A vehicle operation safety evaluation method and system

Abstract

The application discloses a kind of vehicle operation safety evaluation method and system, it is related to vehicle safety monitoring technical field, including, reconstructs the three-dimensional dynamic scene of cockpit using neural radiation field, mapping is carried out in combination with fixation point coordinates, generates attention heat map, collects the vehicle operating parameter of driver, and is converted into operation vector, the matching degree score of operation vector and attention heat map is calculated, historical driver behavior data and vehicle state data are collected to train MDNN architecture, obtain comprehensive safety evaluation model, vehicle state vector and matching degree score are input into comprehensive safety evaluation model, and vehicle operation safety score is generated.The application is ensured by vehicle operation safety evaluation method, the high reliability and adaptability of safety evaluation result, can effectively guarantee driving safety under different driving situations.

Description

Technical Field

[0001] This invention relates to the field of vehicle safety monitoring technology, and in particular to a method and system for assessing vehicle operation safety. Background Technology

[0002] With the rapid development of the automotive industry, vehicle safety has become an important indicator for measuring vehicle quality. In the past, vehicle operation safety assessment mainly relied on physical experiments and rule-based algorithms, including crash tests and behavioral analysis under simulated driving environments. However, in recent years, with the advancement of intelligent sensor technology, computing power, and artificial intelligence algorithms, especially the breakthroughs of deep learning in image recognition and data processing, data-driven vehicle operation safety assessment methods have gradually become possible. This new assessment method can not only monitor the vehicle's operating status in real time, but also predict potential risks through big data analysis, thereby greatly improving driving safety. For example, the onboard sensor network composed of LiDAR, GPS navigation, and cameras can accurately acquire information about the vehicle's surrounding environment. At the same time, neural network models can efficiently handle complex and changing road conditions, providing drivers with timely and accurate safety warnings.

[0003] Nevertheless, existing vehicle operation safety assessment methods still have room for improvement. Most current vehicle safety assessment mechanisms focus on hardware improvements, which undoubtedly increases costs and limits their widespread adoption. Secondly, they do not pay enough attention to driver behavior and fail to fully consider the impact of driver distraction on driving safety. Specifically, existing vehicle operation safety assessment methods usually only focus on the distance and relative speed between the vehicle and external obstacles, while ignoring the driver's reaction time and attention allocation in different situations. This poses a challenge to comprehensively assessing vehicle operation safety. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a vehicle operation safety assessment method to address the problem of insufficient consideration of the impact of driver attention on driving safety.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for assessing vehicle operational safety, comprising,

[0008] Use an onboard intelligent sensor array to collect vehicle location information, raw point cloud data, and signal control time;

[0009] Based on the vehicle's location information, raw point cloud data, and signal control time, the vehicle speed, the distance between the vehicle and the obstacle, and the response delay time of the vehicle's active safety control unit are calculated sequentially to form a vehicle state vector.

[0010] The three-dimensional dynamic scene of the cockpit is reconstructed using neural radiation fields, and then mapped using gaze point coordinates to generate an attention heatmap.

[0011] Collect the driver's vehicle operation parameters and convert them into operation vectors. Calculate the matching score between the operation vectors and the attention heatmap.

[0012] Collect historical driver behavior data and vehicle status data to train the MDNN architecture and obtain a comprehensive safety assessment model;

[0013] The vehicle state vector and matching score are input into the comprehensive safety assessment model to generate a vehicle operation safety score.

[0014] As a preferred embodiment of the vehicle operation safety assessment method of the present invention, the step of using an onboard intelligent sensor group to collect vehicle location information, raw point cloud data, and signal control time specifically involves:

[0015] Use GPS navigation to receive vehicle location information;

[0016] Raw point cloud data is acquired synchronously using LiDAR;

[0017] The vehicle controller area network bus is used to record the control time of signals emitted by the electronic control unit in real time.

[0018] As a preferred embodiment of the vehicle operation safety assessment method of the present invention, the method involves: calculating the vehicle speed, the distance between the vehicle and obstacles, and the response delay time of the vehicle's active safety control unit sequentially based on the vehicle's location information, raw point cloud data, and signal control time, thereby forming a vehicle state vector. Specifically,

[0019] The vehicle speed is obtained by calculating the position information of two adjacent frames.

[0020] After performing Euclidean clustering on the original point cloud data, the nearest neighbor obstacle point set is extracted, and the obstacle centroid coordinates are calculated.

[0021] The distance between the vehicle and the obstacle is deduced based on the centroid coordinates and the current vehicle position information.

[0022] Using the sliding window statistical method, the difference between all signal control times is sorted, and the Nth percentile value is taken as the response delay time of the vehicle active safety control unit.

[0023] The vehicle speed, the distance between the vehicle and the obstacle, and the response delay time of the vehicle's active safety control unit are aligned in time sequence to form a vehicle state vector.

[0024] As a preferred embodiment of the vehicle operation safety assessment method of the present invention, the step of reconstructing the three-dimensional dynamic scene of the cockpit using neural radiation fields and mapping it with gaze point coordinates to generate an attention heatmap is specifically as follows:

[0025] The neural radiation field is used to reconstruct a three-dimensional dynamic scene from a cockpit video sequence captured by an onboard multi-view camera;

[0026] The driver's gaze coordinates are captured using a near-infrared camera;

[0027] After synchronizing the gaze point coordinates with the 3D dynamic scene of the cockpit in time, a Gaussian mixture model is used to estimate the probability density of the gaze point coordinates to obtain an attention heatmap.

[0028] As a preferred embodiment of the vehicle operation safety assessment method of the present invention, the step of collecting the driver's vehicle operation parameters, converting them into operation vectors, and calculating the matching score between the operation vectors and the attention heatmap specifically involves...

[0029] The driver's vehicle operation parameters include steering wheel angle, brake pedal travel, and accelerator pedal opening.

[0030] Pack the steering wheel angle value, brake pedal travel value, and throttle opening value into an operation vector according to the timestamp;

[0031] Establish a mapping relationship between operation vectors and attention heatmaps, and perform 3D convolution operations within a time window to generate a matching score between operation vectors and attention heatmaps;

[0032] As a preferred embodiment of the vehicle operation safety assessment method of the present invention, the step of collecting historical driver behavior data and vehicle status data to train an MDNN architecture and obtain a comprehensive safety assessment model specifically involves...

[0033] Historical driver behavior data and vehicle status data were collected as a training set.

[0034] The training set is input into the MDNN architecture for training. During the training process, the MDNN architecture undergoes nonlinear transformation through a multi-layer fully connected network and activation function. At the same time, the weight coefficients are optimized according to the backpropagation algorithm to minimize the prediction error and complete the training of the MDNN architecture.

[0035] After training the MDNN architecture, cross-validation is used to evaluate its generalization ability, resulting in a comprehensive security evaluation model.

[0036] As a preferred embodiment of the vehicle operation safety assessment method of the present invention, the generation of vehicle operation safety score refers to inputting the vehicle state vector and matching degree score into the comprehensive safety assessment model after standardization processing, and the comprehensive safety assessment model performing feature fusion and nonlinear transformation on the vehicle state vector and matching degree score to obtain the vehicle operation safety score.

[0037] Secondly, the present invention provides a vehicle operation safety assessment system, comprising,

[0038] The data acquisition module uses an onboard intelligent sensor array to collect vehicle location information, raw point cloud data, and signal control time.

[0039] The calculation module, based on the vehicle's location information, raw point cloud data, and signal control time, sequentially calculates the vehicle speed, the distance between the vehicle and obstacles, and the response delay time of the vehicle's active safety control unit, forming a vehicle state vector.

[0040] The mapping module uses neural radiation fields to reconstruct the three-dimensional dynamic scene of the cockpit, and combines the gaze point coordinates for mapping to generate an attention heatmap;

[0041] The matching module collects the driver's vehicle operation parameters, converts them into operation vectors, and calculates the matching score between the operation vectors and the attention heatmap.

[0042] The training module collects historical driver behavior data and vehicle status data to train the MDNN architecture and obtain a comprehensive safety assessment model.

[0043] The safety scoring module inputs the vehicle state vector and matching degree score into the comprehensive safety assessment model to generate a vehicle operation safety score.

[0044] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the vehicle operation safety assessment method as described in the first aspect of the present invention.

[0045] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the vehicle operation safety assessment method as described in the first aspect of the present invention.

[0046] The beneficial effects of this invention are as follows: First, by using a vehicle operation safety assessment method, high-precision perception of the vehicle and its surrounding environment is achieved, providing not only detailed road condition information but also enhancing the ability to identify potential hazards. Second, by conducting in-depth analysis and processing of the collected data, the safety status of the vehicle can be accurately calculated, thus providing a basis for predicting possible dangerous situations. In addition, the introduction of a driver attention monitoring mechanism can dynamically understand the driver's state and issue warnings and take measures when necessary to prevent accidents caused by distraction. Finally, the comprehensive safety assessment model trained using a big data-driven method ensures the high reliability and adaptability of the assessment results, effectively guaranteeing driving safety in different driving situations. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 A flowchart for a vehicle operation safety assessment method;

[0049] Figure 2 A flowchart for forming the vehicle state vector;

[0050] Figure 3 A flowchart for generating a matching score;

[0051] Figure 4 A flowchart for generating vehicle operation safety scores. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a vehicle operation safety assessment method, including the following steps:

[0056] S1. Use the on-board intelligent sensor group to collect vehicle location information, raw point cloud data and signal control time;

[0057] Use GPS navigation to receive vehicle location information;

[0058] It should be noted that GPS navigation receives L1-band radio signals transmitted by navigation satellites. The GPS receiver analyzes the ephemeris data and pseudorange measurements in the L1-band radio signals to obtain the vehicle's current three-dimensional coordinates. The GPS navigation then matches the vehicle's current three-dimensional coordinates with digital map data to determine the vehicle's longitude, latitude, and elevation on the map, thus obtaining the vehicle's location information.

[0059] Raw point cloud data is acquired synchronously using LiDAR;

[0060] It should be noted that the lidar emits pulsed laser beams and achieves 360° field of view coverage in the horizontal direction and 30° field of view coverage in the vertical direction through mechanical rotation scanning. During the acquisition process, after the pulsed laser beams encounter obstacles and are reflected, the avalanche photodiode receiver captures the number of echoes and the reflection intensity. Simultaneously, the lidar's internal high-precision encoder records the horizontal azimuth and vertical elevation angles of each pulsed laser beam in real time. The attitude data provided by the lidar's built-in inertial measurement unit converts the horizontal azimuth and vertical elevation angles of each pulsed laser beam into XYZ coordinates, ultimately generating raw point cloud data including XYZ coordinates, reflection intensity, and the number of echoes.

[0061] The on-board controller area network bus is used to record the control time of signals sent by the electronic control unit in real time.

[0062] It should be noted that the electronic control unit (ECU) broadcasts steering control command messages and braking control command messages at a baud rate (e.g., 500 kbps) via the vehicle controller area network bus. Simultaneously, the gateway node listens for these steering and braking control command messages and captures the hardware timestamp of the transmission start time. It also synchronously receives the UTC time reference provided by the GPS navigation system and records the time deviation between the ECU command issuance time and the vehicle location information acquisition time as the signal control time.

[0063] S2. Based on the vehicle's location information, raw point cloud data, and signal control time, calculate the vehicle speed, the distance between the vehicle and the obstacle, and the response delay time of the vehicle's active safety control unit in sequence to form a vehicle state vector.

[0064] The vehicle speed is obtained by calculating the position information of two adjacent frames.

[0065] It should be noted that the vehicle's position information (longitude, latitude, and elevation values) and WGS-84 ellipsoid parameters are collected for two consecutive frames. These WGS-84 ellipsoid parameters, provided by GPS navigation, include the ellipsoid's semi-minor axis, ellipsoid correction factor, angular distance on the auxiliary sphere, and ellipsoidal correction terms for the angular distance. The longitude and latitude coordinates of the vehicle in two adjacent frames are converted to geodetic distances using the Vincenty inversion formula, expressed as:

[0066] ;

[0067] in, Indicates the distance to the geodetic line. Indicates the minor axis of the ellipsoid. Indicates the ellipsoid correction factor. Indicates the angular distance on the auxiliary sphere. The ellipsoidal correction term represents the angular distance;

[0068] Based on the distance to the earth line and the difference in timestamps between two adjacent vehicle frames, the horizontal displacement is calculated using the following formula:

[0069] ;

[0070] in, Indicates the horizontal displacement. Indicates the distance to the geodetic line. Indicates the timestamp difference;

[0071] The vertical displacement is calculated based on the elevation values ​​of two adjacent frames of the vehicle, and is expressed by the following formula:

[0072] ;

[0073] in, This represents the vertical displacement of the vehicle between two adjacent frames. This represents the vehicle's elevation value in the first frame. This indicates the vehicle's elevation value in the second frame;

[0074] The instantaneous vehicle displacement vector is obtained by combining the distance to the geodetic line, the horizontal displacement, and the vertical displacement in three dimensions. The vehicle speed is obtained by subtracting the instantaneous displacement vector from the time interval. The time interval is determined by the timestamps of two consecutive frames of position information provided by GPS navigation. The vehicle speed calculation result includes eastward, northward, and vertical components.

[0075] After performing Euclidean clustering on the original point cloud data, the nearest neighbor obstacle point set is extracted, and the obstacle centroid coordinates are calculated.

[0076] It should be noted that the raw point cloud data collected by the LiDAR is first subjected to pass-through filtering to retain the point cloud in the area in front of the vehicle (e.g., 20m × 10m). At the same time, a KD-Tree spatial index structure is used to accelerate the neighborhood search, and the Euclidean distance threshold is set to 0.5m as the clustering criterion. The Euclidean distance threshold of 0.5m is set because the average point cloud spacing of the LiDAR within the 20m ranging range is 0.2-0.3m. 0.5m can ensure that adjacent points on the same obstacle surface can be correctly clustered, while effectively separating independent obstacles with a spacing greater than 0.5m.

[0077] The region growing algorithm is used to aggregate point clouds that meet the Euclidean distance threshold into independent obstacle clusters. Obstacle clusters with more than 50 points are selected as effective nearest neighbor obstacle point sets. The arithmetic mean of the XYZ coordinates of all points in the effective nearest neighbor obstacle point set is calculated to obtain the obstacle centroid coordinates.

[0078] The distance between the vehicle and the obstacle is deduced based on the centroid coordinates and the current vehicle position information.

[0079] It should be noted that the vehicle location information and obstacle center coordinates provided by GPS navigation are converted into Cartesian coordinates using a coordinate transformation method. When the obstacle center coordinates and the vehicle position are in the same coordinate system, the straight-line distance between the vehicle position point and the obstacle center point on the horizontal plane is calculated. Based on this straight-line distance, and combined with the difference between the vehicle's altitude and the obstacle center point's height, the distance between the vehicle and the obstacle is obtained. The straight-line distance between the vehicle position point and the obstacle center point on the horizontal plane is expressed by the formula:

[0080] ;

[0081] in, This represents the straight-line distance on the horizontal plane between the vehicle's position and the center of gravity of the obstacle. The X-axis coordinates represent the center of gravity of the obstacle. The Y-axis coordinate represents the center of gravity of the obstacle. The X-axis coordinate represents the vehicle's position. The Y-axis coordinate representing the vehicle's position;

[0082] The distance between a vehicle and an obstacle can be expressed by the formula:

[0083] ;

[0084] ;

[0085] in, Indicates the distance between the vehicle and the obstacle. This represents the straight-line distance on the horizontal plane between the vehicle's position and the center of gravity of the obstacle. This represents the difference between the vehicle's altitude and the center of gravity of the obstacle. Indicates the vehicle's altitude. Indicates the height of the obstacle material;

[0086] Using the sliding window statistical method, the difference between all signal control times is sorted, and the Nth percentile value is taken as the response delay time of the vehicle active safety control unit.

[0087] It should be noted that the signal control times continuously recorded by the vehicle controller area network bus are stored in a circular buffer in chronological order. A sliding window sample set is formed using the most recent (e.g., 100 sets) signal control times. The time interval difference is calculated for each pair of adjacent signal control times in the sliding window sample set. After sorting all time interval differences in ascending order, the time difference corresponding to the 95th percentile of the sequence is selected as the response delay time of the vehicle active safety control unit. It should be noted that the signal transmission of the vehicle controller area network bus may experience extreme delays due to electromagnetic interference and network congestion. The 95th percentile effectively filters out the first 5% of abnormally high-latency data, avoiding the impact of occasional interference on the accuracy of the assessment.

[0088] The vehicle speed, the distance between the vehicle and the obstacle, and the response delay of the vehicle's active safety control unit are aligned in time sequence to form a vehicle state vector;

[0089] Furthermore, millisecond-level timestamps are added to vehicle speed, distance to obstacles, and response delay of the vehicle's active safety control unit. These timestamps are then synchronously interpolated onto a fixed time grid. Each time grid contains a three-dimensional tuple of these three parameters, which are then arranged in chronological order to form a continuous vehicle state vector.

[0090] S3. Reconstruct the three-dimensional dynamic scene of the cockpit using neural radiation fields, and map it in combination with gaze point coordinates to generate an attention heatmap;

[0091] The neural radiation field is used to reconstruct a three-dimensional dynamic scene from a cockpit video sequence captured by an onboard multi-view camera;

[0092] It should be noted that the multi-view camera installed in the cockpit simultaneously captures RGB video sequences from different perspectives of the driver. Each frame in the RGB video sequence is accompanied by a timestamp and internal / external parameters of the camera. The RGB video sequence is input into a neural radiation field, which establishes a mapping relationship between three-dimensional spatial coordinates and color and volume density through a multilayer perceptron network. Simultaneously, new images from arbitrary perspectives are synthesized from the RGB video sequence using a differentiable volume rendering method. By comparing the differences between the new images from arbitrary perspectives and the actual images captured by the multi-view camera, the network parameters are optimized in reverse. After 30 minutes of online training, the neural radiation field can generate a three-dimensional dynamic scene inside the cockpit in real time.

[0093] The reason for choosing 30 minutes of training is that the typical cycle of limb movement in the cockpit is 5-10 seconds, and 30 minutes of training includes 180-360 movement cycles, which is sufficient to capture the statistical regularity of the driver's dynamic characteristics.

[0094] The driver's gaze coordinates are captured using a near-infrared camera;

[0095] It should be noted that after the near-infrared camera is installed in the cockpit facing the driver's face, it emits near-infrared light to illuminate the driver's face. A high-sensitivity sensor captures the reflection point of the driver's eyeball. The near-infrared camera simultaneously acquires images of the driver's eyes. These images are transmitted to the image processing unit, where the Viola-Jones algorithm locates the coordinates of the driver's pupil center. The eye-tracking algorithm calculates the coordinates of the driver's gaze point based on the relative position of the pupil center coordinates and the reflection point of the driver's eyeball. After smoothing and denoising the pupil center coordinates and the driver's gaze point coordinates through a Kalman filter, the final gaze point coordinates are output.

[0096] After synchronizing the gaze point coordinates with the 3D dynamic scene of the cockpit in time, a Gaussian mixture model is used to estimate the probability density of the gaze point coordinates to obtain an attention heatmap.

[0097] Furthermore, after synchronizing the gaze point coordinates with the 3D dynamic scene of the cockpit via timestamps, historical gaze point coordinates are collected as a training set, and the hyperparameters of the Gaussian mixture model (mixture weight coefficients, mean vector, and covariance matrix parameters) are initialized. The training set and the hyperparameters of the Gaussian mixture model are then input into the expectation-maximization algorithm for iterative optimization: In the expectation step, the posterior probability that the historical gaze point coordinates in the training dataset belong to a Gaussian distribution is calculated based on the hyperparameters of the Gaussian mixture model; in the maximization step, the mixture weight coefficients, mean vector, and covariance matrix parameters are updated using the posterior probability of the Gaussian distribution to obtain the optimized hyperparameters (mixture weight coefficients, mean vector, and covariance matrix parameters). The trained Gaussian mixture model performs probability density estimation on the gaze point coordinates, generating continuous probability density values. Specifically, the trained Gaussian mixture model receives the gaze point coordinates and, combined with the optimized mixture weight coefficients, mean, and covariance matrix parameters during training, calculates the probability density of each gaze point coordinate under the mixture Gaussian distribution. The probability densities under each Gaussian mixture distribution are weighted and summed according to the mixture weight coefficients to output continuous probability density values. The mesh vertices of the cockpit 3D dynamic scene are then color-mapped based on the probability density values ​​to generate an attention heatmap.

[0098] S4. Collect the driver's vehicle operation parameters and convert them into operation vectors. Calculate the matching score between the operation vectors and the attention heatmap.

[0099] Pack the steering wheel angle value, brake pedal travel value, and throttle opening value into an operation vector according to the timestamp;

[0100] It should be noted that when the driver engages the vehicle, the steering wheel angle is measured in real time by an angle sensor mounted on the steering column, the brake pedal travel is recorded by a displacement sensor, and the throttle position sensor is used to obtain the throttle opening value. The steering wheel angle, brake pedal travel, and throttle opening values ​​are collected at fixed sampling time points. The steering wheel angle, brake pedal travel, and throttle opening values ​​at each sampling time point are combined into a three-dimensional vector and arranged in chronological order according to their timestamps to form an operation vector.

[0101] Establish a mapping relationship between operation vectors and attention heatmaps, and perform 3D convolution operations within a time window to generate a matching score between operation vectors and attention heatmaps;

[0102] Furthermore, the operation vectors are spatiotemporally aligned with the attention heatmap to ensure a perfect temporal match. Then, within a certain time window, a 3D convolution operation is performed between the 3D features (steering wheel angle, brake pedal travel, and throttle opening values) from the operation vectors and the attention heatmap. Specifically, the 3D convolution kernel performs element-wise multiplication and summation of the 3D features and the attention heatmap along the time axis with a fixed step size at each time window position, generating a spatiotemporal local correlation feature map. The 3D convolution kernel extracts the spatial distribution correlation features between the operation vectors and the attention heatmap in the spatiotemporal local correlation feature map, and transforms them through an activation function to form spatiotemporal feature slices. All spatiotemporal feature slices are stacked along the time axis to form a spatiotemporal feature sequence. After the spatiotemporal feature sequence is compressed in time dimension by a global max pooling layer, it is mapped to scalar values ​​through a fully connected layer. The scalar values ​​are normalized to a matching score in the range of 0 to 1 using the Sigmoid function.

[0103] S5. Collect historical driver behavior data and vehicle status data to train the MDNN architecture and obtain a comprehensive safety assessment model;

[0104] Historical driver behavior data and vehicle status data were collected as a training set.

[0105] It should be noted that historical driver behavior data and vehicle status data are collected and packaged into data segments (e.g., 10 minutes). Each data segment is labeled with the actual driving scenario type (urban road / highway / curve) and safety level label. After data cleaning and outlier removal, a training set covering different climate conditions and traffic conditions is obtained.

[0106] The training set is input into the MDNN architecture for training. During the training process, the MDNN architecture undergoes nonlinear transformation through a multi-layer fully connected network and activation function. At the same time, the weight coefficients are optimized according to the backpropagation algorithm to minimize the prediction error and complete the training of the MDNN architecture.

[0107] It should be noted that the MDNN architecture receives a training set, which is then normalized by the input layer and fed into five fully connected layers. Each layer uses the LeakyReLU activation function for nonlinear transformation. Batch normalization is introduced in the intermediate layers to accelerate convergence. The output layer generates a safety score prediction value in the 0-1 range using the Sigmoid function. During training, the mean squared error loss function is used to calculate the deviation between the predicted value and the safety level label. Based on the deviation, the Adam optimizer is used to execute the backpropagation algorithm to adjust the network weight coefficients of the MDNN architecture. Training is terminated when the number of training epochs reaches 200, resulting in the final trained MDNN architecture. The training epochs follow an early stopping mechanism; after more than 200 training epochs, the accuracy of the MDNN architecture begins to decrease at a rate of 0.1% per epoch. Continuing training at this point will reduce the generalization ability of the MDNN architecture.

[0108] After training the MDNN architecture, cross-validation is used to evaluate the generalization ability of the MDNN architecture, and a comprehensive security evaluation model is obtained.

[0109] Furthermore, the training set is divided into 5 mutually exclusive subsets using the K-fold cross-validation method. One mutually exclusive subset is selected as the validation set for each round of validation (a total of 5 rounds of validation). The remaining 4 mutually exclusive subsets are used to retrain the MDNN architecture, maintaining the same hyperparameter configuration as the main training phase. In each validation process, the absolute error and recall between the predicted value of the MDNN architecture on the validation set and the security level label are calculated. After 5 rounds of cross-validation, the MDNN architecture with an absolute error between the predicted value and the validation set of less than 0.05 and a recall rate of more than 92% is selected from the results of the 5 rounds of cross-validation as the comprehensive security assessment model.

[0110] The selection of an absolute error below 0.05 is to ensure the effectiveness of distinguishing between subsequent adjacent safety levels (such as 0.6 and 0.65). A recall rate above 92% is chosen because it meets the fault detection coverage requirements for ASIL-B level safety functions; furthermore, when the recall rate is below 92%, the false positive rate for nighttime operation increases threefold.

[0111] S6. Input the vehicle state vector and matching degree score into the comprehensive safety assessment model to generate a vehicle operation safety score;

[0112] Generating a vehicle operation safety score refers to inputting the vehicle state vector and matching degree score into a comprehensive safety assessment model after standardization. The comprehensive safety assessment model then performs feature fusion and nonlinear transformation on the vehicle state vector and matching degree score to obtain the vehicle operation safety score.

[0113] It should be noted that the vehicle state vector and the matching score are both Z-score standardized to eliminate dimensional differences before being input into the comprehensive safety assessment model. First, the vehicle state vector and the matching score are concatenated into a joint feature vector through the feature fusion layer of the comprehensive safety assessment model. Then, it undergoes a nonlinear transformation through five fully connected layers, with each layer using the LeakyReLU activation function. Finally, the output layer maps the joint feature vector to a vehicle operation safety score in the 0-1 range using the Sigmoid function. Based on the vehicle operation safety score, a three-level progressive warning strategy is adopted, such as primary warning (0.5 ≤ vehicle operation safety score < 0.6), intermediate warning (0.4 ≤ vehicle operation safety score < 0.5), and advanced warning (vehicle operation safety score < 0.4).

[0114] This embodiment also provides a vehicle operation safety assessment system, including:

[0115] The data acquisition module uses an onboard intelligent sensor array to collect vehicle location information, raw point cloud data, and signal control time.

[0116] The calculation module, based on the vehicle's location information, raw point cloud data, and signal control time, sequentially calculates the vehicle speed, the distance between the vehicle and obstacles, and the response delay time of the vehicle's active safety control unit, forming a vehicle state vector.

[0117] The mapping module uses neural radiation fields to reconstruct the three-dimensional dynamic scene of the cockpit, and combines the gaze point coordinates for mapping to generate an attention heatmap;

[0118] The matching module collects the driver's vehicle operation parameters, converts them into operation vectors, and calculates the matching score between the operation vectors and the attention heatmap.

[0119] The training module collects historical driver behavior data and vehicle status data to train the MDNN architecture and obtain a comprehensive safety assessment model.

[0120] The safety scoring module inputs the vehicle state vector and matching degree score into the comprehensive safety assessment model to generate a vehicle operation safety score.

[0121] This embodiment also provides a computer device applicable to the vehicle operation safety assessment method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the vehicle operation safety assessment method proposed in the above embodiment.

[0122] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0123] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the vehicle operation safety assessment method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0124] In summary, this invention achieves high-precision perception of vehicles and their surrounding environment through a vehicle operation safety assessment method. This not only provides detailed road condition information but also enhances the ability to identify potential hazards. Secondly, by conducting in-depth analysis and processing of the collected data, the safety status of the vehicle can be accurately calculated, thus providing a basis for predicting possible dangerous situations. Furthermore, the introduction of a driver attention monitoring mechanism can dynamically understand the driver's state and issue warnings and take measures when necessary to prevent accidents caused by inattention. Finally, the comprehensive safety assessment model trained using a big data-driven method ensures the high reliability and adaptability of the assessment results, effectively guaranteeing driving safety in different driving scenarios.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for assessing vehicle operational safety, characterized in that: include, Use an onboard intelligent sensor array to collect vehicle location information, raw point cloud data, and signal control time; Based on the vehicle's location information, raw point cloud data, and signal control time, the vehicle speed, the distance between the vehicle and obstacles, and the response delay time of the vehicle's active safety control unit are calculated sequentially to form a vehicle state vector. The specific steps are as follows: The vehicle speed is obtained by calculating the position information of two adjacent frames. After performing Euclidean clustering on the original point cloud data, the nearest neighbor obstacle point set is extracted, and the obstacle centroid coordinates are calculated. The distance between the vehicle and the obstacle is deduced based on the centroid coordinates and the current vehicle position information. Using the sliding window statistical method, the difference between all signal control times is sorted, and the Nth percentile value is taken as the response delay time of the vehicle active safety control unit. The vehicle speed, the distance between the vehicle and the obstacle, and the response delay of the vehicle's active safety control unit are aligned in time sequence to form a vehicle state vector; The three-dimensional dynamic scene of the cockpit is reconstructed using neural radiation fields, and then mapped using gaze point coordinates to generate an attention heatmap. The specific steps are as follows: The neural radiation field is used to reconstruct a three-dimensional dynamic scene from a cockpit video sequence captured by an onboard multi-view camera; The driver's gaze coordinates are captured using a near-infrared camera; After synchronizing the gaze point coordinates with the 3D dynamic scene of the cockpit in time, a Gaussian mixture model is used to estimate the probability density of the gaze point coordinates to obtain an attention heatmap. Collect the driver's vehicle operation parameters and convert them into operation vectors. Calculate the matching score between the operation vectors and the attention heatmap. Collect historical driver behavior data and vehicle status data to train the MDNN architecture and obtain a comprehensive safety assessment model; The vehicle state vector and matching score are input into the comprehensive safety assessment model to generate a vehicle operation safety score.

2. The vehicle operation safety assessment method according to claim 1, characterized in that: The specific steps involve using an onboard intelligent sensor array to collect vehicle location information, raw point cloud data, and signal control time. Use GPS navigation to receive vehicle location information; Raw point cloud data is acquired synchronously using LiDAR; The vehicle controller area network bus is used to record the control time of signals emitted by the electronic control unit in real time.

3. The vehicle operation safety assessment method according to claim 2, characterized in that: The process involves collecting the driver's vehicle operation parameters, converting them into operation vectors, and calculating a matching score between the operation vectors and the attention heatmap. The driver's vehicle operation parameters include steering wheel angle, brake pedal travel, and accelerator pedal opening. Pack the steering wheel angle value, brake pedal travel value, and throttle opening value into an operation vector according to the timestamp; Establish a mapping relationship between operation vectors and attention heatmaps, and perform 3D convolution operations within a time window to generate a matching score between operation vectors and attention heatmaps.

4. The vehicle operation safety assessment method according to claim 3, characterized in that: The process involves collecting historical driver behavior data and vehicle status data to train an MDNN architecture and obtain a comprehensive safety assessment model. Specifically... Historical driver behavior data and vehicle status data were collected as a training set. The training set is input into the MDNN architecture for training. During the training process, the MDNN architecture undergoes nonlinear transformation through a multi-layer fully connected network and activation function. At the same time, the weight coefficients are optimized according to the backpropagation algorithm to minimize the prediction error and complete the training of the MDNN architecture. After training the MDNN architecture, cross-validation is used to evaluate its generalization ability, resulting in a comprehensive security evaluation model.

5. The vehicle operation safety assessment method according to claim 4, characterized in that: The generated vehicle operation safety score refers to the standardized processing of the vehicle state vector and matching degree score, which are then input into the comprehensive safety assessment model. The comprehensive safety assessment model performs feature fusion and nonlinear transformation on the vehicle state vector and matching degree score to obtain the vehicle operation safety score.

6. A vehicle operation safety assessment system, based on the vehicle operation safety assessment method according to any one of claims 1 to 5, characterized in that: include, The data acquisition module uses an onboard intelligent sensor array to collect vehicle location information, raw point cloud data, and signal control time. The calculation module, based on the vehicle's location information, raw point cloud data, and signal control time, sequentially calculates the vehicle speed, the distance between the vehicle and obstacles, and the response delay time of the vehicle's active safety control unit, forming a vehicle state vector. The mapping module uses neural radiation fields to reconstruct the three-dimensional dynamic scene of the cockpit, and combines the gaze point coordinates for mapping to generate an attention heatmap; The matching module collects the driver's vehicle operation parameters, converts them into operation vectors, and calculates the matching score between the operation vectors and the attention heatmap. The training module collects historical driver behavior data and vehicle status data to train the MDNN architecture and obtain a comprehensive safety assessment model. The safety scoring module inputs the vehicle state vector and matching degree score into the comprehensive safety assessment model to generate a vehicle operation safety score.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the vehicle operation safety assessment method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the vehicle operation safety assessment method according to any one of claims 1 to 5.

Images