial inspection box does not meet the requirements or its position is deviated, subsequent high-resolution local reading is prone to truncation and loss of target edge image information, making it difficult for the re-inspection unit to make stable judgments.

[0141] This implementation method corrects the outer boundary by reversing the positioning deviation rate, so that the high-resolution reading range can change with the recent offset. The region expansion parameters can be dynamically adjusted based on the target positioning deviation rate to avoid the truncation problem caused by the long-term fixed reading range.

[0142] In actual deployment, the control and management unit can write the hazard source judgment value, ambient illuminance, continuous trigger rate, target positioning deviation rate and false alarm rate of the current event into the parameter update area after the end of each system cycle, and update the stop processing threshold, preprocessing blocking threshold and area expansion parameters accordingly.

[0143] In the next system cycle, the initial screening unit, the sensor-end preprocessing unit, and the re-inspection unit read the updated parameters and continue working. As a result, the new data on site is not only used for this identification, but also directly affects the processing intensity and reading range of subsequent cycles.

[0144] Furthermore, the system also includes an early warning output unit, which generates an alarm signal and triggers an external alarm prompt when the hazard source category output by the re-inspection unit matches the preset category and the hazard source judgment value is greater than or equal to the preset alarm threshold; it does not trigger an alarm when the preset alarm threshold is not reached.

[0145] In this embodiment, the system is also equipped with an early warning output unit. This unit receives the hazard source category, hazard source judgment value and location result output by the re-inspection unit. When the hazard source category matches the preset hazard source category and the hazard source judgment value reaches the preset alarm threshold, an alarm signal is generated and the external alarm device is controlled to issue an alarm prompt signal. When the hazard source judgment value is lower than the preset alarm threshold, no alarm signal is generated, thereby avoiding low-reliability results from directly triggering external alarms.

[0146] Furthermore, based on the above implementation method, the early warning output process is explained in detail;

[0147] The system acquires the hazard source category, hazard source judgment value, and location result output by the re-inspection unit. Based on the preset hazard source category and preset alarm threshold, it determines whether to generate an alarm signal and generates an early warning output result. Among them, the preset hazard source category refers to the types of hazard sources that the system pre-sets to trigger an alarm, such as personnel, equipment, or other dangerous targets that need to be focused on in construction site monitoring; the preset alarm threshold refers to the confidence threshold used to determine whether to trigger an external alarm action; the location result refers to the regional coordinate information of the hazard source in the image, which is used for subsequent location prompts.

[0148] This implementation uses both hazard source category matching and hazard source judgment threshold as output conditions, which can intercept results that do not belong to key categories or have insufficient credibility outside the alarm link, reducing invalid alarms; it can realize early warning judgment based on dual conditions of category matching and alarm threshold, avoiding low credibility results from directly triggering external devices;

[0149] The system obtains the determination result that the alarm signal generation conditions are met, encapsulates the alarm signal according to the location result and the hazard source category, and controls the external alarm device to issue an alarm prompt signal; wherein, the alarm signal contains at least the hazard source category and location result, and the location result can be used to indicate the approximate area of ​​the target in the monitoring screen;

[0150] In its specific implementation, the warning output unit can write the alarm signal into the output register or control buffer, and then the control management unit can send a control signal to the external alarm device. The external alarm device can be an audible and visual alarm device or other device that receives the alarm signal and performs a prompting action. If the hazard source judgment value is lower than the preset alarm threshold, the control management unit will keep the corresponding output status bit unchanged and will not trigger the external alarm device.

[0151] This implementation method continues to send the results to the external device control action, forming an actual output from recognition to prompt; it can realize on-site prompts based on alarm signal encapsulation and external device control, avoiding the recognition results remaining internal and failing to form an effective warning;

[0152] In actual deployment, once the re-inspection unit outputs the hazard source category, hazard source judgment value and location result that meet the conditions, the early warning output unit can generate an alarm signal within the current system cycle and control the external alarm device to execute the alarm prompt; if the conditions are not met, only the identification result is recorded and no alarm is triggered; thus, the newly generated hazard source identification result on site can be directly converted into an executable early warning action.

[0153] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A low-power intelligent camera recognition system for detecting hazardous sources at construction sites, characterized in that, include: The system includes a pyroelectric infrared sensor, an image sensor, an image signal processor, memory, and a recognition processor; a data bus is used to connect the image sensor and the recognition processor. as well as, The sensor-end preprocessing unit, integrated into the image sensor end, is used to perform motion energy statistics based on the continuous frame image data output by the image sensor to generate an active area after receiving the micro wake-up signal output by the pyroelectric infrared sensor, and output a first-level wake-up flag according to preset feature judgment conditions. The initial screening unit, deployed on the recognition processor, is used to acquire the low-resolution global image output by the image sensor and the active area after receiving the first-level wake-up flag, and extract the regional image features to evaluate the background interference value and the target determination value; when the background interference value does not exceed the preset background interference threshold and the target determination value reaches the target existence threshold, the unit outputs the second-level wake-up flag and the initial detection area set. The re-inspection unit is activated upon receiving the secondary wake-up flag, retrieves high-resolution image data of the corresponding region of the initial inspection region set, performs hazard source category analysis, and outputs hazard source category, hazard source judgment value, and location result. The control and management unit is used to uniformly manage the state switching of each component of the system, terminate the subsequent identification operation when the background interference value exceeds the preset background interference threshold or the target judgment value does not reach the target existence threshold; and is used to adjust the historical sample records and system judgment thresholds based on the identification results output by the re-inspection unit.

2. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 1, characterized in that, The sensor-end preprocessing unit is configured to: perform pixel-level difference based on the continuous frame image data to generate a motion mask, mesh the current global image corresponding to the continuous frame image data to count the mesh motion energy, calculate the motion discrete value characterizing the motion energy distribution law, and perform connected component aggregation operation based on the motion mask to obtain the active region.

3. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 2, characterized in that, The specific logic for the sensor-end preprocessing unit to output a first-level wake-up flag based on preset feature determination conditions is as follows: When the motion discrete value is less than the preset preprocessing blocking threshold, and the maximum area of ​​the active region and the number of active regions both meet the preset minimum effective area threshold and the upper limit of the number of connected regions, it is determined that there is suspected entity motion, and the first-level wake-up flag and the active region are output; otherwise, the current frame data is discarded at the image sensor and the sleep state is restored.

4. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 1, characterized in that, The primary screening unit is configured as follows: Generate a region filtering mask based on the circumscribed coordinates of the active region; Based on the region filtering mask, the low-resolution global image is subjected to background filtering to obtain target region image data; The pixel distribution features or edge features of the target region image data are extracted to calculate the background interference value and the target determination value. When the background interference value does not exceed a preset background interference threshold and the target determination value reaches the target existence threshold, the initial detection region set containing low-resolution coordinates is generated.

5. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 1, characterized in that, The re-inspection unit is configured as follows: Obtain preset scale mapping coefficients and preset region expansion parameters, map and expand the low-resolution coordinates in the initial detection region set to the original high-resolution coordinates, so as to obtain a high-definition reading region set. The hazard source category is determined based on the high-resolution image data corresponding to the set of high-definition reading areas.

6. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 1, characterized in that, The feedback adjustment function of the control management unit includes: The system determines whether an alarm is real or false based on the comparison between the hazard source determination value and the preset confirmation threshold, thereby updating the system's historical false alarm rate. Obtain the ambient illuminance parameters of the system currently in operation; Based on the historical false alarm rate and the ambient illumination parameter, the background interference threshold of the primary screening unit is dynamically adjusted.

7. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 6, characterized in that, The logic of the control management unit to dynamically adjust the background interference threshold is as follows: based on the preset basic exit threshold, compensation adjustment is performed based on the historical false alarm rate and the ambient illuminance parameter.

8. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 3, characterized in that, The logic of the control management unit to dynamically adjust the preprocessing blocking threshold is as follows: based on the preset basic discrete threshold, compensation adjustment is performed by comprehensively considering the historical false alarm rate data and continuous trigger rate data of the system.

9. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 5, characterized in that, The control and management unit is also used to compare the coordinate differences between the initial inspection area set and the high-definition reading area set to calculate the target positioning deviation rate, and dynamically update the area expansion parameters used by the re-inspection unit according to the target positioning deviation rate.

10. The low-power intelligent camera recognition system for detecting hazardous sources at construction sites according to claim 1, characterized in that, The system also includes an early warning output unit, which generates an alarm signal and triggers an external alarm notification when the hazard source category output by the re-inspection unit matches a preset category and the hazard source judgment value is greater than or equal to a preset alarm threshold; it does not trigger an alarm notification if the preset alarm threshold is not reached.

Images