Efficient micro-model generation method and system based on field video data

Abstract

The invention relates to the field of visual identification, in particular to an efficient micro-model generation method and system based on field video data, and the method comprises the following steps: 1, obtaining the field video data, and carrying out the frame extraction, object labeling and sign extraction on the field video data, thereby obtaining sample data; step 2, training the sample data by using a micro-model algorithm to obtain an initial model, the micro-model algorithm comprising three core sub-networks: Backbone, Neck and Head, the Backbone comprising a plurality of CSP modules, and the Neck comprising a spatial pyramid pooling module; and 3, performing parameter tuning on the initial model to obtain a final micro-model. The micro-model constructed by the method has economic and social benefits, and economic loss and social resource waste caused by false alarms are reduced by reducing false alarms and improving the detection accuracy. And meanwhile, by quickly deploying and applying the AI technology, the technical innovation is accelerated, the social operation efficiency is improved, and the method has a positive promotion effect on social and economic development.

Description

Technical Field

[0001] The present invention relates to the field of visual recognition technology, and in particular to an efficient micro-model generation method and system based on on-site video data. Background Art

[0002] Convolutional neural networks (CNNs) in deep learning are usually composed of multiple key components, including convolutional layers, activation functions, pooling layers, and fully connected layers. In order to train these models, thousands of images are usually required. After the model is trained, it can be converted into a forward propagation model, using GPU or NPU to accelerate the reasoning process, process single or multiple images, and finally output key information such as target classification, object coordinates, similarity, etc.

[0003] Due to the large number of layers of convolutional neural networks, the huge training data set, the complex detection targets, and the large number of training batches and iterations required, the model training process often takes a long time. Generally speaking, it may take hours, dozens of hours, or even days to complete the training of a model.

[0004] The training workload of CNN models is huge, including the collection, labeling, and cleaning of training data, which often require manual intervention. In addition, deploying the trained model to the application system involves a long chain, a cumbersome validity verification process, and a long model iteration cycle, which makes it difficult to automate the trained model and push it to the actual application scenario in real time. Generally, after the CNN model is trained, it takes more than several weeks to update it to the AI ​​algorithm application system.

[0005] In terms of the selection of model training materials, although the range of materials is wide, the focus on specific scenarios is insufficient. This results in the trained model being able to adapt to a variety of scenarios, but in actual monitoring, there may be a high false alarm rate and low accuracy. Summary of the invention

[0006] To achieve the above purpose, the present invention proposes a stability lobe diagram drawing method based on an improved full discrete method. This method not only significantly improves the calculation accuracy, but also effectively improves the calculation efficiency, and can better meet the needs in practical applications. The present invention provides the following technical solutions:

[0007] An efficient micro-model generation method based on live video data, comprising:

[0008] The first step is to obtain on-site video data and extract frames, mark objects, and extract vital signs to obtain sample data;

[0009] In the second step, the sample data is trained using the micro-model algorithm to obtain the initial model. The micro-model algorithm includes three core sub-networks: Backbone, Neck, and Head. Backbone is a deep convolutional network that integrates multiple CSP (Cross Stage Partial) modules. These modules are specially designed to efficiently extract image features while reducing the amount of computation, laying a solid foundation for subsequent detection tasks. Neck includes a spatial pyramid pooling module, which enables the network to flexibly handle inputs of different sizes and extract features at multiple scales. In addition, through the path aggregation network, Neck realizes the fusion of multi-layer features and enhances the expressiveness of features, especially playing an important role in improving the detection performance of small objects. Head is responsible for the final detection task, which predicts the bounding box, object category, and object confidence. This part is the key to the model output and is directly related to the accuracy and reliability of target detection. The micro-model with this specific structure achieves lightweight and fast training of the model while maintaining high detection performance. This design makes the micro-model more suitable for scenarios that require rapid deployment and real-time processing, such as instant analysis and processing of on-site video data. This simplification and optimization of the micro-model not only improves efficiency, but also reduces the demand for computing resources, enabling it to perform well even in resource-constrained environments.

[0010] The third step is to tune the parameters of the initial model to obtain the final micro-model.

[0011] As a further solution of the present invention: the RequestCheck interface is used to obtain the on-site video data. The interface is used to receive single-frame video data that needs to be detected for the target, and return a "correct" or "wrong" result. It receives three parameters: the requested image data (in the form of a Base64 string), the detection target category (i.e., label, such as English labels such as car and person), and the target object coordinates (x1, y1, x2, y2). The general AI algorithm system can verify whether the target detection object inferred by the model is a real alarm record based on the result returned by this interface. If the result is "correct", it means that the reasoning result of this AI model is accurate; if the result is "wrong", it means that the reasoning of this AI model is a false alarm.

[0012] As a further solution of the present invention: in the third step, parameter tuning is to use the RequestData interface to receive the false alarm data after the micro-model detection and hand it over to the micro-model for re-training. The RequestData interface does not need to return the detection result, but only returns the result of the interface call: success or failure. The parameters it receives include: image data (Base64 string), detection target category (i.e. label, such as English logos such as car and person), target object coordinates (x1, y1, x2, y2) and similarity (value between 0-1). After receiving the request, the interface saves the data to the local disk, and records the image name, target category, target object coordinates, similarity and other data to the local file. The interface does not perform model reasoning, but waits for the micro-model system to retrieve the image, and realizes automatic labeling and automatic training.

[0013] As a further solution of the present invention: extracting physical signs is to identify the detection object, calculate the similarity and determine the object coordinate value.

[0014] As a further solution of the present invention: the total number of Backbone, Neck and Head layers is about 10 layers. The micro-model is an optimization and simplification based on the traditional CNN network architecture. The micro-model simplifies the originally complex and multi-layered neural network structure to a convolutional neural network of about 10 layers. This simplification not only reduces the number of parameters and training materials required for model training, but also significantly improves the speed of model training, thereby greatly shortening the model training time.

[0015] An efficient micro-model generation system based on live video data, comprising:

[0016] A sample generation module is used to obtain live video data and extract frames, object annotations and vital signs to obtain sample data;

[0017] The initial generation module is used to train the sample data using the micro-model algorithm to obtain the initial model;

[0018] The final generation module is used to tune the parameters of the initial model to obtain the final micro-model.

[0019] Compared with the prior art, the present invention has the following beneficial effects:

[0020] The micro-model constructed by the present invention can effectively reduce the false alarm rate, improve the accuracy rate, and can autonomously perform image annotation, training and iterative model. The longer the micro-model is used, the more accurate the micro-model's judgment of the result of CNN model reasoning in the general AI algorithm system will be, making the entire general AI algorithm system more effective;

[0021] The micro-model constructed by the present invention can improve detection accuracy and reduce false alarm rate: by simplifying the traditional CNN network structure, it can reduce training parameters and materials, accelerate model training speed, and shorten training time. This simplification makes the model more focused on specific scenarios and improves detection accuracy. Specifically, the micro-model performs secondary verification on the recognition results of the general AI model through the RequestCheck interface. Only the alarm records that are verified to be correct will be saved and reported, thereby significantly reducing the false alarm rate. Based on the secondary verification of the micro-model, the accuracy of the general AI algorithm system can be quickly improved to more than 95%, and the false alarm rate can be reduced to between 1% and 5%;

[0022] The micro-model constructed by the present invention can optimize resource utilization and cost efficiency: the simplified network structure reduces the demand for computing resources, so that small-scale, low-resource teams can also give full play to the potential of AI models. Compared with training models from scratch, the micro-model uses the basic knowledge of pre-trained models to effectively reduce the dependence on large-scale data and computing resources. The automatic training mechanism of the micro-model shortens the model training process to a few minutes, and reduces the number of training materials from tens of thousands or hundreds of thousands to dozens, saving a lot of time for manual labeling, model training and model updating, shortening the time from several weeks to multiple model training, model iteration and model updating within one day;

[0023] The micro-model constructed by the present invention has economic and social benefits: by reducing false alarms and improving detection accuracy, it reduces economic losses and waste of social resources caused by false alarms. At the same time, by rapidly deploying and applying AI technology, it accelerates technological innovation, improves social operation efficiency, and has a positive role in promoting social and economic development. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Flow chart of an efficient micro-model generation method based on live video data in an embodiment of the present invention.

[0025] Figure 2 4 is a structural block diagram of the composition of an efficient micro-model generation system based on live video data in an embodiment of the present invention.

[0026] Figure 3 Schematic diagram of the working principle of the micro-model obtained by the efficient micro-model generation method based on on-site video data in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The present invention aims to simplify the AI ​​model training process, reduce manual intervention, and accelerate the model iteration process. The system can automatically capture on-site video data and perform intelligent annotation. Through a customized convolutional neural network, i.e., a micro-model, an automated training process is realized, and the trained model is automatically converted into a forward model to achieve efficient reasoning for a single picture. The key is that the training materials in the micro-model are all derived from the false alarm pictures of the general AI algorithm system. Therefore, when the CNN network of the micro-model is inferred, if the result is a high similarity, it can be confirmed that the alarm picture is a false alarm. At a lower similarity, or when the similarity is 0, the alarm record is likely to be a correct alarm record. The present invention can be quickly integrated with the existing AI algorithm application system through a standardized API interface, effectively solving the problem of high false alarm rate and low accuracy of the general CNN model in practical applications. Through this method, not only the efficiency and accuracy of model training are improved, but also a more flexible and efficient solution is provided for the practical application of AI models.

[0029] The specific implementation of the present invention is described in detail below in conjunction with specific embodiments.

[0030] See also Figure 1-Figure 3This product is based on the Linux operating system, uses the C++ development language, and builds a service program based on the HTTP protocol. By specifying a port (for example: 8080), the system provides an API interface service that complies with the Restful protocol standard, so that any system that follows the standard can call it. These API interfaces are mainly responsible for processing requests based on network communication from other AI algorithm systems, middle-end systems, etc., and output the results of image detection by micro-models, thereby reducing the false alarm rate in traditional CNN models and improving the accuracy rate. The present invention includes two main interfaces: RequestCheck interface and RequestData interface. The RequestCheck interface is used to receive single-frame video data that requires target detection and return a "correct" or "wrong" result. It receives three parameters: the requested image data (in the form of a Base64 string), the detection target category (i.e., label, such as English labels such as car and person), and the target object coordinates (x1, y1, x2, y2). The general AI algorithm system can verify whether the target detection object inferred by the model is a real alarm record based on the results returned by this interface. If the result is "correct", it means that the reasoning result of this AI model is accurate; if the result is "wrong", it means that the reasoning of this AI model is a false alarm.

[0031] The RequestData interface is used to receive false positive data after detection by the general AI model and hand it over to the micro-model for retraining. This interface does not need to return the detection result, but only returns the result of the interface call: success or failure. The parameters it receives include: image data (Base64 string), detection target category (i.e. label, such as English labels such as car and person), target object coordinates (x1, y1, x2, y2), and similarity (value between 0-1). After receiving the request, the interface saves the data to the local disk and records the image name, target category, target object coordinates, similarity and other data to the local file. This interface does not perform model inference, but waits for the micro-model system to retrieve the image and realize automatic labeling and automatic training. Through the RequestCheck interface and the RequestData interface, the micro-model can effectively interact with external systems to improve the accuracy of target detection, while retraining false positive data to optimize model performance.

[0032] The micro-model constructed by the present invention is independently deployed on hardware devices equipped with GPU or NPU. These devices can run independently and freely network to achieve efficient data processing, have the ability to read the video stream of each camera in real time, and can push video images of 25 frames per second and above to the pre-trained CNN model to perform reasoning operations and obtain reasoning results. Each micro-model is assigned an independent IP address and port, and communicates with the outside world through the HTTP protocol to ensure the flexibility and security of data transmission. The system has built-in API interface services as described above, enabling it to play the role of a central service system in multiple AI algorithm systems. Under this architecture, each general AI algorithm system can easily configure the IP address and port of the micro-model to initiate a request to it. These requests are intended to verify the recognition results of the AI ​​model and ensure accuracy. At the same time, they are also responsible for pushing the false alarm video frame data and related parameters generated by the AI ​​algorithm system. In this way, the micro-model can not only perform secondary verification on the output of the AI ​​model to reduce false alarms, but also use these data for self-learning and optimization to further improve the detection accuracy. The independence and flexibility of the micro-model make it an indispensable part of the AI ​​algorithm system. The micromodel provides a powerful central processing platform for each AI algorithm system through API interface services. This platform can not only process large amounts of video frame data, but also intelligently identify and correct false positives, thereby improving the reliability and efficiency of the entire system. In addition, the micromodel's automatic labeling and automatic training functions enable it to continuously adapt to new data and scenarios, maintaining the model's advancement and accuracy.

[0033] The micro-model improves the accuracy of target detection and reduces false positives by simplifying and accelerating the CNN network reasoning process. The micro-model processes the video stream: the general AI model algorithm system first reads the video stream and inputs each frame of video image into the CNN network for reasoning; the CNN network analyzes the image content, identifies the detection object, calculates similarity, determines key information such as the object coordinate value, etc.

[0034] Micro-model generates alarm records: Based on the dangerous events and detection thresholds preset by the AI ​​algorithm system, the micro-model generates alarm records. These records include pictures with detection boxes, labels of target detection objects, object similarity, and coordinate values ​​of objects in the image.

[0035] Micro-model verification of alarm records: After the alarm record is generated, the micro-model transmits the alarm image, similarity, coordinate value and target object data through the RequestCheck function of the API interface. Only when the interface returns "correct" will the alarm record be saved in the system, indicating that the alarm record has been verified twice by the micro-model and confirmed as a valid alarm. At this time, the micro-model will push the alarm record out through the reporting mechanism. If the interface returns "error", the alarm is considered to be a false alarm, and the alarm record is ignored and not reported.

[0036] Micro-model processing false alarm data: When the general AI model records multiple alarm records, these alarm events will be manually identified. Once some alarm records are determined to be false alarms, the micro-model will transmit the image data, detection target category, target object coordinates and similarity in the false alarm record through the RequestData function of the API interface. The micro-model records and saves this data, uses the automatic training mechanism to extract negative sample data for training, and updates the trained model for RequestCheck interface calls and reasoning.

[0037] Optimization of micro-models and reduction of false alarms: As the general AI model system continuously requests the RequestCheck interface and pushes more false alarm data, the CNN network of the micro-model gradually learns and optimizes the error reasoning parameters in the current specific scenario. This means that the micro-model is increasingly accurate in identifying false alarm images. When the AI ​​general model requests the RequestCheck interface again, the micro-model can provide more accurate judgments, thereby reducing false alarms and improving the accuracy of detection. Through the above workflow, the micro-model not only improves the accuracy of alarm records, but also continuously adapts to new detection scenarios through continuous learning and optimization, reduces false alarms, and improves the overall performance of the AI ​​model.

[0038] Compared with traditional monitoring, the present invention has the following advantages: 1. Low-cost upgrade without changing the camera: Common and mainstream high-definition cameras can meet the upgrade requirements, and the upgrade can be easily achieved by using a server adapted to the algorithm. 2. Intelligent data analysis: Based on existing cameras, 50 standard algorithms and customized models are used to achieve real-time data analysis and key event recognition, with a response in seconds. 3. Scenario-based early warning: Identify events such as intrusion, illegal parking, and abnormal behavior, and actively trigger alarms. 4. Dynamic control capability: Target tracking and dynamic adjustment of monitoring strategies are achieved through scene recognition. 5. Deep learning closed-loop optimization: AI cameras continuously improve recognition accuracy and reliability by continuously learning field data. The present invention is a closed-loop application of algorithms, models, integrated engineering and deep re-learning.

[0039] The first phase of this product was launched simultaneously in Qingdao and Chengdu, with a construction period of only one month, achieving seamless integration with the daily operations of the park. The Beijing headquarters accesses national data in real time through the central management platform to ensure full coverage of security management. The recognition rate of this product has been increased to more than 96%, the false alarm rate has been reduced to 4%, and the missed alarm rate is close to 0, and it is being continuously optimized; the alarm response time has been shortened from the traditional hourly calculation to the fastest 3 minutes, and the safety efficiency has been significantly improved; it supports multi-dimensional alarm information synchronization, and the monthly report generation and analysis are more transparent.

[0040] Example 1: In an emergency at a Chengdu campus, the AI ​​camera identified the intruding vehicle in real time and triggered an alarm quickly. The security team intercepted the vehicle in just 3 minutes, successfully avoiding a possible safety accident. Customers commented: "The AI ​​system is not only a monitoring device, but also a 24 / 7 security manager."

[0041] Example 2: In a potential spontaneous combustion incident in a transport trailer in Qingdao, the AI ​​camera immediately identified the abnormal situation and triggered an alarm through its smoke detection function. The management team responded quickly and put out the small flame, avoiding possible major property losses. Customer comments: "The AI ​​camera's accurate identification and immediate alarm saved us from a possible major property loss, and the effect is far better than traditional monitoring equipment."

[0042] Example 3: Monitoring the operating status of large equipment in the Chengdu Park (such as manual / automatic mode) usually requires deep background integration and complex configuration. This product directly obtains the device status through visual recognition, without the need for additional hardware or cumbersome settings, which greatly reduces time and labor costs, while improving management efficiency. Customers commented: "Such technological improvements make us full of confidence in digital transformation, and we hope to explore more AI applications in the future."

[0043] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. An efficient micro-model generation method based on live video data, characterized in that: The following steps are involved: The first step is to obtain on-site video data and extract frames, mark objects, and extract vital signs to obtain sample data; In the second step, the sample data is trained using the micro-model algorithm to obtain the initial model. The micro-model algorithm includes three core sub-networks: Backbone, Neck and Head. Backbone includes multiple CSP modules, and Neck includes a spatial pyramid pooling module. The third step is to tune the parameters of the initial model to obtain the final micro-model.

2. The efficient micro-model generation method based on on-site video data according to claim 1 is characterized in that: In the first step, the RequestCheck interface is used to obtain the on-site video data.

3. The efficient micro-model generation method based on on-site video data according to claim 1 or 2, characterized in that: The parameter tuning in the third step is to use the RequestData interface to receive the false positive data after the micro-model detection and hand it over to the micro-model for re-training.

4. The efficient micro-model generation method based on on-site video data according to claim 1 or 2, characterized in that: The physical sign extraction is to identify the detected object, calculate the similarity and determine the object coordinate value.

5. The efficient micro-model generation method based on on-site video data according to claim 1 is characterized in that: The total number of layers of Backbone, Neck and Head is 10.

6. The efficient micro-model generation method based on on-site video data according to claim 3 is characterized in that: The false alarm data received by the RequestData interface includes false alarm image data, false alarm detection target category, false alarm target object coordinates and false alarm similarity.

7. An efficient micro-model generation system based on live video data, characterized in that: The system comprises: A sample generation module is used to obtain live video data and extract frames, object annotations and vital signs to obtain sample data; The initial generation module is used to train the sample data using the micro-model algorithm to obtain the initial model; The final generation module is used to tune the parameters of the initial model to obtain the final micro-model.

Images