A 5G visual alarm service method and system

Abstract

The present invention relates to the technical field of alarm systems, and specifically to a 5G visual alarm service method and system, including: a preview video image acquisition module, which acquires the video preview screen of the alarm person and collects on-site image information therefrom; a voice information acquisition module, which obtains an alarm voice information set; an alarm analysis module, which outputs the voice information analysis result through the analysis of a deep neural network; an accident evaluation module, which is used to input the voice information analysis result into the neural network and output the accident evaluation result; a sequence adjustment module, which comprehensively integrates the on-site image information and the accident evaluation result to adjust the processing order of the video alarm tasks in the queue. Through the present invention, on the one hand, an interactive service in the queue processing process is provided in the existing alarm service, and the difficulty of subsequent processing is reduced through early data collection. On the other hand, the processing order can be adjusted for more urgent situations, thereby obtaining more timely and effective processing.

Description

Technical Field

[0001] The present invention relates to the technical field of alarm systems, and in particular to a 5G visual alarm service method and system. Background Art

[0002] In the 4G era, wireless communication resources are limited and cannot fully meet the needs of comprehensive video communication. Emergency command centers cannot easily apply video communication technology to improve their emergency response capabilities. With the completion of the construction of domestic 5G networks, bandwidth bottlenecks no longer exist, the video capabilities of mobile terminals are changing with each passing day, and native video call functions have matured. Major telecom operators are widely promoting video call services, and high-definition video calls will become the mainstream call form in the 5G era. Based on the development and progress of the above technologies, the obstacles that originally restricted the command center from widely using video technology no longer exist, and the command center already has the technical possibility to enter the omnimedia era.

[0003] As for traffic accidents, they are affected by the morning and evening rush hours, and there will be obvious high-incidence periods. Video alarms can effectively improve the efficiency of accident handling during these periods. However, it is still inevitable that video alarm services will accumulate when the police capacity is limited. This situation will cause untimely accident handling and easily cause traffic congestion during the morning and evening rush hours.

[0004] How to take advantage of video alarms and optimize services during the time period when alarm services are queued for access has become a breakthrough point for further improving the capabilities of video alarm services. Summary of the invention

[0005] The present invention provides a 5G visual alarm service method and system, thereby effectively solving the problems pointed out in the background technology.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A 5G visual alarm service system is used for the phased service when the alarm person is making a video alarm and queuing for access, including:

[0008] The preview video image acquisition module obtains the video preview screen of the alarm person and collects the scene image information from it;

[0009] A voice information collection module collects voice information based on the alarm person's responses to specific questions to obtain an alarm voice information set;

[0010] An alarm analysis module, used for inputting the alarm voice information set into a deep neural network for analysis and outputting the voice information analysis result;

[0011] An accident evaluation module, used for inputting the speech information analysis result into a neural network and outputting an accident evaluation result;

[0012] The sequence adjustment module adjusts the processing order of the video alarm tasks in the queue according to the on-site image information and the accident evaluation results.

[0013] A 5G visual alarm service method is used to implement a phased service when an alarm person is making a video alarm and queuing for access, including:

[0014] Obtain the video preview of the person who called the police and collect on-site image information from it;

[0015] Collecting voice information from the alarm person by asking specific questions to obtain an alarm voice information set;

[0016] Inputting the alarm voice information set into a deep neural network for analysis, and outputting the voice information analysis result;

[0017] Inputting the speech information analysis result into a neural network for analysis, and outputting an accident assessment result;

[0018] Based on the on-site image information and accident assessment results, the processing order of the video alarm tasks in the queue is adjusted.

[0019] Furthermore, the processing order of the video alarms in the queue is adjusted based on the on-site image information and the accident assessment result, including:

[0020] Outputting a sequence adjustment requirement based on the accident evaluation result;

[0021] Comparing the on-site image information of the currently processed task that needs to be queued up with the original task that is adjusted after the currently processed task;

[0022] According to the comparison result, it is determined whether the order adjustment requirement is met.

[0023] Furthermore, the alarm voice information set is input into a deep neural network for analysis, and the voice information analysis result is output, including:

[0024] Select a convolutional neural network and establish a shared layer in the network structure to extract low-level feature representations;

[0025] Adding a task layer on top of the shared layer to perform speech recognition and sentiment analysis tasks respectively;

[0026] Use a joint loss function to simultaneously consider the speech recognition and sentiment analysis tasks when training the model, share parameters and minimize the loss functions of the two tasks;

[0027] The alarm voice information is input into the trained convolutional neural network, and the voice information analysis results are output.

[0028] Furthermore, the alarm voice information set is input into a deep neural network for analysis, and the voice information analysis result is output, including:

[0029] Build a speech recognition model;

[0030] The speech recognition model is trained, and during the training process, the sentiment analysis label is used to optimize the training strategy.

[0031] Furthermore, the speech recognition model is trained, and during the training process, the sentiment analysis label is used to optimize the training strategy, including:

[0032] Establishing a training set, wherein the training set includes speech information and corresponding sentiment analysis labels;

[0033] Combine the speech recognition task and sentiment analysis task and define a multi-task loss function, including speech recognition loss and sentiment analysis loss.

[0034] According to the input speech information, a deep neural network model is used to perform forward propagation calculation to obtain a first prediction result, and a second prediction result of the sentiment analysis label is also obtained to calculate the sentiment analysis loss;

[0035] Based on all the prediction results and true labels, the back propagation algorithm is used to update the model parameters so that the loss function value is reduced;

[0036] According to the performance of the training set and the validation set during the speech recognition model training process, the model structure, hyperparameters and training strategy are adjusted.

[0037] Furthermore, the sentiment analysis labels include three types: positive, negative and neutral.

[0038] Furthermore, establishing a speech recognition model includes:

[0039] Collect and prepare training and test datasets, including sound signals and corresponding text labels respectively;

[0040] Converting the sound signal into a high-dimensional vector representation using a feature extraction algorithm;

[0041] Select a deep neural network model and set the corresponding hyperparameters;

[0042] Use the training data set to train the model, and update the model parameters through the back propagation algorithm to improve the performance of the model on the training set;

[0043] Use a test dataset independent of the training and validation sets to evaluate the generalization and robustness of the model, at least calculating the model's prediction accuracy and recall metrics;

[0044] The model structure, hyperparameters and training strategies are adjusted according to the test results to obtain the speech recognition model.

[0045] Furthermore, the on-site image information is collected, including:

[0046] Segment and extract features from the scene image to obtain background environment features;

[0047] and, analyzing and identifying the posture features of the person to obtain the person's features;

[0048] Wherein, the on-site image information is collected according to guidance prompts.

[0049] Furthermore, an incentive layer is embedded in the deep neural network to optimize the deep learning model.

[0050] The technical solution of the present invention can achieve the following technical effects:

[0051] Through the present invention, on the one hand, an interactive service in the queuing process is provided in the existing alarm service, and the difficulty of subsequent processing is reduced through early data collection. On the other hand, the processing sequence can be adjusted for more urgent situations, thereby obtaining more timely and effective processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0053] Figure 1 A flowchart of the 5G visual alarm service method;

[0054] Figure 2 A flowchart for adjusting the processing order of queued video alarms in order to integrate on-site image information and accident assessment results;

[0055] Figure 3 A flowchart for inputting an alarm voice information set into a deep neural network for analysis and outputting the voice information analysis results;

[0056] Figure 4 Another implementation method for inputting an alarm voice information set into a deep neural network for analysis and outputting the voice information analysis result;

[0057] Figure 5 A flowchart for training a speech recognition model and optimizing the training strategy using sentiment analysis labels during the training process;

[0058] Figure 6 Flowchart for building a speech recognition model. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0060] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0061] Embodiment 1

[0062] A 5G visual alarm service system is used for the phased service when the alarm person is making a video alarm and queuing for access, including:

[0063] The preview video image acquisition module obtains the video preview screen of the alarm person and collects the scene image information from it;

[0064] Specifically, frames can be captured from a video stream and converted into an image format through an appropriate programming framework; of course, a simpler way is to take a screenshot of the video screen; the above two methods are different concepts, but both can achieve the purpose of the present invention. Specifically, capturing a frame means obtaining a specific frame from a video stream and converting it into a static picture, while screenshot means selecting a specific moment in a video or image and saving it as a static picture, which is usually used to capture a moment in the picture. In short, capturing a frame is the process of obtaining a certain number of frame sequences from a video stream and converting them into a static image, while screenshot is the process of quickly capturing a single static picture during video playback.

[0065] The voice information collection module collects voice information through the alarm person's responses to specific questions to obtain an alarm voice information set; the specific questions referred to here can be set according to actual work experience, taking traffic accidents as an example, including but not limited to the following questions:

[0066] (1) Time and location of the accident: The traffic police need to know the exact time and location of the accident. (2) Nature and circumstances of the accident: The traffic police need to know the nature and circumstances of the accident, such as the collision site, vehicle speed, etc. (3) Information about the parties and vehicles: The traffic police need to know the identity information and vehicle information of the parties, including name, phone number, license plate, etc. (4) Whether there are casualties: If there are casualties, the traffic police need to know the situation of the injured and how to deal with them. (5) Whether it affects traffic: The traffic police need to know whether the accident affects traffic and take corresponding measures to ensure traffic safety.

[0067] This part of the work does not require human intervention, and basic information collection can be achieved through pre-question setting and output.

[0068] The alarm analysis module is used to input the alarm voice information set into the deep neural network for analysis and output the voice information analysis results;

[0069] In specific implementation, when analyzing the alarm voice information set, deep neural network technology can be used for speech recognition and sentiment analysis, including but not limited to: (1) Speech recognition: identifying the text transcription results of the alarm voice information set, such as identifying "no casualties" and "no impact on traffic".

[0070] (2) Emotional analysis: Analyze the emotional colors contained in the alarm voice information set, such as warning, urgency, anxiety, etc., and output the emotional state expressed by the alarm information.

[0071] In speech recognition, deep learning-based speech recognition models can be used, such as recurrent neural networks (RNNs), convolutional neural networks (CNNs), and transcribers, etc. These models can process sound signals, convert them into symbol sequences, and then further classify and label the sequences; speech recognition models require a large amount of training data and computing resources to optimize model parameters in order to obtain more accurate recognition results.

[0072] In terms of sentiment analysis, deep learning-based sentiment classification models can be used, such as long short-term memory networks (LSTM), convolutional neural networks (CNN), and recurrent neural networks (RNN). These models can identify and classify different emotions expressed in speech, such as anger, fear, sadness, etc.; sentiment analysis models also require a large amount of training data and computing resources to obtain more accurate sentiment classification results.

[0073] Whether it is speech recognition or sentiment analysis, the speech signal needs to be preprocessed, such as denoising, downsampling, normalizing, etc., in order to better extract features.

[0074] The accident evaluation module is used to input the speech information analysis results into the neural network and output the accident evaluation results;

[0075] The accident evaluation module is a very important component that can provide fast and accurate accident evaluation results for various types of data, thereby helping people better understand and respond to accidents.

[0076] The sequence adjustment module integrates the on-site image information and accident assessment results to adjust the processing order of the queued video alarms.

[0077] In the process of queuing up a large number of processing tasks, how to reasonably adjust the processing order determines the response speed of emergency situations, such as the rapid response speed to major casualties and severe traffic congestion. The emergency situations referred to here mainly focus on the safety of personnel and the orderliness of traffic operations.

[0078] Embodiment 2:

[0079] like Figure 1 As shown, a 5G visual alarm service method is used to implement a phased service when the alarm person is making a video alarm and queuing for access, including:

[0080] S100: Obtain a video preview of the person who called the police, and collect on-site image information from it;

[0081] S200: collecting voice information from the alarm person by asking specific questions to obtain an alarm voice information set;

[0082] S300: Input the alarm voice information set into the deep neural network for analysis, and output the voice information analysis result;

[0083] S400: Input the speech information analysis result into the neural network and output the accident evaluation result;

[0084] S500: Comprehensively integrate the on-site image information and the accident assessment results, and adjust the processing order of the video alarms in the queue.

[0085] The technical effect achieved by the 5G visual alarm service method in this embodiment is the same as the technical effect that can be achieved by the system in Example 1, and will not be repeated here.

[0086] In the present invention, after the accident evaluation result is obtained, the scene image information and the accident evaluation result are integrated to adjust the processing order of the video alarms to be processed in the queue, such as Figure 2 As shown, including:

[0087] S510: Outputting sequence adjustment requirements based on accident evaluation results;

[0088] When this requirement is raised, the entire service process still does not require human intervention, and the result can be automatically achieved through program settings. Of course, the tasks that need to be processed in advance must obtain more urgent accident evaluation results; of course, if the accident evaluation results of the tasks that have been queued are the same as those of the tasks to be queued, the tasks to be queued are preferably arranged after the tasks that have been queued;

[0089] S520: Comparing the on-site image information of the currently processed task that needs to be queued and the original task that is adjusted to be placed after the currently processed task;

[0090] In this step, human participation is required to determine whether the order adjustment is finally achieved. During the implementation process, the collection of on-site image information can be carried out under the guidance of automatic voice prompts, so as to obtain more targeted and representative image information. In the present invention, a preview video is first obtained through the video request of the alarm person. Therefore, the collection of on-site image information is necessarily feasible. Of course, whether the image is clear or whether it ultimately meets the needs of use is largely related to the alarm person's understanding ability, the conditions of the equipment used, and the actual situation on the scene. Therefore, in the present invention, at this step, it is preferred to compare the on-site image information manually.

[0091] S530: Determine whether the sequence adjustment requirement is met based on the comparison result.

[0092] When it is determined that the requirements for sequence adjustment are met, the sequence of the alarm tasks is automatically increased through confirmation of the requirements, otherwise the original sequence is executed.

[0093] Among them, the alarm voice information set is input into the deep neural network for analysis, and the voice information analysis results are output, such as Figure 3 As shown, including:

[0094] S310: Select a convolutional neural network and establish a shared layer in the network structure to extract low-level feature representations;

[0095] S320: adding a task layer on top of the shared layer to perform speech recognition and sentiment analysis tasks respectively;

[0096] Depending on the task, different classifiers can be used, such as fully connected layer, softmax layer, etc., and corresponding loss functions can be set for each task, such as cross entropy loss function;

[0097] S330: Use a joint loss function to simultaneously consider the speech recognition and sentiment analysis tasks when training the model, share parameters and minimize the loss functions of the two tasks;

[0098] S340: Input the alarm voice information into the trained convolutional neural network and output the voice information analysis result.

[0099] The above method can realize the stacking of speech recognition model and sentiment analysis model, thereby improving the performance and generalization ability of the model.

[0100] As a preferred embodiment of the above, the alarm voice information set is input into the deep neural network for analysis, and the voice information analysis result is output, such as Figure 4 As shown, it may also include:

[0101] S350: Establishing a speech recognition model;

[0102] S360: Train the speech recognition model and optimize the training strategy using sentiment analysis tags during the training process.

[0103] S350~S360 here can achieve the same technical purpose as S310~S340, but adopts a different approach. S310~S340 implements a stacked model, while S350~S360 implements joint training.

[0104] Specifically, the speech recognition model is trained, and during the training process, the sentiment analysis tags are used to optimize the training strategy, such as Figure 5 As shown, it may include:

[0105] S361: Establish a training set, where the training set includes voice information and corresponding sentiment analysis labels, where the sentiment analysis labels include three types: positive, negative, and neutral;

[0106] S362: Combine the speech recognition task and the sentiment analysis task and define a multi-task loss function, including speech recognition loss and sentiment analysis loss.

[0107] Usually, weighted average, cross entropy and other methods are used to calculate the loss value, where the sentiment analysis label can be represented by One-hot encoding or Embedding.

[0108] S363: Perform forward propagation calculation using a deep neural network model according to the input speech information to obtain a first prediction result, and also obtain a second prediction result of the sentiment analysis label for calculating the sentiment analysis loss;

[0109] S364: Based on all the prediction results and true labels, the back propagation algorithm is used to update the model parameters so that the loss function value is reduced; at the same time, gradient clipping, regularization and other methods are also required to prevent the model from overfitting or underfitting;

[0110] S365: Adjust the model structure, hyperparameters, and training strategy based on the performance of the training set and validation set during speech recognition model training.

[0111] In step S350, Figure 6 As shown, establishing a speech recognition model includes:

[0112] S3111: Collect and prepare training and test data sets, including sound signals and corresponding text labels respectively. This step is the data preparation process;

[0113] S3112: Use a feature extraction algorithm to convert the sound signal into a high-dimensional vector representation. The above steps implement feature extraction. The vector referred to may include MFCC, FBANK, etc., as well as feature vectors based on deep learning such as Deep Feature;

[0114] S3113: Select a deep neural network model and set the corresponding hyperparameters; the hyperparameters of the deep neural network model include learning rate, batch size, number of iterations, number of layers, number of hidden units, etc. These hyperparameters will directly affect the performance and training speed of the model;

[0115] S3114: Use the training data set to train the model, and update the model parameters through the back propagation algorithm to improve the performance of the model on the training set; at the same time, the validation set needs to be evaluated and adjusted to prevent overfitting and underfitting;

[0116] S3115: Use a test dataset independent of the training and validation sets to evaluate the generalization and robustness of the model, at least calculating the prediction accuracy and recall metrics of the model;

[0117] S3116: Adjust the model structure, hyperparameters and training strategy according to the test results to obtain a speech recognition model.

[0118] Through the above method, a trained speech recognition model can be obtained, and the trained model can be applied in actual scenarios to output speech information analysis results.

[0119] Of course, sentiment analysis models can also be built, including:

[0120] S3121: Collect and prepare training and test datasets, including text data and corresponding labels for positive, neutral and negative, respectively;

[0121] S3122: Use feature extraction algorithms to convert text data into high-dimensional vector representations;

[0122] S3123: Select a deep neural network model and set corresponding hyperparameters;

[0123] S3124: Use the training data set to train the model, and update the model parameters through the back propagation algorithm to improve the performance of the model on the training set;

[0124] S3125: Use a test dataset independent of the training and validation sets to evaluate the generalization and robustness of the model, at least calculating the prediction accuracy and recall metrics of the model;

[0125] S3126: Adjust the model structure, hyperparameters and training strategy according to the test results to obtain a speech recognition model.

[0126] As a preferred embodiment of the above, collecting on-site image information includes:

[0127] Segment and extract features from the scene image to obtain background environment features;

[0128] and, analyzing and identifying the posture features of the person to obtain the person's features;

[0129] Among them, the on-site image information is collected according to the guidance prompts.

[0130] Guidance prompts can be used to guide the person reporting the incident to adjust the captured image. For example, through posture recognition, it can be determined whether the person can move freely and whether the set actions can be completed as required. This can help determine the urgency of the task.

[0131] In order to better achieve the technical purpose of the present invention, an incentive layer is embedded in the deep neural network model to optimize the deep learning model. The incentive layer is an important component of the deep learning model. Its function is to introduce nonlinear transformations to the neural network and enhance the model's expression and approximation capabilities. Specifically, the incentive layer can achieve the following functions:

[0132] Introducing nonlinearity: The excitation layer uses some nonlinear functions (such as Sigmoid, ReLU, Tanh, etc.) to perform nonlinear transformation on the input signal, thereby enhancing the expression and approximation capabilities of the neural network.

[0133] Compressed features: When the input feature dimension is high or redundant, the excitation layer can compress the features, reduce the number of neural network parameters, reduce computational complexity, and improve training efficiency.

[0134] Achieve normalization: The incentive layer can normalize the input, accelerate model convergence, and improve the robustness of the model.

[0135] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A 5G visual alarm service system, characterized in that: It is used for the phased service when the alarm user is making a video alarm and queuing for access, including: The preview video image acquisition module obtains the video preview screen of the alarm person and collects the scene image information from it; A voice information collection module collects voice information based on the alarm person's responses to specific questions to obtain an alarm voice information set; An alarm analysis module, used for inputting the alarm voice information set into a deep neural network for analysis and outputting the voice information analysis result; An accident evaluation module, used for inputting the speech information analysis result into a neural network and outputting an accident evaluation result; The sequence adjustment module adjusts the processing order of the video alarm tasks in the queue according to the on-site image information and the accident evaluation results.

2. A 5G visual alarm service method, characterized in that: It is used to implement phased services when the alarm user makes a video alarm and queues for access, including: Obtain the video preview of the person who called the police and collect on-site image information from it; Collecting voice information from the alarm person by asking specific questions to obtain an alarm voice information set; Inputting the alarm voice information set into a deep neural network for analysis, and outputting the voice information analysis result; Inputting the speech information analysis result into a neural network for analysis, and outputting an accident assessment result; Based on the on-site image information and accident assessment results, the processing order of the video alarm tasks in the queue is adjusted.

3. The 5G visual alarm service method according to claim 2, characterized in that: Based on the on-site image information and the accident assessment results, the processing order of the video alarms in the queue is adjusted, including: Outputting a sequence adjustment requirement based on the accident evaluation result; Comparing the on-site image information of the currently processed task that needs to be queued up with the original task that is adjusted after the currently processed task; According to the comparison result, it is determined whether the order adjustment requirement is met.

4. The 5G visual alarm service method according to claim 2, characterized in that: The alarm voice information set is input into a deep neural network for analysis, and the voice information analysis result is output, including: Select a convolutional neural network and establish a shared layer in the network structure to extract low-level feature representations; Adding a task layer on top of the shared layer to perform speech recognition and sentiment analysis tasks respectively; Use a joint loss function to simultaneously consider the speech recognition and sentiment analysis tasks when training the model, share parameters and minimize the loss functions of the two tasks; The alarm voice information is input into the trained convolutional neural network, and the voice information analysis results are output.

5. The 5G visual alarm service method according to claim 2, characterized in that: The alarm voice information set is input into a deep neural network for analysis, and the voice information analysis result is output, including: Build a speech recognition model; The speech recognition model is trained, and during the training process, the sentiment analysis label is used to optimize the training strategy.

6. The 5G visual alarm service method according to claim 5, characterized in that: The speech recognition model is trained, and during the training process, the sentiment analysis label is used to optimize the training strategy, including: Establishing a training set, wherein the training set includes speech information and corresponding sentiment analysis labels; Combine the speech recognition task and sentiment analysis task and define a multi-task loss function, including speech recognition loss and sentiment analysis loss. According to the input speech information, a deep neural network model is used to perform forward propagation calculation to obtain a first prediction result, and a second prediction result of the sentiment analysis label is also obtained to calculate the sentiment analysis loss; Based on all the prediction results and true labels, the back propagation algorithm is used to update the model parameters so that the loss function value is reduced; According to the performance of the training set and the validation set during the speech recognition model training process, the model structure, hyperparameters and training strategy are adjusted.

7. The 5G visual alarm service method according to claim 6, characterized in that: The sentiment analysis labels include three types: positive, negative and neutral.

8. The 5G visual alarm service method according to claim 5, characterized in that: Building a speech recognition model includes: Collect and prepare training and test datasets, including sound signals and corresponding text labels respectively; Converting the sound signal into a high-dimensional vector representation using a feature extraction algorithm; Select a deep neural network model and set the corresponding hyperparameters; Use the training data set to train the model, and update the model parameters through the back propagation algorithm to improve the performance of the model on the training set; Use a test dataset independent of the training and validation sets to evaluate the generalization and robustness of the model, at least calculating the model's prediction accuracy and recall metrics; The model structure, hyperparameters and training strategies are adjusted according to the test results to obtain the speech recognition model.

9. The 5G visual alarm service method according to claim 2, characterized in that: Collect on-site image information, including: Segment and extract features from the scene image to obtain background environment features; and, analyzing and identifying the posture features of the person to obtain the person's features; Wherein, the on-site image information is collected according to guidance prompts.

10. The 5G visual alarm service method according to claim 9, characterized in that: An incentive layer is embedded in the deep neural network to optimize the deep learning model.

Images