Intelligent identification and real-time alarm system for video monitoring

Abstract

The invention provides an intelligent identification and real-time alarm system for video monitoring. The intelligent identification and real-time alarm system comprises a video sensor, a real-time preprocessing module, a deep learning module, a risk assessment module and an early warning module, the video sensor collects video information in an environment and sends the video information to the real-time preprocessing module, and the video sensor adopts one of a network camera IPC, an analog camera AVC and an NVR video processor; the real-time preprocessing module preprocesses the video data, the preprocessing comprises clipping, size adjustment, enhancement and denoising, standard video image formation, resolution and frame rate unification and video noise reduction, and the standard video image obtained after preprocessing is sent to the deep learning module; and the deep learning module performs target identification, target detection and abnormal behavior detection on the standard video image through a deep learning model, and analyzes the video according to the output detection target and the detection abnormal result.

Description

Technical Field

[0001] The present invention belongs to the field of identification and warning, and particularly relates to an intelligent identification and real-time alarm system for video surveillance. Background Art

[0002] The intelligent identification and real-time alarm system for video surveillance, as an important technology in the modern security surveillance field, has been widely applied and developed in recent years. However, although the system has achieved remarkable results in improving the surveillance efficiency and accuracy, there are still many deficiencies in its existing technologies.

[0003] Traditional video surveillance systems often rely on manual monitoring, which is not only inefficient but also vulnerable to fatigue, resulting in frequent false alarms and missed alarms. The introduction of intelligent identification technology has alleviated this problem to a certain extent. By automatically analyzing surveillance videos through algorithms, abnormal behaviors or targets can be identified. However, current intelligent identification algorithms still struggle to handle complex scenarios, especially in situations such as light changes, occlusions, and angle changes, where the identification accuracy will drop significantly.

[0004] Although the real-time alarm system can issue alarms in a timely manner when abnormalities are detected, there are still bottlenecks in its transmission method and processing mechanism. On the one hand, due to network latency or unstable transmission, alarm information may not be delivered to relevant personnel in a timely manner, resulting in a slow response speed; on the other hand, the data processing ability of the alarm system is limited, and it is unable to effectively analyze and mine a large amount of surveillance data, thus restricting its application in complex scenarios.

[0005] More critically, there are obvious shortcomings in the cross-departmental and cross-regional collaboration of existing video surveillance intelligent identification and real-time alarm systems. The data formats and standards between different systems are not unified, resulting in difficulties in effectively integrating and sharing data, which affects the ability to cooperate in combat. At the same time, due to the lack of a unified early warning standard and process in the system, there are problems such as untimely and inaccurate release and dissemination of early warning information, further weakening the practicality of the system.

[0006] The construction and maintenance costs of the system are also a major challenge. High-quality sensors, high-performance computing devices, and a stable network environment are the basis for the normal operation of the system, but all of these require a large amount of capital investment. In some remote or economically underdeveloped areas, due to a shortage of funds, the construction of the system lags behind and cannot meet the actual needs. The intelligent identification and real-time alarm system for video surveillance is of great significance in improving the efficiency of security surveillance, but there are still many deficiencies in its existing technologies, and continuous improvement and innovation are needed in aspects such as algorithm optimization, data transmission, data fusion, cross-departmental collaboration, and cost control. Summary of the Invention

[0007] The present invention proposes an intelligent recognition and real-time alert system for video surveillance, which solves the technical problems such as low efficiency of manual surveillance, untimely response to abnormal behaviors, and inaccurate risk assessment in traditional video surveillance systems. By integrating video sensors, real-time preprocessing, deep learning, risk assessment, and warning modules, intelligent analysis and real-time warning of surveillance videos are realized, improving the surveillance efficiency and accuracy.

[0008] The technical solution of the present invention is realized as follows: In the intelligent recognition and real-time alert system for video surveillance, data interaction is carried out among a video sensor, a real-time preprocessing module, a deep learning module, a risk assessment module, and a warning module;

[0009] The video sensor collects video information in the environment and sends it into the real-time preprocessing module. The video sensor adopts one of an IP camera (IPC), an analog camera (AVC), and an NVR video processor;

[0010] The real-time preprocessing module preprocesses the video data. The preprocessing includes clipping, size adjustment, enhancement, and denoising to form a standard video image, unify the resolution and frame rate, reduce video noise, and send the standard video image obtained after preprocessing to the deep learning module;

[0011] The deep learning module performs object recognition, object detection, and abnormal behavior detection on the standard video image through a deep learning model, analyzes the video according to the output detection targets and detection abnormal results, and sends the analysis results to the risk assessment module and the warning module respectively;

[0012] The risk assessment module analyzes the video, calculates the numerical value of the risk factor, obtains the distribution of the risk status, and sends the analysis result to the warning module;

[0013] The warning module compares the numerical value of the risk factor with the set standard threshold. When it exceeds the standard threshold, all risk points are marked and video surveillance warning information is synchronously generated and sent to the preset recipient.

[0014] Traditional video surveillance technologies mainly rely on manual surveillance and simple video recording functions, and there are many deficiencies. Manual surveillance requires a large amount of manpower and material resources, and is easily affected by factors such as fatigue and distraction of surveillance personnel, resulting in low surveillance efficiency and untimely response to abnormal behaviors. The traditional system lacks intelligent analysis functions and cannot deeply mine video content and detect abnormal behaviors. In addition, in terms of risk assessment, traditional methods often rely on empirical judgments or simple rule matching, and it is difficult to accurately assess the risk status.

[0015] Compared with the prior art, this technical solution has significant differences. In terms of data acquisition, this solution adopts advanced video sensor technologies, including network cameras IPC, analog cameras AVC, and NVR video processors, etc., which can efficiently collect high-quality video information. In the preprocessing stage, through processing means such as cropping, resizing, enhancement, and denoising, standard video images are formed, laying a foundation for subsequent intelligent analysis.

[0016] Most importantly, in terms of intelligent analysis, this solution introduces a deep learning module, which uses a deep learning model to perform object recognition, object detection, and abnormal behavior detection on standard video images, greatly improving the accuracy and efficiency of analysis. In addition, the risk assessment module calculates the values of risk factors and obtains the distribution of risk states, providing a scientific basis for early warning. The early warning module can compare the values of risk factors with the set standard thresholds, and generate early warning information in a timely manner and send it to the preset recipients when the thresholds are exceeded, realizing real-time monitoring and rapid response.

[0017] These differences make this technical solution more comprehensive in function and more superior in performance. Especially in terms of intelligent analysis and risk assessment, this solution breaks through the limitations of traditional technologies, realizes in-depth mining and accurate assessment of surveillance videos, and brings a revolutionary change to the field of video surveillance.

[0018] As a preferred embodiment, the video sensor includes at least one of a network camera IPC, an analog camera AVC, and an NVR video processor:

[0019] As a preferred embodiment, the deep learning module includes a dataset construction unit, a training unit, a testing unit, and an identification and detection unit;

[0020] The dataset construction unit establishes a dataset according to object recognition, object detection, and abnormal behavior detection, and divides it into a training set and a testing set;

[0021] The training unit performs network training based on convolutional neural network object recognition and object detection models and abnormal behavior detection models to obtain trained model parameters;

[0022] The testing unit calls the test data in the training unit and inputs it into the trained model to perform adaptive fine-tuning on the model parameters;

[0023] The identification and detection unit inputs the video processed by the real-time preprocessing module into the trained and optimized model, performs object recognition and detection and saves the detection results, and at the same time performs abnormal behavior detection and records abnormal behavior data.

[0024] As a preferred embodiment, the real-time preprocessing module is built with an enhancement unit. By enhancing the target and the contour, the enhancement and detail highlighting of the image are achieved. The enhancement unit performs convolution operations, applies a convolution kernel to slide and cover all pixels of the original image, and generates an extracted feature map. Then, by adjusting the threshold parameters, a low threshold and a high threshold are set. For the part located above the high threshold, its original gray value is retained. For the part located below the low threshold, it is set to 0. For the part between the thresholds, the proportion in the low threshold part is extracted to generate a new gray value. By setting different threshold parameters and repeating the above processing, the image features at different gray levels are matched. The output image is re-formed according to the retained part and the enhanced part.

[0025] As a preferred embodiment, the risk assessment module includes an environmental risk assessment unit and an abnormal risk assessment unit. The environmental risk assessment unit obtains the image data of illegal vehicles, illegal personnel, illegal machines, illegal wearables, and suspicious items based on the abnormal behavior detection results of the deep learning module, calculates the environmental risk factor according to the image data, and gives the distribution of the environmental risk status. The abnormal risk assessment unit obtains the number of detected abnormal behaviors based on the abnormal behavior detection results of the deep learning module. The abnormal behaviors include loitering abnormal behaviors, conflict abnormal behaviors, group behaviors, and dangerous goods transportation, calculates the abnormal risk factor, and gives the distribution of the abnormal risk status.

[0026] As a preferred embodiment, the environmental risk factor is the probability weighted value of the environmental risk status distribution. The environmental risk status distribution includes the probability of illegal vehicles, the probability of illegal personnel, the probability of illegal machines, the probability of illegal wearables, and the probability of suspicious items. The probability of illegal vehicles is the proportion of vehicles with inconsistent license plate numbers among all vehicles in the video image. The probability of illegal personnel is the proportion of personnel wearing inappropriate clothing among all dressed personnel in the video image. The probability of illegal machines is the proportion of machines not operating according to regulations among all machines in the video image. The probability of illegal wearables is the proportion of other personnel not following the dress code among all personnel in the video image. The probability of suspicious items is the proportion of people carrying illegal items among all personnel in the video image.

[0027] As a preferred embodiment, the abnormal risk factor is the probability weighted value of the abnormal risk state distribution, and the abnormal risk state distribution includes a wandering risk factor, a conflict risk factor, and a group behavior risk factor; the calculation method of the value of the wandering risk factor is that within a set time, if the target is tracked to a certain area in the video image and stays for more than a preset time, it is confirmed that the target wanders within the area; according to the number of times tracked within the set time, its wandering probability is calculated; the calculation method of the value of the conflict risk factor is to calculate the ratio of the overlapping time of the time period when the target is tracked or stationary in the video image to another target to the sum of the tracking time period and the stationary time period, and calculate its conflict probability; the calculation method of the value of the group behavior risk factor is that when the time period when the target is tracked or stationary in the video image overlaps with another target, and the number of tracked targets is greater than one, and the overlapping time period exceeds the preset time, it is confirmed that the target exhibits group behavior; according to the number of times of group behavior within the set time, the group behavior probability is calculated.

[0028] After adopting the above technical solutions, the beneficial effects of the present invention are as follows: by introducing an intelligent analysis and real-time warning mechanism, the burden of manual monitoring is greatly reduced, and the monitoring efficiency is improved. At the same time, the intelligent analysis can achieve a rapid response to abnormal behaviors, effectively reducing the security risks. The introduction of the deep learning model has greatly improved the accuracy of target recognition and abnormal behavior detection. Compared with traditional methods, the deep learning model has stronger adaptability and generalization ability, and can better adapt to the complex and changeable monitoring environment.

[0029] This solution obtains the distribution of the risk state by calculating the numerical values of the risk factors, providing a scientific basis for early warning. This method is more objective and accurate than traditional empirical judgment or rule matching, and can better reflect the actual situation. Through the real-time monitoring and rapid response mechanism of this technical solution, users can timely understand the risk status of the monitored area and take corresponding countermeasures. At the same time, the generation and sending of early warning information also greatly improve the convenience and satisfaction of users.

[0030] While solving the problems existing in traditional video monitoring technologies, this technical solution also brings many beneficial effects. It performs excellently in terms of monitoring efficiency, accuracy, risk assessment, and user experience, injecting new vitality into the development of the video monitoring field. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0032] Figure 1 This is the system flow block diagram of the present invention. Specific implementation manners

[0033] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present invention.

[0034] Embodiment:

[0035] As Figure 1 shown, the intelligent recognition and real-time alarm system for video surveillance is a comprehensive system integrating advanced video acquisition, real-time data processing, deep learning algorithms, risk assessment models, and early warning mechanisms. Through the close cooperation among various modules, the system realizes the intelligent recognition of target objects in surveillance videos, the detection of abnormal behaviors, the assessment of risk states, and the real-time sending of early warning information, greatly improving the intelligent level and response speed of video surveillance.

[0036] The system first collects video information in the environment through video sensors, namely network cameras IPC, analog cameras AVC, or NVR video processors. These video sensors can capture the pictures of the surveillance area in real time and convert them into digital signals for transmission. The network camera IPC usually supports network transmission and can be accessed and controlled remotely; the analog camera AVC needs to transmit signals through physical media such as video cables and then be converted into digital signals through a video encoder; while the NVR video processor integrates video recording and network transmission functions and can process the video signals of multiple cameras simultaneously.

[0037] The collected video data is sent to the real-time preprocessing module. This module performs a series of preprocessing operations on the video data, including cropping, size adjustment, enhancement, and denoising, etc. The cropping operation can remove the unnecessary parts in the video picture, such as borders or irrelevant backgrounds; the size adjustment ensures that all video data has a unified resolution for subsequent processing; the enhancement operation can improve the clarity and contrast of the video picture; the denoising operation is used to reduce the noise interference in the video and improve the video quality. After preprocessing, the video data is converted into standard video images, providing a basis for subsequent intelligent analysis.

[0038] The standard video image is sent to the deep learning module. This module uses a deep learning model to perform object recognition, object detection, and abnormal behavior detection on the video image. Object recognition can identify specific objects in the video, such as people, vehicles, animals, etc.; object detection can locate the positions of these objects in the video frame; abnormal behavior detection can identify abnormal behaviors that do not conform to the normal behavior pattern, such as people breaking into restricted areas, missing items, etc. The deep learning model can learn the characteristics of these behaviors by training a large number of sample data and achieve accurate analysis of new video data. The analysis results, including the detected object types, positions, and abnormal behavior types, are sent to the risk assessment module and the early warning module.

[0039] The risk assessment module receives the analysis results from the deep learning module and assesses the risk status in the video. This module calculates the value of the risk factor to obtain the distribution of the risk status. The risk factors may include the types, quantities, speeds, directions of the target objects, etc., as well as the types, durations, and influence scopes of the abnormal behaviors. The risk assessment module calculates the risk level and possible impacts based on these risk factors and in combination with a preset risk assessment model. The assessment results are sent to the early warning module.

[0040] Suppose a specific working scenario is a large industrial park. There are multiple surveillance cameras installed in the park to monitor the security situation in the park. The park management department hopes to improve the security management level of the park by introducing an intelligent recognition and real-time alarm system.

[0041] System deployment and initialization: First, install video sensors, such as network cameras IPC, in key areas of the park. These cameras can capture the images in the park in real time and transmit them to the real-time preprocessing module of the system through the network. At the same time, perform initialization settings on the system, including configuring camera parameters, setting early warning thresholds, defining risk factors, etc.

[0042] Video acquisition and preprocessing: After the system starts, the video sensors begin to collect video information in the park. These video data are transmitted to the real-time preprocessing module in real time. The preprocessing module performs operations such as cropping, resizing, enhancing, and denoising on the video data to form standard video images. These images have a unified resolution and frame rate, reducing the interference of video noise and providing a high-quality data basis for subsequent intelligent analysis.

[0043] Deep learning analysis: The standard video images are sent to the deep learning module. This module uses the trained deep learning model to perform object recognition, object detection, and abnormal behavior detection on the video images. For example, the system can identify target objects such as people and vehicles in the park and locate their positions in the video frame. At the same time, the system can also detect abnormal behaviors such as people breaking into restricted areas and vehicles speeding. These analysis results are sent to the risk assessment module and the early warning module in real time.

[0044] Risk assessment and early warning: The risk assessment module receives the analysis results from the deep learning module and evaluates the risk status in the park. For example, when the system detects that someone has broken into a restricted area, the risk assessment module will calculate the risk level and possible impacts based on risk factors such as the number, speed, and direction of the intruders. When the risk level exceeds the preset threshold, the early warning module will trigger the early warning mechanism. The early warning information, including the specific location of the intruders, the risk level, and the recommended countermeasures, is sent to the security management personnel in the park in real time.

[0045] Countermeasures and feedback: After receiving the early warning information, the security management personnel immediately take countermeasures. For example, they dispatch security personnel to the scene to deal with the intruders and activate the emergency plan to ensure the safety of the park. During the handling process, the security management personnel can view the live video of the scene in real time through the system to understand the progress of the handling. After the handling is completed, the security management personnel can feedback the handling results to the system so that the system can evaluate and optimize the early warning effect.

[0046] Through this workflow, the intelligent recognition and real-time alert system of video surveillance can achieve real-time monitoring and early warning of the security status in the park. The system can automatically identify abnormal behaviors, evaluate the risk status, and trigger the early warning mechanism, providing strong technical support for the security management of the park. At the same time, the real-time and intelligent level of the system also greatly improves the work efficiency and response speed of the security management personnel.

[0047] The video sensor includes at least one of an IP camera (IPC), an analog camera (AVC), and an NVR video processor. In an intelligent security monitoring system, the video sensor, as a key component for information collection, its diversity and flexibility are crucial. For example, in a large industrial park, in order to comprehensively monitor multiple key areas such as entrances and exits, production lines, and storage areas, IP cameras (IPC) for high-definition remote monitoring, analog cameras (AVC) for cost-sensitive or areas with incomplete network coverage, and NVR (Network Video Recorder) video processors can be deployed to centrally manage and store video data. This combination not only improves the monitoring coverage but also enables efficient data management and remote access capabilities through the NVR.

[0048] Combining multiple types of video sensors to meet the monitoring requirements of different scenarios, improving the adaptability and flexibility of the system. The application of NVR realizes the unified storage and management of video data, simplifies system maintenance, and supports remote access, enhancing the monitoring efficiency. Flexibly select sensor types according to the importance and security level of different regions, optimizing the cost investment.

[0049] The deep learning module includes a dataset construction unit, a training unit, a testing unit, and an identification and detection unit;

[0050] The dataset construction unit establishes a dataset based on object recognition, object detection, and abnormal behavior detection, and divides it into a training set and a testing set;

[0051] The training unit conducts network training based on the convolutional neural network object recognition and object detection models and the abnormal behavior detection model to obtain the trained model parameters;

[0052] The testing unit calls the test data in the training unit and inputs it into the trained model to perform adaptive fine-tuning on the model parameters;

[0053] The identification and detection unit inputs the video processed by the real-time preprocessing module into the trained and optimized model for object recognition and detection, saves the detection results, and simultaneously conducts abnormal behavior detection and records the abnormal behavior data.

[0054] In the traffic management of smart cities, the deep learning module is used in intelligent traffic monitoring systems. By constructing a multi-category dataset including vehicles, pedestrians, traffic signs, etc., the model is trained to identify traffic violations (such as running red lights, driving over the line), abnormal behaviors (such as pedestrians entering the lane), vehicle types, license plate recognition, etc. After preprocessing, the real-time video data is directly input into the trained model for instant detection and recognition, and the results are saved for subsequent analysis or law enforcement basis. High degree of automation: From dataset construction to model training, testing, and then to actual application, the whole process is highly automated, reducing manual intervention and improving processing efficiency. Accurate identification: The model based on convolutional neural network has higher accuracy and robustness for object recognition and abnormal behavior detection in complex scenarios. Continuous optimization: The fine-tuning mechanism of the model parameters by the testing unit ensures that the model can continuously maintain high performance when facing new scenarios or data.

[0055] The real-time preprocessing module is built-in with an enhancement unit. By enhancing the target and contour, image enhancement and detail highlighting are achieved. The enhancement unit applies a convolution kernel to slide and cover all pixels of the original image through convolution operations to generate an extracted feature map. Then, by adjusting the threshold parameters, a low threshold and a high threshold are set. For the part located above the high threshold, its original gray value is retained. For the part located below the low threshold, it is set to 0. For the part between the thresholds, the proportion in the low threshold part is extracted to generate a new gray value. By setting different threshold parameters and repeating the above process, the image features at different gray levels are matched. The output image is re-formed according to the retained part and the enhanced part.

[0056] The risk assessment module includes an environmental risk assessment unit and an abnormal risk assessment unit. The environmental risk assessment unit obtains the image data of illegal vehicles, illegal personnel, illegal machines, illegal wearables, and suspicious items based on the abnormal behavior detection results of the deep learning module, calculates the environmental risk factor according to the image data, and gives the distribution of the environmental risk status. The abnormal risk assessment unit obtains the number of detected abnormal behaviors based on the abnormal behavior detection results of the deep learning module. The abnormal behaviors include loitering abnormal behaviors, conflict abnormal behaviors, group behaviors, and dangerous goods transportation, calculates the abnormal risk factor, and gives the distribution of the abnormal risk status.

[0057] In a monitoring environment at night or under low light conditions, the real-time preprocessing module significantly improves the video quality through image enhancement technology. For example, in the night monitoring of a port terminal, the enhancement unit highlights the details of the ship's contour and personnel activities. Even in the absence of sufficient light, the target can be clearly identified, effectively preventing security incidents such as theft and illegal intrusion. Dynamic threshold adjustment: By setting a low threshold and a high threshold, and dynamically adjusting the threshold parameters, fine matching of image features at different gray levels is achieved, improving the flexibility and effect of image enhancement. Detail highlighting: Focusing on the enhancement of the target and contour not only improves the overall brightness of the image, but more importantly, retains important details, which is beneficial for subsequent target recognition and abnormal behavior detection.

[0058] The environmental risk factor is the probability weighted value of the environmental risk state distribution, and the environmental risk state distribution includes the probability of a vehicle violation, the probability of a personnel violation, the probability of a machine violation, the probability of a dressing violation, and the probability of a suspicious item; the probability of a vehicle violation is the proportion of vehicles with inconsistent license plate numbers among all vehicles in the video image; the probability of a personnel violation is the proportion of personnel in proper attire who do not meet the dressing requirements among all attired personnel in the video image; the probability of a machine violation is the proportion of machines that do not operate according to regulations among all machines in the video image; the probability of a dressing violation is the proportion of other personnel who do not comply with the dressing regulations among all personnel in the video image; the probability of a suspicious item is the proportion of people carrying illegal items among all personnel in the video image.

[0059] In the security management of large transportation hubs such as airports and railway stations, the risk assessment module can, based on the abnormal behavior detection results of the deep learning module, evaluate the environmental risk and abnormal risk in real time. For example, by identifying illegally carried items, suspicious loitering personnel, conflict behaviors, etc., it can quickly respond and take measures to effectively prevent security incidents such as terrorist attacks and thefts. By combining the environmental risk and abnormal risk, a comprehensive assessment is carried out from multiple dimensions such as vehicle violations, personnel, machines, dressing to abnormal behaviors, providing a more comprehensive and detailed risk view. Quantifying the risk state through the probability weighted value makes the risk assessment more scientific and accurate, providing strong support for decision-making.

[0060] The abnormal risk factor is the probability weighted value of the abnormal risk state distribution, and the abnormal risk state distribution includes the loitering risk factor, the conflict risk factor, and the group behavior risk factor; the value calculation method of the loitering risk factor is that within a set time, if the target is tracked to a certain area in the video image and stays for more than the preset time, it is confirmed that the target loiters in the area; according to the number of times the target is tracked within the set time, its loitering probability is calculated; the value calculation method of the conflict risk factor is to overlap the time period when the target is tracked or stationary in the video image with another target, and calculate the ratio of the overlapping time to the sum of the tracking time period and the stationary time period to calculate its conflict probability; the value calculation method of the group behavior risk factor is that when the time period when the target is tracked or stationary in the video image overlaps with another target, and the number of tracked targets is greater than one and the overlapping time period exceeds the preset time, it is confirmed that the target exhibits group behavior; according to the number of times of group behavior within the set time, the group behavior probability is calculated.

[0061] In the safety monitoring of factories, environmental risk factors such as the probability of violating vehicles and the probability of violating personnel can timely detect and warn of potential safety hazards, such as unauthorized vehicles entering restricted areas and employees not wearing protective clothing as required, effectively preventing accidents such as chemical leaks and fires. In campus safety monitoring, abnormal risk factors such as the wandering risk factor and the conflict risk factor can identify and warn of abnormal behaviors such as conflicts among students and the wandering of off-campus personnel, and take timely measures to ensure campus safety. By refining risk factors, accurate identification and quantitative assessment of different types of risks are achieved, improving the pertinence and effectiveness of risk management. By refining risk factors, accurate identification and quantitative assessment of different types of risks are achieved, improving the pertinence and effectiveness of risk management.

[0062] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. Intelligent identification and real-time alarm system for video surveillance, characterized by: Including data interaction between video sensors, real-time preprocessing modules, deep learning modules, risk assessment modules and early warning modules; The video sensor collects video information in the environment and sends it to the real-time preprocessing module; The real-time preprocessing module preprocesses the video data, including cropping, resizing, enhancement and denoising, to form a standard video image, unify the resolution and frame rate to reduce video noise, and send the standard video image obtained after preprocessing to the deep learning module; The deep learning module performs target recognition, target detection and abnormal behavior detection on standard video images through a deep learning model, analyzes the video according to the output detection target and detection abnormality results, and sends the analysis results to the risk assessment module and the early warning module respectively; The risk assessment module analyzes the video, calculates the value of the risk factor, obtains the distribution of the risk status, and sends the analysis results to the early warning module; The warning module compares the value of the risk factor with the set standard threshold. When the standard threshold is exceeded, all risk points are marked and video surveillance warning information is synchronously generated and sent to the preset recipient.

2. The intelligent identification and real-time alarm system for video surveillance according to claim 1 is characterized in that: The video sensor includes at least one of an IP camera IPC, an analog camera AVC and an NVR video processor.

3. The intelligent identification and real-time alarm system for video surveillance as claimed in claim 1, characterized in that: The deep learning module includes a data set construction unit, a training unit, a testing unit, and a recognition and detection unit; The data set construction unit establishes a data set according to target recognition, target detection, and abnormal behavior detection, and divides the data set into a training set and a test set; The training unit performs network training based on the convolutional neural network target recognition and target detection model and the abnormal behavior detection model to obtain trained model parameters; The testing unit calls the test data in the training unit and inputs it into the trained model to perform adaptive fine-tuning on the model parameters; The recognition and detection unit inputs the video processed by the real-time preprocessing module into the trained and optimized model, performs target recognition and detection and saves the detection results, and simultaneously performs abnormal behavior detection and records abnormal behavior data.

4. The intelligent identification and real-time alarm system for video surveillance as claimed in claim 1, characterized in that: The real-time preprocessing module is built with an enhancement unit, which enhances the target and the contour to achieve image enhancement and detail highlighting. The enhancement unit uses a convolution kernel to slide and cover all pixels of the original image through a convolution operation to generate an extracted feature map; then, the low threshold and the high threshold are set by adjusting the threshold parameters; for the part located at the high threshold, its original grayscale value is retained, and for the part located at the low threshold, it is set to 0; for the part between the thresholds, the proportion of the low threshold part is extracted to generate a new grayscale value; the above process is repeated by setting different threshold parameters to match image features of different grayscale levels; The output image is re-formed from the retained portion and the enhanced portion.

5. The intelligent identification and real-time alarm system for video surveillance as claimed in claim 1, characterized in that: The risk assessment module includes an environmental risk assessment unit and an abnormal risk assessment unit; the environmental risk assessment unit obtains image data of illegal vehicles, illegal persons, illegal machines, illegal clothing and suspicious items based on the abnormal behavior detection results of the deep learning module, and calculates environmental risk factors based on the image data to give the distribution of environmental risk states; the abnormal risk assessment unit obtains the number of abnormal behaviors detected based on the abnormal behavior detection results of the deep learning module, and the abnormal behaviors include wandering abnormal behaviors, conflicting abnormal behaviors, group behaviors and dangerous goods transportation, calculates abnormal risk factors, and gives the distribution of abnormal risk states.

6. The intelligent identification and real-time alarm system for video surveillance as claimed in claim 5, characterized in that: The environmental risk factor is a probability-weighted value of the environmental risk state distribution, and the environmental risk state distribution includes the probabilities of illegal vehicles, illegal persons, illegal machines, illegal wearing and suspicious items; the illegal vehicle probability is the proportion of vehicles with inconsistent license plates among all vehicles in the video image; the illegal person probability is the proportion of all dressed persons in the video image who are dressed in a way that does not comply with the dress code requirements; the illegal machine probability is the proportion of all machines that do not operate in accordance with regulations in the video image; the illegal wearing probability is the proportion of other persons who do not comply with the dress code regulations among all persons in the video image; the suspicious item probability is the proportion of all persons carrying illegal items in the video image.

7. The intelligent identification and real-time alarm system for video surveillance as claimed in claim 5, characterized in that: The abnormal risk factor is a probability weighted value of the abnormal risk state distribution, and the abnormal risk state distribution includes a wandering risk factor, a conflict risk factor, and a group behavior risk factor; the wandering risk factor value is calculated by tracking the target in a certain area in the video image within a set time, and staying for more than a preset time, and confirming that the target is wandering in the area; according to the number of times it is tracked within the set time, the wandering probability is calculated; The method for calculating the value of the conflict risk factor is to overlap the time period of tracking or stillness of the target in the video image with another target, calculate the ratio of the overlapping time to the sum of the tracking time period and the stillness time period, and calculate the conflict probability; The method for calculating the value of the group behavior risk factor is to determine that the target has group behavior when the time period during which the target is tracked or stationary in the video image overlaps with another target, and when the number of tracked targets is greater than one and the overlapping time period exceeds a preset time; and calculate the group behavior probability based on the number of group behaviors within the set time.

Images