Dynamic image real-time recognition alarm system combined with artificial intelligence and machine learning

Abstract

This invention relates to the field of image recognition technology, and more particularly to a dynamic image real-time recognition alarm system combining artificial intelligence and machine learning. The system includes a dynamic recognition alarm center, an image acquisition module, an initial analysis module, a machine learning optimization module, a false alarm elimination module, an alarm decision module, and a remote alarm module. This invention collects dynamic image data from a monitored area and generates standardized image data through preprocessing. Simultaneously, it performs target recognition and preliminary screening on the standardized image data to construct a labeling library and a normal sample library, providing high-quality input for subsequent recognition. Through a closed-loop process of "dynamic sample screening - distributed incremental training - accuracy verification," it achieves continuous optimization of the AI ​​recognition model. Furthermore, by constructing a four-dimensional confidence evaluation system of "environmental parameters + historical data + model accuracy + initial recognition" through data collaboration, it reduces the false alarm rate from the source, solving the problems of single-threshold alarms and false alarms in existing systems.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a dynamic image real-time recognition alarm system that combines artificial intelligence and machine learning. Background Technology

[0002] With the acceleration of urbanization and the upgrading of security needs, dynamic image monitoring systems have been widely used in smart communities, industrial plants, transportation hubs, public places and other scenarios, becoming a core component of the security system. The traditional monitoring system, which relies on manual duty and post-event tracing, can no longer meet the needs of "real-time early warning and proactive prevention".

[0003] However, these systems have revealed many technical shortcomings in practical applications, making it difficult to meet the precise and real-time security needs in complex scenarios. Specific problems are as follows:

[0004] Firstly, the model has poor adaptability and a high false alarm rate: fixed models are difficult to cope with dynamic interference in monitoring scenarios, such as changes in lighting (e.g., strong light, backlight, nighttime) and target occlusion (e.g., tree occlusion, crowd occlusion), which leads to unstable recognition accuracy. At the same time, the lack of comprehensive analysis of environmental parameters and historical alarm data makes it easy to misjudge normal behavior as abnormal. A large number of false alarms increase the workload of on-duty personnel and reduce the reliability of the system.

[0005] Secondly, the sample utilization rate is low: no dynamic sample library has been established, and low-confidence recognition results cannot be converted into effective training data. The model cannot be iteratively optimized over time, and the recognition performance is prone to degradation after long-term use.

[0006] To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a real-time dynamic image recognition alarm system that combines artificial intelligence and machine learning to solve the aforementioned technical deficiencies.

[0008] The objective of this invention can be achieved through the following technical solution: a dynamic image real-time recognition alarm system combining artificial intelligence and machine learning, including a dynamic recognition alarm center, an image acquisition module, an initial analysis module, a machine learning optimization module, a false alarm elimination module, an alarm decision module, and a remote alarm module;

[0009] The image acquisition module is used to acquire dynamic image data of the monitored area and preprocess the dynamic image data to obtain standardized image data.

[0010] The initial analysis module is used to perform target recognition and segmentation analysis on standardized image data, and obtain the annotation library and normal sample library based on the obtained exclusion signal or abnormal signal;

[0011] The machine learning optimization module is used to dynamically optimize and evaluate the AI ​​recognition model, obtain a temporary updated model, analyze the training effectiveness of the temporary updated model, and output the model update confidence score.

[0012] The false alarm elimination module is used to perform multi-dimensional false alarm confidence assessment and analysis on the alarm accuracy rate within a preset historical time period to obtain the confidence level of historical data and the confidence level of environmental parameters;

[0013] The alarm decision module is used to perform multi-dimensional fusion alarm analysis on standardized image data and output alarm feedback list, regular signals or alarm signals.

[0014] Preferably, the analysis process of the initial analysis module is as follows:

[0015] Standardized image data is input into a pre-set AI recognition model to obtain the recognition result, which includes the target category, location coordinates, and initial confidence level.

[0016] A preliminary comparative analysis is performed based on the initial confidence level in the recognition results. If the initial confidence level is less than the preset initial confidence level threshold, an exclusion signal is generated, and the standardized image data corresponding to the exclusion signal is stored as a sample in the annotation library. If the initial confidence level is greater than or equal to the preset initial confidence level threshold, an abnormal signal is generated, and the standardized image data corresponding to the abnormal signal is stored as a sample in the normal sample library.

[0017] Preferably, the analysis process of the machine learning optimization module is as follows:

[0018] S1: Set the standardized image data corresponding to the abnormal signal as a high-confidence suspicious frame;

[0019] S2: Construct transmission data packets based on high-confidence suspicious frames;

[0020] S3: Simultaneously push the annotation library to the annotators, correct target category errors and position calibration of the samples in the annotation library, obtain the corrected samples and store them in the valid sample library;

[0021] S4: Divide the effective sample library according to the target dimension, scenario dimension, and time dimension.

[0022] Preferably, it also includes S41: stratified sampling from the effective sample library according to scene type, extracting i samples to construct an incremental training set, where i is a natural number greater than zero;

[0023] S42: Distribute the incremental training set to 3 local training nodes. Each node trains the AI ​​recognition model based on the incremental training set to obtain a temporary updated model.

[0024] S43: Randomly select k samples from the normal sample library to construct a validation set, where k is a natural number greater than zero, and test the recognition accuracy of the temporarily updated model;

[0025] S44: Obtain the value obtained by subtracting the recognition accuracy before the update from the updated recognition accuracy, and set the value obtained by subtracting the recognition accuracy before the update from the updated recognition accuracy as the update improvement value. Perform discrimination processing on the update improvement value to obtain an effective signal or retraining instruction.

[0026] Responding to the retraining instruction, the temporarily updated model is retrained until a valid signal is generated;

[0027] S45: When a valid signal is generated, the updated recognition accuracy is set as the model update confidence level based on the temporary update model.

[0028] Preferably, the analysis process of the false alarm elimination module is as follows:

[0029] T1: Based on the recognition results, obtain the alarm accuracy rate for the same scene and the same behavior type within a historical preset time period;

[0030] T2: Set the alarm accuracy to the confidence level of historical data;

[0031] T3: Simultaneously acquire the illumination intensity of the standardized image data, and perform discrimination processing on the illumination intensity and preset illumination intensities Gmax and Gmin to obtain the normalized value of the illumination intensity.

[0032] Preferably, it also includes T4: obtaining the area of ​​the occluded region of the target and the overall area of ​​the target based on standardized image data;

[0033] T5: Set the ratio of the area of ​​the occluded region of the target to the total area of ​​the target as the target occlusion rate;

[0034] T6: Obtain the weight coefficient a corresponding to the normalized value of illumination intensity and the weight coefficient b corresponding to the target occlusion rate. Both a and b are greater than zero. Calculate the confidence level of the environmental parameters based on the normalized value of illumination intensity × weight coefficient a + (1 - target occlusion rate) × weight coefficient b.

[0035] Preferably, the analysis process of the alarm decision module is as follows:

[0036] Obtain the confidence scores of environmental parameters, model update, historical data, and initial confidence scores. Set the final confidence score as the value calculated by initial confidence score × weight coefficient c1 + model update confidence score × weight coefficient c2 + historical data confidence score × weight coefficient c3 + environmental parameter confidence score × weight coefficient c4.

[0037] The final confidence level is then processed to obtain an alarm feedback list or monitoring signal.

[0038] Preferably, when generating a monitoring signal, a curve of the change in final confidence level is constructed based on the time series. Based on the curve of the change in final confidence level, g final confidence levels are randomly selected, where g is a natural number greater than 3. The average value of the selected final confidence level is obtained, and the average value of the final confidence level is processed to obtain a regular signal or an alarm signal.

[0039] The beneficial effects of this invention are as follows:

[0040] (1) This invention collects dynamic image data of the monitoring area and generates standardized image data through preprocessing to provide high-quality input for subsequent identification and analysis. At the same time, it performs target identification and preliminary screening on the standardized image data, constructs a library to be labeled and a normal sample library, and achieves continuous optimization of the AI ​​recognition model based on abnormal signals and manually corrected samples and through a closed-loop process of "dynamic sample screening - distributed incremental training - accuracy verification", avoiding model forgetting and computing power waste.

[0041] (2) This invention constructs a four-dimensional confidence assessment system of "environmental parameters + historical data + model accuracy + initial identification" through data collaboration, which reduces the false alarm rate from the source, solves the problem of single threshold alarm and false alarm in existing systems, and helps to provide core technical support for the transformation of the security monitoring industry from "passive defense" to "active intelligent early warning". Attached Figure Description

[0042] The invention will now be further described with reference to the accompanying drawings;

[0043] Figure 1 This is a flowchart of the system of the present invention;

[0044] Figure 2 This is an analysis diagram of the effective sample library of the present invention;

[0045] Figure 3 This is a partial analytical diagram of Embodiment 1 of the present invention;

[0046] Figure 4 This is a partial analysis diagram of Embodiment 2 of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments;

[0049] Example 1: Please refer to Figures 1 to 4 As shown, the present invention is a dynamic image real-time recognition alarm system combining artificial intelligence and machine learning, including a dynamic recognition alarm center, an image acquisition module, an initial analysis module, a machine learning optimization module, a false alarm elimination module, an alarm decision module, and a remote alarm module;

[0050] The dynamic identification alarm center has a one-way communication connection with the image acquisition module, the image acquisition module has a one-way communication connection with the initial analysis module, the initial analysis module has a one-way communication connection with the dynamic identification alarm center, the dynamic identification alarm center has a two-way communication connection with the machine learning optimization module, the dynamic identification alarm center has a one-way communication connection with the alarm decision module, the machine learning optimization module has a one-way communication connection with the false alarm elimination module, the false alarm elimination module has a one-way communication connection with the alarm decision module, and the alarm decision module has a one-way communication connection with the remote alarm module.

[0051] The image acquisition module is used to acquire dynamic image data of the monitored area and preprocess the dynamic image data to obtain standardized image data. The preprocessing includes noise reduction, illumination equalization, etc.

[0052] The initial analysis module is used to perform target recognition and segmentation analysis on standardized image data. The specific target recognition and segmentation analysis process is as follows:

[0053] Standardized image data is input into a pre-set AI recognition model to obtain recognition results that include information such as target category, location coordinates, and initial confidence level.

[0054] A preliminary comparison analysis is performed based on the initial confidence level in the recognition results. If the initial confidence level is less than the preset initial confidence threshold, an exclusion signal is generated, and the standardized image data corresponding to the exclusion signal is stored as a sample in the annotation library.

[0055] If the initial confidence level is greater than or equal to the preset initial confidence threshold, an abnormal signal is generated, and the standardized image data corresponding to the abnormal signal is stored as a sample in the normal sample library.

[0056] The annotation library and the normal sample library are sent to the dynamic identification alarm center for storage;

[0057] When an abnormal signal is generated, the machine learning optimization module is used to dynamically optimize and evaluate the AI ​​recognition model. The specific dynamic optimization and evaluation process is as follows:

[0058] S1: Set the standardized image data corresponding to the abnormal signal as a high-confidence suspicious frame;

[0059] S2: Construct transmission data packets based on high-confidence suspicious frames. The transmission data packets include target metadata (such as timestamps, device locations, etc.) and short time-series video clips (such as short videos from 2 seconds before to 3 seconds after the target appears).

[0060] S3: Simultaneously push the annotation library to the annotators, correct target category errors (such as misclassifying "suitcase" as "package") and position calibration (such as adjusting the bounding box coordinates to accurately surround the target) for the samples in the annotation library, obtain the corrected samples and store them in the valid sample library;

[0061] S4: Divide the effective sample library according to the target dimension (such as people, vehicles, items, etc.), scene dimension (such as light intensity: strong light / weak light / backlight; occlusion degree: no occlusion / partial occlusion / severe occlusion; time: day / night, etc.), and time dimension (such as recording the sample collection timestamp and dividing the storage interval by week / month);

[0062] S41: Stratify sampling from the effective sample library according to scene type, and extract i samples to construct an incremental training set, where i is a natural number greater than zero;

[0063] S42: Distribute the incremental training set to 3 local training nodes (such as edge computing devices), and each node trains the AI ​​recognition model based on the incremental training set to obtain a temporary updated model;

[0064] S43: Randomly select k samples from the normal sample library to construct a validation set, where k is a natural number greater than zero, and test the recognition accuracy of the temporarily updated model;

[0065] S44: Obtain the value obtained by subtracting the recognition accuracy before the update from the updated recognition accuracy, and set the value obtained by subtracting the recognition accuracy before the update from the updated recognition accuracy as the update improvement value. Perform discrimination processing on the update improvement value. If the update improvement value is greater than the preset update improvement value threshold, generate a valid signal. If the update improvement value is less than or equal to the preset update improvement value threshold, generate a retraining instruction. Immediately respond to the retraining instruction to retrain the temporary updated model until a valid signal is generated.

[0066] S45: When a valid signal is generated, the updated recognition accuracy is set as the model update confidence level based on the temporary update model, and the model update confidence level is sent to the alarm decision module for storage.

[0067] The machine learning optimization module achieves continuous optimization of the AI ​​recognition model through a closed-loop process of "precise screening of dynamic sample library - manual correction - classification and storage" and "distributed training of incremental learning - parameter aggregation - accuracy verification". This ensures the timeliness and diversity of sample data, and avoids model forgetting and computing power waste through lightweight incremental training. As a result, the recognition accuracy of the system in complex scenarios gradually improves over time, ultimately achieving a real-time alarm effect with low false alarms and high robustness.

[0068] Example 2: The false alarm elimination module is used to perform multi-dimensional false alarm confidence assessment analysis on the alarm accuracy rate within a preset historical time period. The specific multi-dimensional false alarm confidence assessment analysis process is as follows:

[0069] T1: Based on the recognition results, obtain the alarm accuracy rate for the same scene and the same behavior type (such as climbing over fences, running, etc.) within a preset historical time period;

[0070] Alarm accuracy rate = Number of correct alarms within the preset historical time period / Total number of alarms within the preset historical time period;

[0071] T2: Set the alarm accuracy to the confidence level of historical data;

[0072] T3: Simultaneously acquire the illumination intensity of the standardized image data, and perform discrimination processing on the illumination intensity and preset illumination intensities Gmax and Gmin. If the illumination intensity > Gmax, then set the illumination intensity normalization value to 1; if Gmin ≤ illumination intensity ≤ Gmax, then set the illumination intensity normalization value to 0.6; if the illumination intensity < Gmin, then set the illumination intensity normalization value to 0.1.

[0073] T4: Based on standardized image data, the area of ​​the occluded region of the target and the overall area of ​​the target are obtained;

[0074] The total area of ​​the target region is the area of ​​the bounding rectangle of the target in the pre-set AI recognition model.

[0075] The area of ​​the occluded region of the target is calculated by obtaining the areas missing from the target outline or covered by the background / other targets.

[0076] T5: The ratio between the area of ​​the occluded region of the target and the area of ​​the overall target region is set as the target occlusion rate. The target occlusion rate is between 0 and 1, including 0 and 1.

[0077] T6: Obtain the weight coefficient a corresponding to the normalized value of illumination intensity and the weight coefficient b corresponding to the target occlusion rate. Both a and b are greater than zero.

[0078] The confidence level of environmental parameters is calculated based on the normalized value of light intensity × weighting coefficient a + (1 - target occlusion rate) × weighting coefficient b.

[0079] Historical data confidence levels and environmental parameter confidence levels are sent to the alarm decision module for storage;

[0080] By dynamically collecting and analyzing real-time environmental parameters of the monitored scene, the degree of interference of the environment on the recognition accuracy is converted into a quantifiable confidence value (range 0-1). The closer the value is to 1, the less interference the current environment has on the recognition result and the higher the recognition reliability; the closer the value is to 0, the greater the environmental interference and the lower the recognition reliability.

[0081] The alarm decision module is used to perform multi-dimensional fusion alarm analysis on standardized image data. The specific multi-dimensional fusion alarm analysis process is as follows:

[0082] Obtain the confidence scores for environmental parameters, model update confidence scores, historical data confidence scores, and initial confidence scores;

[0083] The final confidence level is calculated as the value obtained by multiplying the initial confidence level by the weight coefficient c1, the model update confidence level by the weight coefficient c2, the historical data confidence level by the weight coefficient c3, and the environmental parameter confidence level by the weight coefficient c4.

[0084] The final confidence level is then processed. If the final confidence level is greater than or equal to the preset final confidence level threshold, an alarm signal is generated. An alarm feedback list is generated based on the alarm signal. The alarm feedback list includes the alarm type (such as climbing over the fence, gathering of people, etc.) and the alarm location (such as 30 meters from the fence at the east gate, etc.).

[0085] If the final confidence level is less than the preset final confidence level threshold, a monitoring signal is generated. When a monitoring signal is generated, the target is marked with a yellow box and the target's movement trajectory is displayed in real time. At the same time, a curve of the change in final confidence level is constructed based on the time series. Based on the curve of the change in final confidence level, g final confidence levels are randomly selected, where g is a natural number greater than 3. The average value of the selected final confidence level is obtained and processed. If the average final confidence level is less than the preset average final confidence level threshold, a normal signal is generated. If the average final confidence level is greater than or equal to the preset average final confidence level threshold, an alarm signal is generated.

[0086] The remote alarm module is used to respond to alarm feedback lists, regular signals, or alarm signals, and immediately display the alarm feedback list or perform preset early warning operations corresponding to regular signals or alarm signals. The preset early warning operations corresponding to regular signals are: marking the target as a low-confidence identification event in the log, including time, location, etc. The preset early warning operations corresponding to alarm signals are: pushing reminders to the mobile phones of on-duty personnel.

[0087] In summary, by collecting dynamic image data of the monitored area and generating standardized image data through preprocessing, high-quality input is provided for subsequent identification and analysis. Simultaneously, target identification and preliminary screening are performed on the standardized image data to construct a labeling library and a normal sample library. Based on abnormal signals and manually corrected samples, and through a closed-loop process of "dynamic sample screening - distributed incremental training - accuracy verification," the AI ​​recognition model is continuously optimized, avoiding model forgetting and wasted computing power. Furthermore, by constructing a four-dimensional confidence assessment system of "environmental parameters + historical data + model accuracy + initial identification" through data collaboration, the false alarm rate is reduced from the source, solving the problems of single-threshold alarms and false alarms in existing systems. This provides core technical support for the security monitoring industry's transformation from "passive defense" to "proactive intelligent early warning."

[0088] The threshold is set for result comparison and analysis to determine whether it is good or bad. The value of the threshold is determined by a combination of large-scale model analysis of sample data and human experience. It can also be adjusted appropriately based on seasonal or common-sense influencing factors.

[0089] The size of the coefficient is a specific value obtained by quantifying each parameter to facilitate subsequent comparison. The size of the coefficient depends on the amount of sample data and the corresponding operating coefficient initially set by those skilled in the art for each set of sample data; as long as it does not affect the proportional relationship between the parameter and the quantified value.

[0090] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic image real-time recognition alarm system combining artificial intelligence and machine learning, characterized in that: It includes a dynamic identification alarm center, an image acquisition module, an initial analysis module, a machine learning optimization module, a false alarm elimination module, an alarm decision module, and a remote alarm module; The image acquisition module is used to acquire dynamic image data of the monitored area and preprocess the dynamic image data to obtain standardized image data. The initial analysis module is used to perform target recognition and segmentation analysis on standardized image data, and obtain the annotation library and normal sample library based on the obtained exclusion signal or abnormal signal; The machine learning optimization module is used to dynamically optimize and evaluate the AI ​​recognition model, obtain a temporary updated model, analyze the training effectiveness of the temporary updated model, and output the model update confidence score. The false alarm elimination module is used to perform multi-dimensional false alarm confidence assessment and analysis on the alarm accuracy rate within a preset historical time period to obtain the confidence level of historical data and the confidence level of environmental parameters; The alarm decision module is used to perform multi-dimensional fusion alarm analysis on standardized image data and output alarm feedback list, regular signal or alarm signal; The analysis process of the initial analysis module is as follows: Standardized image data is input into a pre-set AI recognition model to obtain the recognition result, which includes the target category, location coordinates, and initial confidence level. A preliminary comparative analysis is performed based on the initial confidence level in the recognition results. If the initial confidence level is less than the preset initial confidence level threshold, an exclusion signal is generated, and the standardized image data corresponding to the exclusion signal is stored as a sample in the annotation library. If the initial confidence level is greater than or equal to the preset initial confidence level threshold, an abnormal signal is generated, and the standardized image data corresponding to the abnormal signal is stored as a sample in the normal sample library. The analysis process of the machine learning optimization module is as follows: S1: Set the standardized image data corresponding to the abnormal signal as a high-confidence suspicious frame; S2: Construct transmission data packets based on high-confidence suspicious frames; S3: Simultaneously push the annotation library to the annotators, correct target category errors and position calibration of the samples in the annotation library, obtain the corrected samples and store them in the valid sample library; S4: Divide the effective sample library according to the target dimension, scenario dimension, and time dimension; It also includes S41: stratified sampling from the effective sample library according to scene type, extracting i samples to construct an incremental training set, where i is a natural number greater than zero; S42: Distribute the incremental training set to 3 local training nodes. Each node trains the AI ​​recognition model based on the incremental training set to obtain a temporary updated model. S43: Randomly select k samples from the normal sample library to construct a validation set, where k is a natural number greater than zero, and test the recognition accuracy of the temporary update model; S44: Obtain the value obtained by subtracting the recognition accuracy before the update from the updated recognition accuracy, and set the value obtained by subtracting the recognition accuracy before the update from the updated recognition accuracy as the update improvement value. Perform discrimination processing on the update improvement value to obtain an effective signal or retraining instruction. Responding to the retraining instruction, the temporarily updated model is retrained until a valid signal is generated; S45: When a valid signal is generated, the updated recognition accuracy is set as the model update confidence level based on the temporary update model. The analysis process of the false alarm elimination module is as follows: T1: Based on the recognition results, obtain the alarm accuracy rate for the same scene and the same behavior type within a historical preset time period; T2: Set the alarm accuracy to the confidence level of historical data; T3: Simultaneously acquire the illumination intensity of the standardized image data, and perform discrimination processing on the illumination intensity and the preset illumination intensity Gmax and preset illumination intensity Gmin to obtain the normalized value of the illumination intensity. It also includes T4: obtaining the area of ​​the occluded region of the target and the overall area of ​​the target based on standardized image data; T5: Set the ratio of the area of ​​the occluded region of the target to the total area of ​​the target as the target occlusion rate; T6: Obtain the weight coefficient a corresponding to the normalized value of illumination intensity and the weight coefficient b corresponding to the target occlusion rate. Both a and b are greater than zero. Calculate the confidence level of the environmental parameters based on the normalized value of illumination intensity × weight coefficient a + (1 - target occlusion rate) × weight coefficient b. The analysis process of the alarm decision module is as follows: Obtain the confidence scores of environmental parameters, model update, historical data, and initial confidence scores. Set the final confidence score as the value calculated by initial confidence score × weight coefficient c1 + model update confidence score × weight coefficient c2 + historical data confidence score × weight coefficient c3 + environmental parameter confidence score × weight coefficient c4. The final confidence level is then processed to obtain an alarm feedback list or monitoring signal; When a monitoring signal is generated, a curve showing the change of the final confidence level is constructed based on the time series. Based on the curve showing the change of the final confidence level, g final confidence levels are randomly selected, where g is a natural number greater than 3. The average value of the selected final confidence levels is obtained, and the average value of the final confidence level is processed to obtain a regular signal or an alarm signal.

Images