A Method and System for Intelligent Analysis and Comprehensive Scoring of Mine Driver Behavior

Abstract

This invention discloses a method and system for intelligent analysis and comprehensive scoring of mine driver behavior. The method includes: a terminal device collecting driver behavior data and vehicle driving information; an edge device determining whether the driver has committed a violation and whether the terminal device has malfunctioned based on a trained lightweight behavior recognition model. The lightweight behavior recognition model is trained on a cloud server, and the convolutional layers in the head network of the lightweight behavior recognition model are depthwise separable convolutional layers, with all weights in the weighted layers being quantized weights. If the driver commits a violation or the terminal device malfunctions, the terminal device sends a violation warning or anomaly warning to the driver, and the cloud server evaluates the driver based on the violation data to obtain a comprehensive score for the driver. This invention's model has lower computational complexity and better real-time behavior monitoring.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of driving behavior recognition, and particularly relates to a mine driver behavior intelligent analysis and comprehensive scoring method and system. BACKGROUND

[0002] With the acceleration of the intelligent transformation of the mining industry, the safety control and efficiency improvement of mine operations have become the top priority of the industry development. In view of the complex mine operation environment, the operation behavior of the driver is directly related to the safety of the entire operation process, so it is particularly crucial to monitor and analyze the driver behavior in real time and accurately.

[0003] However, the traditional driving behavior monitoring system is not suitable for the mine scene. The traditional monitoring system usually monitors in a good communication environment, such as monitoring the behavior of bus or private car drivers, uploads the driver behavior data to the cloud for behavior result analysis to obtain the behavior analysis result. However, the communication equipment construction on the mine is relatively insufficient, and if the behavior data is uploaded to the cloud for behavior analysis, it will produce a large delay, making the system real-time poor and the response speed slow. SUMMARY

[0004] The mine driver behavior intelligent analysis and comprehensive scoring method and system provided by the embodiments of the application can solve the problem of poor real-time performance of the current driving behavior monitoring system when applied to the mine scene.

[0005] In a first aspect, the mine driver behavior intelligent analysis and comprehensive scoring method provided by the embodiments of the application is applied to a mine driver behavior intelligent analysis and comprehensive scoring system, the system includes a terminal device, an edge device and a cloud server, and the method includes:

[0006] The terminal device collects behavior data and vehicle driving information of the driver;

[0007] The edge device determines whether the driver has violated the rules and whether the terminal device has appeared abnormities based on the trained light-weight behavior recognition model and the behavior data and vehicle driving information, wherein the light-weight behavior recognition model is trained on the cloud server, the convolution layer of the head network in the light-weight behavior recognition model is a depth separable convolution layer, and the weights of all weight layers in the light-weight behavior recognition model are quantized weights;

[0008] If the driver has violated the rules or the terminal device has appeared abnormities, the terminal device sends a rule violation warning or an abnormality warning to the driver, and the edge device saves the rule violation data or the terminal device abnormality data;

[0009] The cloud server evaluates the driver according to the violation behavior data to obtain a comprehensive score of the driver.

[0010] In a second aspect, the embodiment of the present application provides a mine driver behavior intelligent analysis and comprehensive scoring system, comprising a terminal device, an edge device and a cloud server.

[0011] The terminal device is used for collecting behavior data of a driver and vehicle driving information.

[0012] The edge device is used for determining whether the driver produces a violation behavior and whether the terminal device is abnormal based on a trained lightweight behavior recognition model and the behavior data and the vehicle driving information, wherein the lightweight behavior recognition model is trained on the cloud server, a convolution layer of a head network in the lightweight behavior recognition model is a depth separable convolution layer, and weights of all weight layers in the lightweight behavior recognition model are quantized weights.

[0013] If the driver produces a violation behavior or the terminal device is abnormal, the terminal device is further used for sending a violation warning or an abnormal warning to the driver, and the edge device is further used for saving violation behavior data or terminal device abnormal data.

[0014] The cloud server is used for evaluating the driver according to the violation behavior data to obtain a comprehensive score of the driver.

[0015] Compared with the prior art, the embodiment of the present application has the beneficial effects that the system provided by the present application adopts a cloud-edge-end collaborative architecture, and combines the powerful computing capacity of the cloud server with the real-time response capacity of the edge device. The edge device is responsible for real-time acquisition of driver behavior data, and performs preliminary processing and analysis by using a lightweight deep learning model. The cloud server is responsible for receiving data uploaded by the edge device, performing deeper analysis, model training and optimization, and storing and managing large-scale data. This architecture can guarantee the real-time performance of the system, and can realize comprehensive and in-depth analysis of data. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A structure schematic diagram of a lightweight behavior recognition model provided by the embodiment of the present application;

[0017] Figure 2 A structure schematic diagram of a mine driver behavior intelligent analysis and comprehensive scoring system provided by the embodiment of the present application;

[0018] Figure 3 An implementation flowchart of a mine driver behavior intelligent analysis and comprehensive scoring method provided by the embodiment of the present application;

[0019] Figure 4 A schematic diagram of a data acquisition page provided in an embodiment of the present invention;

[0020] Figure 5 A schematic diagram of driver behavior data provided in an embodiment of the present invention;

[0021] Figure 6 This is a schematic diagram of vehicle driving information provided in an embodiment of the present invention;

[0022] Figure 7 This is a schematic diagram illustrating a format for storing valid information in violation data, as provided in an embodiment of the present invention. Detailed Implementation

[0023] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0024] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0025] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0026] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0027] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0029] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0030] Example 1

[0031] Figure 1 The diagram shown is a structural schematic of a lightweight behavior recognition model provided in an embodiment of the present invention.

[0032] As an example, and not a limitation, Model 100 can be an improvement on the traditional YOLOv5 model. See also Figure 1 Similar to the traditional YOLOv5 model, Model 100 can also include a backbone network, a neck network, and a head network. The difference is that Model 100 replaces the convolutional layers of the head network in the YOLOv5 model with depthwise separable convolutional layers and performs quantization on the weights of all weighted layers in the model; this improves the model's running efficiency and inference speed while significantly reducing memory usage and the number of model parameters.

[0033] For example, in Model 100, the convolutional layers in the Dual IoU Perceptual Decoupled Head Module (DDH) of the head network in the YOLOv5 model can be replaced with depth-separable convolutional layers.

[0034] For example, the weighted layers in model 100 may include all modules with weights, such as convolutional layers, depthwise separable convolutional layers, and upsampling layers.

[0035] In one possible implementation, depthwise separable convolution operations can consist of depthwise convolution and pointwise convolution.

[0036] In one example, the depthwise convolution process can be represented as:

[0037]

[0038] Where Yd(i′,j′) is the value of the i′ row and j′ column of the feature map output by the depthwise convolution, H is the height of the feature map input by the depthwise convolution, W′ is the width of the feature map input by the depthwise convolution, X(i′+m′,j′+n) is the value of the i′+m′ row and j′+n column of the feature map input by the depthwise convolution, and Dk(m′,n) is the value of the m′ row and n column of the depthwise convolution kernel of the k-th channel.

[0039] In one example, the pointwise convolution process can satisfy the following formula:

[0040]

[0041] Where Y(i′,j′) is the value of the pixel in the i′ row and j′ column of the pointwise convolution output feature map, C is the total number of input feature channels, and Pk(i′,j′) is the value of the element in the i′ row and j′ column of the pointwise convolution kernel of the k-th channel.

[0042] In one possible implementation, the quantized weights of Model 100 can satisfy the following formula:

[0043]

[0044] Where Q(W) is the quantized weight, W is the original weight, Qmax and Qmin are the maximum and minimum values ​​of the quantization range, max(W) and min(W) are the maximum and minimum values ​​of W, and the round function is used to round the scaled floating-point value to an integer.

[0045] This invention significantly reduces model size and computational complexity while maintaining high recognition accuracy through quantization and network structure optimization. This enables edge devices to quickly and accurately identify driver violations even under resource-constrained conditions.

[0046] Example 2

[0047] Figure 2 The diagram shown illustrates the structure of a mine driver behavior intelligent analysis and comprehensive scoring system provided in an embodiment of the present invention. See also, as an example and not a limitation. Figure 2 The system may include terminal devices, edge devices, and cloud servers.

[0048] For example, the terminal device can collect driver behavior data and vehicle driving information and send them to an edge device. The edge device can then input the behavior data and vehicle driving information into a built-in, trained, lightweight behavior recognition model. The trained model 100 can then determine whether the driver has committed a violation or whether the terminal device is malfunctioning. If the driver commits a violation or the terminal device malfunctions, the terminal device will send a violation warning or anomaly warning to the driver and save the violation data or anomaly data to the cloud server. The cloud server can then evaluate the driver based on the violation data to obtain a comprehensive driver score, or alert maintenance personnel to inspect the terminal device based on the anomaly data.

[0049] In one possible implementation, the terminal device may include an infrared camera, a color camera, a display screen, an audio output module, and a positioning device.

[0050] For example, an embedded operating system and applications can be installed on the terminal device to support communication and data interaction with edge devices.

[0051] In one example, an infrared camera, a color camera, and a positioning device enable the terminal device to collect data. The infrared and color cameras can collect driver behavior data, capturing facial expressions and gestures; they can also capture images of road conditions. The positioning device can obtain the vehicle's real-time location information via the Global Navigation Satellite System (GNSS).

[0052] For example, both the infrared camera and the color camera can be installed in the driver's cab. The infrared camera can be located at a first preset height slightly below and directly in front of the driver's seat, or at a position 20 to 45 degrees to the left or right of the driver's seat. The color camera can be located at a second preset height directly above the centerline of the driver's cab.

[0053] For example, the values ​​of the first preset height and the second preset height can be determined based on the shape and structure of the cab. The offset angle of the infrared camera can be determined based on the left and right steering wheel configuration of the driver's seat.

[0054] The placement of the camera significantly impacts the recognition accuracy of Model 100. Traditional behavior recognition systems typically mount the camera directly in front of the driver's seat, a configuration that is difficult to adapt to different models of mining trucks. Furthermore, vehicles are divided into right-hand drive and left-hand drive types, further reducing the number of qualified images captured by the camera for model recognition. This invention, through the aforementioned configuration, staggers the placement of the infrared and color cameras, and adjusts the infrared camera position according to the shape and structure of the mining truck's cab and the vehicle's left-hand drive configuration, thereby improving detection performance.

[0055] Optionally, the terminal device can also integrate facial recognition login functionality, using infrared cameras or color cameras to identify the driver's identity.

[0056] Specifically, the terminal device can integrate two advanced face detection algorithms, RetinaFace and ArcFace, for facial recognition. This design ensures that the system accurately establishes a binding relationship with a specific driver, thereby enabling a highly personalized management solution.

[0057] In one example, the audio output module and display screen can enable the alarm function of the terminal device, sending violation warnings and abnormal warnings to the driver through voice alarms, alarm displays and other means to provide feedback on driving behavior.

[0058] In one possible implementation, the edge device can integrate functions such as data reception, data transmission, real-time computing, data processing, and data storage.

[0059] For example, the edge device can be an edge computing device that supports deep learning acceleration, such as the Rockchip series of development boards.

[0060] Specifically, edge devices can receive driver behavior data and vehicle driving information; perform real-time calculations and data processing using built-in Model 100 and MediaPipe algorithms; send violation data or terminal device anomaly data to the cloud server; send calculation results (i.e., behavior recognition results) to the terminal device; and store violation data or terminal device anomaly data.

[0061] In one possible implementation, the cloud server can integrate functions such as data reception, data transmission, data processing, driving behavior assessment, and model training.

[0062] For example, the cloud server can be a high-performance cloud computing platform, such as Alibaba Cloud or Tencent Cloud. A deep learning model training and optimization platform can be built on it, supporting online training and optimization of Model 100.

[0063] In one example, the cloud server can train model 100 to obtain a trained model 100; then the trained model 100 is deployed to an edge device.

[0064] For example, cloud servers can use the Rockchip NN (RKNN) inference engine for model transformation and optimization.

[0065] Specifically, the cloud server first uses the RKNN toolkit to convert the PyTorch format model 100 to RKNN format; secondly, it configures the runtime environment of the edge device to ensure that the RKNN runtime environment and dependent libraries (such as OpenCV, TensorFlow Lite, etc.) are installed and correctly configured; finally, it uploads the converted RKNN model file to the edge device and uses the corresponding inference code written by the user to load and run the model.

[0066] Optionally, after the system has been running for a period of time, the cloud server can periodically update and optimize model 100; or obtain user feedback and retrain model 100 based on that feedback; or it can conduct targeted training and adjustments to model 100 based on changes in the mine's operating environment and driver behavior characteristics. This helps improve the model's recognition accuracy and real-time performance.

[0067] In one example, the cloud server can also receive violation data or terminal device anomaly data via a secure API interface. Based on the violation data, driving behavior is assessed to obtain a comprehensive driver score. The terminal device anomaly data is processed to evaluate the terminal device's performance and to alert maintenance personnel to inspect the terminal device.

[0068] For example, the violation data can be in JSON or CSV format for easy parsing and storage.

[0069] In one example, the cloud server can also generate driver evaluation reports based on the driver's overall rating and send all driver evaluation reports to the administrator for safety management, driver training, and performance evaluation.

[0070] The system provided by this invention adopts a cloud-edge-device collaborative architecture, combining the powerful computing capabilities of cloud servers with the real-time response capabilities of edge devices. Edge devices are responsible for acquiring driver behavior data in real time and performing preliminary processing and analysis using lightweight deep learning models; the cloud server is responsible for receiving data uploaded by edge devices, performing deeper analysis, model training and optimization, and storing and managing large-scale data. This architecture ensures both the real-time performance of the system and enables comprehensive and in-depth analysis of the data.

[0071] Example 3

[0072] Figure 3 The diagram illustrates the implementation flow of a mine driver behavior intelligent analysis and comprehensive scoring method according to an embodiment of the present invention. As an example and not a limitation, this method can be applied to the aforementioned system. The method may include steps S301-S307, which are described below.

[0073] S301, the terminal device collects driver behavior data and vehicle driving information.

[0074] For example, see Figure 4 The data collection page shown may include driver behavior data, such as images of the driver (see [link]). Figure 5 Vehicle driving information may include road conditions (see...). Figure 6 (a) in the middle), the location of the mine car (see Figure 6 The satellite positioning map shown in (b) is an example.

[0075] S302, the terminal device sends driver behavior data and vehicle driving information to the edge device.

[0076] Accordingly, edge devices receive driver behavior data and vehicle driving information.

[0077] S303, the edge device uses a trained lightweight behavior recognition model to determine whether the driver has committed a violation and whether the terminal device has malfunctioned, based on behavior data and vehicle driving information.

[0078] For example, types of violations may include driving while fatigued (see...). Figure 5 (a) and Figure 5 (b) of the above, distraction, using a mobile phone, smoking, obstructing a camera, not wearing a seatbelt (see also...) Figure 5 (c) etc.

[0079] For example, abnormal situations of terminal devices include camera damage, camera connection problems, etc.

[0080] In one possible implementation, if the driver commits a violation or the terminal device malfunctions, the following steps S304 or S305 can be performed.

[0081] In one example, if a driver commits a violation, step S304 can be performed.

[0082] In another example, if the terminal device malfunctions, the following step S305 can be performed.

[0083] For example, when the camera malfunctions, the acquired image of the driver may have no image input or excessive image noise. When the behavior model recognizes that the input image is unqualified, it can determine that the terminal device is malfunctioning.

[0084] In another example, if the terminal device malfunctions and the driver commits a violation, steps S304 and S305 can be performed simultaneously.

[0085] In another possible implementation, if the driver does not commit any violation and the terminal device does not malfunction, step S301 can be performed to continue collecting the driver's behavior data for the next moment.

[0086] S304, The edge device sends the behavior recognition results to the terminal device so that the terminal device can send a violation warning to the driver.

[0087] In one example, once the edge device detects a driver's violation, the system immediately triggers an early warning mechanism, providing the driver with real-time feedback via the terminal device, including audible prompts and visual warnings, to prompt the driver to correct the inappropriate behavior promptly. Simultaneously, step S306 can be performed to upload the violation data to a cloud server for subsequent processing and analysis by the mine administrator.

[0088] For example, violation data may include the type of violation, the time of occurrence, the location of occurrence, and the duration of the violation.

[0089] For example, the terminal device can display warning information on the screen, such as "Please pay attention to driving safety and avoid driving while fatigued," and play sound prompts through the audio output device.

[0090] S305, the edge device sends the terminal device identification result to the terminal device so that the terminal device can send an abnormal warning to the driver.

[0091] For example, if a camera connection error occurs, the terminal device can send an alert to the driver, reminding them to check the camera's network connection and power supply. Simultaneously, it can send abnormal data from the terminal device to a cloud server for device malfunction detection.

[0092] S306, Edge devices send violation data to cloud servers.

[0093] Accordingly, the cloud server receives data on violations.

[0094] S307, the cloud server evaluates drivers based on violation data to obtain a comprehensive score for the drivers.

[0095] In one possible implementation, the cloud server can first analyze the violation data, using data processing and machine learning algorithms to extract valuable information. Then, based on a scoring model, the driver can be evaluated using the information extracted from the violation data to obtain a comprehensive driver score.

[0096] In one example, cloud servers can employ data processing methods such as data cleaning, removing duplicate and invalid data entries to ensure data accuracy.

[0097] In one example, the information extracted from violation data may include the number of violations, the time and location of the violations, etc.

[0098] For example, the number of violations can be calculated using the following formula:

[0099]

[0100] Where Nb is the total number of violations, I(bj) is an indicator function that returns 1 when the j-th violation bjbn occurs, otherwise returns 0, and m is the total number of violation types.

[0101] For example, the time of occurrence of the violation can be represented as: Lb={(lat1,lon1),(lat2,lon2),…,(latk,lonk)}, where Lb is the set of locations where the violation occurred, latk is the location of the violation, and lonk is the location of the violation.

[0102] Optionally, the cloud server can be configured as follows: Figure 7 The format shown stores information extracted from violation data.

[0103] In one example, the scoring model can satisfy the following formula:

[0104]

[0105] Where S is the driver's overall score, wj is the weight of the j-th violation, Nj is the number of times the j-th violation occurs, and m is the total number of violation types.

[0106] For example, the weight of the j-th violation can be preset according to the severity of the violation; for example, the weight of driving while fatigued can be higher than that of making a phone call.

[0107] Optionally, the method may further include the following step S308.

[0108] S308, the cloud server generates an evaluation report for the driver based on the comprehensive score; and sends the evaluation report to the administrator so that the administrator can carry out safety management, driver training and performance evaluation.

[0109] For example, the cloud server can send the generated assessment report and the driver's overall score to the administrator via email or a management platform interface. The administrator can then use this information for safety management, driver training, and performance evaluation.

[0110] The system provided by this invention adopts a cloud-edge-device collaborative architecture, combining the powerful computing capabilities of cloud servers with the real-time response capabilities of edge devices. Edge devices are responsible for acquiring driver behavior data in real time and performing preliminary processing and analysis using lightweight deep learning models; the cloud server is responsible for receiving data uploaded by edge devices, performing deeper analysis, model training and optimization, and storing and managing large-scale data. This architecture ensures both the real-time performance of the system and enables comprehensive and in-depth analysis of the data.

[0111] Furthermore, the cloud server generates detailed driving behavior assessment reports based on the collected driver behavior data and assigns quantitative scores to drivers. These reports and scores not only reflect the current behavioral status of drivers but also provide mine managers with long-term driving behavior trend analysis, offering strong support for safety management, driver training, and performance evaluation.

[0112] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

Claims

1. A mine driver behavior intelligent analysis and comprehensive scoring method, characterized in that, The method is applied to a mine driver behavior intelligent analysis and comprehensive scoring system, the system comprising a terminal device, an edge device and a cloud server, and the method comprising: The terminal device collects driver behavior data and vehicle driving information; The edge device determines whether the driver has committed a violation and whether the terminal device has an abnormality based on a trained lightweight behavior recognition model according to the behavior data and vehicle driving information, wherein the lightweight behavior recognition model is trained on the cloud server, the convolution layer of the head network in the lightweight behavior recognition model is a depth separable convolution layer, and the weights of all weight layers in the lightweight behavior recognition model are quantized weights; If the driver has committed a violation or the terminal device has an abnormality, the terminal device sends a violation warning or an abnormality warning to the driver, and the edge device saves violation behavior data or terminal device abnormality data; The cloud server evaluates the driver based on the violation behavior data to obtain a comprehensive score of the driver; The quantized weights satisfy the following formula: ， wherein is the quantized weight of the weight layer, is the original weight of the weight layer, , are the maximum and minimum values of the quantization range, respectively, , are the maximum and minimum values of , respectively, is a function for rounding a scaled floating point value to an integer.

2. The method of claim 1, wherein, The terminal device comprises an infrared camera, a color camera, a display screen, an audio output module and a positioning device; The terminal device collects driver behavior data and vehicle driving information, comprising: The infrared camera, the color camera and the positioning device collect the driver behavior data and the vehicle driving information; The terminal device sends a violation warning or an abnormality warning to the driver, comprising: The display screen and the audio output module send a violation warning or an abnormality warning to the driver.

3. The method of claim 2, wherein, The infrared camera is installed in the driver's cabin at a position 20 to 45 degrees to the left or right of the main driving position or at a position offset downward from the front of the main driving position at a first preset height, and the color camera is installed in the driver's cabin at a position directly above the center line of the driver's cabin at a second preset height.

4. The method of claim 1, wherein, The types of violation behaviors include fatigue driving, distraction, playing with a mobile phone, smoking, obstructing the camera and not wearing a seat belt.

5. The method of claim 1, wherein, Violation behavior data includes the type, time, location and duration of the violation behavior.

6. The method of claim 5, wherein, The cloud server evaluates the driver based on the violation behavior data to obtain a comprehensive score of the driver, comprising: The cloud server evaluates the driver based on a scoring model according to the violation behavior data to obtain a comprehensive score of the driver; The scoring model satisfies the following formula: ， wherein, is a combined score for the driver, is a weight for the type of violation, is a number of occurrences of the type of violation, is a total number of types of violations.

7. The method of claim 1, wherein, The method further comprises: The cloud server generates an evaluation report of the driver based on the comprehensive score, and sends the evaluation report to an administrator for safety management, driver training and performance evaluation.

8. The method of claim 1, wherein, The method further comprises: The cloud server obtains feedback from users and re-trains the lightweight behavior recognition model based on the feedback.

9. A mine driver behavior intelligent analysis and comprehensive scoring system, characterized in that, The method comprises a terminal device, an edge device and a cloud server; The terminal device is used to collect driver behavior data and vehicle driving information; The edge device is configured to determine whether the driver produces a violation behavior and whether the terminal device is abnormal according to the behavior data and vehicle driving information based on a trained lightweight behavior recognition model, wherein the lightweight behavior recognition model is trained on the cloud server, a convolution layer of a head network in the lightweight behavior recognition model is a depth separable convolution layer, and weights of all weight layers in the lightweight behavior recognition model are quantized weights; If the driver produces a violation behavior or the terminal device is abnormal, the terminal device is further configured to send a violation warning or an abnormality warning to the driver, and the edge device is further configured to save violation behavior data or terminal device abnormality data; The cloud server is configured to evaluate the driver according to the violation behavior data to obtain a comprehensive score of the driver; The quantized weights satisfy the following formula: ， wherein is the quantized weight of the weight layer, is the original weight of the weight layer, , are the maximum and minimum values of the quantization range, respectively, , are the maximum and minimum values of , respectively, is a function for rounding a scaled floating point value to an integer.

Images