Method for automatically measuring driving sight distance at highway curve

Abstract

The invention discloses a method for automatically measuring the driving sight distance of a highway curve. The method comprises the following steps: S1, acquiring three-dimensional space data and vehicle motion parameters of a curve area through a multi-modal sensor; the method comprises the steps of S1, obtaining three-dimensional space data and vehicle motion parameters, S2, carrying out multi-source data fusion on the three-dimensional space data and the vehicle motion parameters, S3, constructing a driving safety sight distance calculation model based on a dynamic vehicle kinematics model, and S4, dynamically correcting a safety sight distance calculation result according to real-time environment parameters. In the frequency domain, double-domain filtering is guided through attention to effectively suppress interference, the purity and accuracy of signals are improved, compared with a traditional single-domain processing method, the double-domain filtering mode can more accurately remove signal interference generated by dynamic interference sources such as construction marks, the reliability of subsequent data processing and decision making is ensured, and the method is suitable for popularization and application. More accurate data support is provided for curve identification and sight distance measurement, and the robustness of the system in a complex environment is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed intelligent driving, and in particular to a method for automatically measuring sight distance when driving on a curved highway. Background Art

[0002] Existing technologies for highway sight-distance measurement suffer from numerous shortcomings. Traditional technologies often rely on single sensors, such as cameras, lidar, or millimeter-wave radar. These technologies suffer from poor robustness in complex environments, such as rain, snow, fog, and low light conditions at night. For example, cameras alone are prone to failure in backlighting and occlusion, while millimeter-wave radars have low resolution and struggle to accurately identify curve edge features.

[0003] In terms of dynamic calibration of multi-sensor fusion, existing solutions have the following problems: First, static calibration has obvious limitations. For example, some traditional calibration methods rely on pre-static calibration parameters, such as the pitch angle and focal length of the camera. However, these parameters may be offset due to factors such as bumps and load changes during vehicle driving. The existing technology lacks an effective real-time dynamic calibration mechanism, resulting in serious error accumulation in long-term use; second, cross-sensor collaboration is insufficient. Although some multi-sensor calibration methods, such as the iKalibr framework, can jointly calibrate IMUs and radars, these methods often require manual intervention or preset targets, and fail to properly solve the key problem of spatiotemporal synchronization of sensor data during dynamic driving, which greatly reduces the fusion accuracy.

[0004] Traditional curve sight-distance feature point extraction algorithms that are adaptive to curve curvature have poor adaptability to curvature changes and suffer from the following drawbacks: First, they rely on fixed models. Some methods, for example, based on preset curve straight line models, divide the image into straight / curved areas with a fixed ratio. This makes them incapable of detecting curves of varying radii, such as continuous S-bends or sharp turns, resulting in large lane line fitting errors. Second, they are limited by geometric projection. Methods based on Hough transform or inverse perspective transform rely on the assumption of an ideal road plane, ignoring the actual three-dimensional curvature changes of the curve, resulting in significant feature point offsets after projection. Third, dynamic obstacle interference. In real-world scenarios, interference factors such as temporary construction signs and vehicle lane changes often disrupt the continuity of sight-distance feature points, making it difficult for traditional filtering methods, such as median filtering, to effectively remove noise. Summary of the Invention

[0005] In view of the above shortcomings of the prior art, the present invention provides a method for automatically measuring the sight distance when driving on a curved highway.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for automatically measuring sight distance on a curved highway, comprising the following steps: S1: Collecting three-dimensional spatial data and vehicle motion parameters of the curve area through a multimodal sensor, wherein the multimodal sensor includes a laser radar, a camera device, and an inertial measurement unit; S2: Curve recognition, performing multi-source data fusion on the three-dimensional spatial data and vehicle motion parameters to extract curve curvature, obstacle distribution, and lane line geometry features; S3: Constructing a safe driving sight distance calculation model based on a dynamic vehicle kinematic model, which integrates curve geometry parameters, vehicle real-time speed, and braking response time; S4: Dynamically correct the safe sight distance calculation result according to the real-time environmental parameters, generate a driving warning signal and output it to the vehicle control system.

[0007] Furthermore, the dynamic configuration of the multimodal sensor includes: the laser radar adaptively adjusting the scanning density according to the curvature of the curve; The camera device uses image optimization algorithms to enhance lane edge feature capture; The sampling frequency of the inertial measurement unit is dynamically matched to the vehicle acceleration.

[0008] Furthermore, the dynamic configuration method also includes triggering the curve enhancement scanning mode of the laser radar when it is detected that the curvature of the curve exceeds a set threshold, and dynamically adjusting the color space conversion parameters of the camera device according to the ambient light intensity.

[0009] Furthermore, the multi-source data fusion in S2 uses a spatiotemporal registration algorithm to align the lidar point cloud and visual image data, and uses a filtering algorithm to fuse inertial measurement data and lane line geometric features to compensate for measurement errors, establish an obstacle motion trajectory prediction model, and dynamically update the threat level.

[0010] Furthermore, the safety sight distance calculation model formula in S3 is:

[0011] Among them, v is the real-time speed of the vehicle, μ is the road adhesion coefficient, θ is the curve slope angle, t reaction is the driver's reaction time, L vehicle For the safety margin distance, the functional relationship is determined by the vehicle dynamics constraints.

[0012] Furthermore, the position information of traffic participants within the set range ahead is obtained through vehicle-road cooperative communication, and the road adhesion coefficient is updated in real time by integrating meteorological sensor data, and the calculation weight of the reaction time is adjusted according to the driver's gaze direction.

[0013] Furthermore, the driving warning signal logic in S4 is to trigger a multi-level progressive warning when the measured sight distance value is lower than the safety threshold: Primary warning, showing the safe curve area boundary through the head-up display; Intermediate warning, applying tactile feedback; Advanced warning, linked to the vehicle control system to implement active braking.

[0014] Furthermore, it also includes data verification, cross-modal consistency check of multi-sensor data and elimination of abnormal data. When a single sensor fails, the line-of-sight prediction mode based on historical data is activated.

[0015] Furthermore, the curve recognition also includes using a region growing algorithm to segment the curve area point cloud data, defining seed points based on the lane line curvature characteristics, and reconstructing broken lane lines and eliminating interference features through an interpolation algorithm.

[0016] Furthermore, the road adhesion coefficient μ is dynamically updated through the tire-road friction model, and its calculation formula is:

[0017] 0.8 is a constant term, which represents the reference value of the road adhesion coefficient under standard conditions; -0.003T is a term that is proportional to the tire temperature T, where T is the tire temperature, and the data comes from the tire pressure monitoring sensor.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. The spatiotemporal registration algorithm of the present invention aligns data from different sensors and effectively suppresses interference in the frequency domain through attention-guided dual-domain filtering, improving signal purity and accuracy. Compared with traditional single-domain processing methods, this dual-domain filtering approach can more accurately remove signal interference generated by dynamic interference sources such as construction signs, ensuring the reliability of subsequent data processing and decision-making, providing more accurate data support for curve recognition and line-of-sight measurement, and enhancing the system's robustness in complex environments.

[0019] 2. The present invention's safe sight distance calculation model, constructed based on a dynamic vehicle kinematic model, comprehensively considers multiple factors, including curve geometry, vehicle real-time speed, and brake response time. Compared to models with fixed parameters, this model more accurately reflects dynamic changes during actual driving, calculates the center of mass slip angle in real time, and adjusts the safe sight distance accordingly, providing more accurate warning information to the driver or vehicle control system, effectively improving driving safety. Furthermore, the road adhesion coefficient is dynamically updated through the tire-road friction model, further enhancing the accuracy of safe sight distance calculation, ensuring that the vehicle maintains an appropriate safe distance under varying road conditions.

[0020] 3. The proposed Joint Temporal and Spatial Synchronous Calibration Network (JTSN) achieves dynamic calibration of sensor parameters through the combined optimization of an extended Kalman filter and an online sparse autoencoder. Compared to traditional static calibration methods, this dynamic calibration mechanism can compensate for sensor parameter drift caused by factors such as vehicle turbulence and load changes during driving in real time, thus avoiding error accumulation, improving the precision of multi-sensor data fusion, and ensuring the accuracy of sight distance measurements on curved roads.

[0021] 4. The Dynamic Heterogeneous Compute Engine (DHCE) proposed in this invention combines FPGA parallel acceleration with neural network model compression technology to achieve dual optimization of algorithm complexity and hardware resources. By optimizing task allocation using a genetic algorithm and dynamically pruning the model, this improves system efficiency and reduces computing resource usage while ensuring data processing accuracy. This enables the system to more quickly respond to dynamic changes in vehicle movement, providing a powerful guarantee for real-time sight distance measurement and early warning during cornering.

[0022] 5. The Multi-Scale Bezier-Transformer Network (MBT-Net) constructed in this invention addresses the poor adaptability of traditional geometric projection methods to curvature changes. By dividing curves into differentiable piecewise Bezier curves and dynamically adjusting control points using deep reinforcement learning, this method achieves adaptive modeling and precise recognition of curves of varying radii. Furthermore, bicubic spline interpolation is used to reconstruct broken lane lines and eliminate interfering features, further improving the accuracy and reliability of curve recognition and providing more accurate curve geometry information for sight distance measurement. DETAILED DESCRIPTION

[0023] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with 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 making creative efforts are within the scope of protection of the present invention.

[0024] Example 1: This example provides a method for automatically measuring sight distance on a curved highway, comprising the following steps: S1: Collecting three-dimensional spatial data and vehicle motion parameters of the curve area through a multimodal sensor, wherein the multimodal sensor includes a lidar, a camera device, and an inertial measurement unit (IMU); The laser radar uses a 1550nm wavelength MEMS scanning laser radar. A non-uniform scanning strategy is set in the curved area (horizontal field of view angle 120°, vertical field of view angle 30°). When the curvature radius of the curve is less than 300m, the enhanced scanning mode is triggered (the point cloud density is increased to 0.1°×0.1°). The trigger logic of the enhanced scanning mode is based on the real-time curvature estimation error threshold control. When the curvature radius estimation standard deviation σR>5% (corresponding to the confidence interval ±2σ), the enhanced scanning is activated. The scanning density improvement formula is:

[0025] Where R is the curvature radius (m), k = 0.03 is the adjustment coefficient to achieve a Sigmoid-type smooth transition of the scanning resolution; The camera is equipped with an 8-megapixel global shutter CMOS sensor, which uses an adaptive exposure algorithm (dynamic range of 120dB) to optimize the capture of corner edges. In HDR mode, it simultaneously collects visible light and near-infrared band data (850nm) to cope with backlit scenes. The IMU integrates a three-axis accelerometer (±16g), a gyroscope (±2000dps), and a magnetometer. The sampling frequency is dynamically adjusted based on the vehicle's lateral acceleration (increased to 200Hz when the acceleration is greater than 0.3g). Dynamic frequency adjustment is implemented on the Xilinx Zynq UltraScale+ MPSoC's PL side, building a double-buffered ping-pong architecture. When the acceleration threshold is triggered, the DMA channel automatically switches to 200Hz sampling mode, increasing data throughput to 8.4GB / s and keeping latency within 5μs. The sensor clock is synchronized in microseconds through the PTPv2 protocol, and a hard synchronization trigger device is used to align the lidar point cloud and camera image acquisition at the beginning of each frame. The IMU data is fused with the wheel speed meter information through Kalman filtering to compensate for the coordinate offset caused by the vehicle's pitch / roll.

[0026] S2: Curve recognition, performing multi-source data fusion on the three-dimensional spatial data and vehicle motion parameters to extract curve curvature, obstacle distribution, and lane line geometry features; Multi-source data fusion uses a spatiotemporal registration algorithm to align lidar point clouds and visual image data. A filtering algorithm is then used to fuse inertial measurement data with lane geometry to compensate for measurement errors. This allows for the establishment of an obstacle trajectory prediction model and dynamic update of the threat level. In the spatiotemporal registration algorithm, by integrating attention-guided dual-domain filtering, interference suppression is effectively implemented in the frequency domain to improve signal purity and accuracy. The specific implementation method is as follows:

[0027] in: S(f) is the frequency domain representation of the original signal, which includes the target signal and various interference components; (f) is the signal obtained after filtering, which effectively suppresses interference and more clearly highlights the target signal characteristics; G(f) is an adaptive frequency-domain mask. It is not fixed in advance but is dynamically generated based on the motion characteristics of the obstacle. For example, when an obstacle moves at a specific speed and acceleration, it will exhibit a unique distribution pattern in the frequency domain. The algorithm can analyze these characteristics in real time and adjust the shape and amplitude of G(f) to ensure that the mask accurately covers the target signal area while avoiding the frequency band of interfering signals. αk is the interference coefficient, representing the strength of the kth interference source. These coefficients are estimated in real time using the Kalman filter. The Kalman filter uses its prediction-update mechanism to continuously refine its estimate of interference strength. This allows for rapid adjustments even when interference source strength fluctuates, ensuring the stability and adaptability of the filtering effect.

[0028] δ(ff k ) is the Dirac function, which is used to represent the frequency f k The interference source is at frequency f k The value at is infinite and zero at other frequencies, thus accurately locating the interference source in the frequency domain.

[0029] This frequency-domain interference suppression method based on attention-guided dual-domain filtering can effectively filter out signal interference from dynamic interference sources such as construction signs, improve the robustness and accuracy of the spatiotemporal registration algorithm in complex environments, and provide a more reliable signal foundation for subsequent data processing and decision-making. The frequency domain mask G(f) is generated using dynamic prediction using a convolutional neural network. The network input includes the 40-dimensional Mel cepstral coefficients of the current frame signal spectrum S(f) and the historical movement trajectory of the interference source frequency points (tracked by particle filtering). The network structure consists of three layers of 1D convolution (with kernel size 5 and number of channels 64) followed by LSTM time series modeling. The output layer uses Sigmoid activation to generate mask values in the range of 0-1.

[0030] Voxel filtering of the lidar point cloud (voxel size 0.1m 3 ) noise reduction, segmenting the curve area through the improved region growing algorithm (the seed point is the lane line curvature mutation point, and the growth threshold Δ curvature is > 0.05m -1 ); Bicubic spline interpolation is used to reconstruct broken lane lines and eliminate interference from bridge guardrails (curvature continuity constraint: second-order derivative error < 0.1m) -2 ); Using a spatiotemporal registration algorithm, the lidar point cloud and visual image are aligned. Lane line pixels are separated using the HSV color space (yellow saturation threshold: S>0.7, V>0.6). IMU angular velocity data is integrated to compensate for point cloud distortion caused by vehicle yaw, and an obstacle trajectory prediction model is established (Kalman filtering combined with an interactive multi-model algorithm). The obstacle trajectory prediction model consists of three sub-models: Constant velocity model (CV), Q CV =diag([0.5 2 ,0.1 2 ]); Constant acceleration model (CA), Q CA =diag([0.8 2 ,0.3 2 ,0.2 2 ]); Coordinated turning model (CT), ω-U(-0.3, 0.3)rad / s; The transition probability matrix is:

[0031] S3: Constructing a safe driving sight distance calculation model based on a dynamic vehicle kinematic model, which integrates curve geometry parameters, vehicle real-time speed, and braking response time; The bicycle model is used to derive the lateral dynamics equation and calculate the sideslip angle of the center of mass in real time:

[0032] l f is the distance from the front axle to the center of mass of the vehicle, that is, the front wheelbase; l r is the distance from the rear axle to the center of mass of the vehicle, that is, the rear wheelbase; δ is the steering angle of the vehicle, that is, the deflection angle of the front wheel relative to the vehicle's forward direction; This ratio represents the ratio of the rear wheelbase to the total wheelbase of the vehicle (front wheelbase plus rear wheelbase). It reflects the position of the vehicle's center of mass relative to the front and rear axles. r >l f , this ratio is greater than 0.5, indicating that the center of mass is closer to the rear axle; Tan(δ) is the tangent of the steering angle δ, which describes the severity of the front wheel deflection. The larger the steering angle, the larger the value of tan(δ), indicating that the vehicle is turning more sharply. This formula is used to calculate the vehicle's center of mass slip angle in real time based on the vehicle's wheelbase parameters and steering angle. The center of mass slip angle is a crucial parameter in vehicle dynamic analysis and control system design, affecting vehicle stability, handling, and ride comfort. Accurately calculating and monitoring the center of mass slip angle allows for a better understanding and prediction of vehicle behavior, providing a crucial basis for vehicle control and safety.

[0033] The formula for safe sight distance is:

[0034] is the braking distance required for the vehicle to stop completely from the current speed v, μ is the road adhesion coefficient (dynamically corrected, such as integrating meteorological sensor data), g is the acceleration of gravity, cosθ is the curve slope correction factor (slope angle θ affects the effective friction force), v·t reaction The driver or system reaction time t reaction The distance the vehicle continues to travel, t reaction Adaptive adjustment for driver status, L vehicle The length of the vehicle or an additional safety margin (e.g. 1.5 times the vehicle length) to avoid contact with the vehicle ahead / obstacle after braking; Parameters are updated dynamically, and the road adhesion coefficient μ is calculated using the tire-road friction model (e.g., μ = 0.8-0.003·T, where T is the tire temperature in °C). The vehicle speed v is obtained from the CAN bus, and the braking response time t_reaction is adjusted according to the driver's gaze direction (a 0.5s redundancy is added when the gaze deviation is greater than 2s); The curve slope angle θ is calculated using the LiDAR point cloud elevation map (differential elevation resolution ±5 cm).

[0035] S4: Dynamically correct the safe sight distance calculation result based on real-time environmental parameters, generate a driving warning signal and output it to the vehicle control system (ESP); The driving warning signal logic is to trigger a multi-level progressive warning when the measured sight distance value is lower than the safety threshold: Primary warning, showing the safe curve area boundary through the head-up display; Intermediate warning, applying tactile feedback; Advanced warning, linked to vehicle control system to implement active braking The position of traffic participants 200 meters ahead is acquired through V2X communication (DSRC protocol, positioning accuracy ±0.5m), and the μ value is updated by integrating meteorological sensor data (the μ attenuation coefficient is 0.15 when rainfall is greater than 20mm / h). Blockchain-based abnormal data verification: Generate SHA-256 hash values for multi-sensor data, store them in the Hyperledger Fabric distributed ledger, and provide three-level progressive warnings: Level 1 warning (S measured < 1.2S_safe): AR-HUD projects the curve safety zone boundary (green dynamic arc, brightness 300cd / m 2 ); Level 2 warning (S measured < 1.0S_safe): Pulse tactile feedback (frequency 8 Hz, amplitude 0.3 mm) is applied to the steering wheel. Level 3 warning (S measured < 0.8S_safe): The ESP system is linked to implement graded braking (deceleration steps increased to 0.4g).

[0036] Furthermore, the dynamic configuration of the multimodal sensor is as follows: the laser radar adaptively adjusts the scanning density according to the curvature of the curve; the camera device enhances the capture of lane edge features through image optimization algorithm; the sampling frequency of the inertial measurement unit is dynamically matched with the vehicle acceleration; The dynamic configuration method also includes triggering the curve enhancement scanning mode of the laser radar when it is detected that the curvature of the curve exceeds a set threshold, and dynamically adjusting the color space conversion parameters of the camera device according to the ambient light intensity.

[0037] Furthermore, the position information of traffic participants within the set range ahead is obtained through vehicle-road cooperative communication, and the road adhesion coefficient is updated in real time by integrating meteorological sensor data, and the calculation weight of the reaction time is adjusted according to the driver's gaze direction.

[0038] Furthermore, it also includes data verification, cross-modal consistency check of multi-sensor data and elimination of abnormal data. When a single sensor fails, the line-of-sight prediction mode based on historical data is activated.

[0039] Furthermore, the curve recognition also includes using a region growing algorithm to segment the curve area point cloud data, defining seed points based on the lane line curvature characteristics, and reconstructing broken lane lines and eliminating interference features through an interpolation algorithm.

[0040] The road adhesion coefficient μ is dynamically updated through the tire-road friction model, and its calculation formula is:

[0041] 0.8 is a constant term, which represents the reference value of the road adhesion coefficient under standard conditions; −0.003T is a term that is proportional to the tire temperature T, where T is the tire temperature. The data comes from the tire pressure monitoring sensor.

[0042] Example 2: The hard synchronization trigger device in Example 1 is further integrated with a joint spatiotemporal synchronization calibration network (JTSN) to dynamically compensate for the sensor pose offset caused by vehicle bumps through joint optimization of an extended Kalman filter and an online sparse autoencoder. Specifically, construct the sensor state space model:

[0043] in: X t Contains the laser radar rotation angle (±0.02°) and camera focal length parameters (±0.1mm), the state vector X t It is used to represent the state of the system at time t, and covers the set of sensor parameters that need to be dynamically calibrated. For example, the installation pitch angle (θ lidar ), the focal length parameter of the camera (f camera ), IMU and radar coordinate system offset (Δx, Δy, Δz), sensor temperature drift coefficient (α temp ) etc. In this embodiment, these parameters may be dynamically offset due to vehicle bumps, load changes, or ambient temperature fluctuations, and must be calibrated in real time to ensure the accuracy and reliability of the sensor data; The state transfer function f(·) describes the dynamic law of the system state changing over time. Its input includes the state parameter X at the previous moment. t−1 and external control input u t , where u t These typically come from measurements from other sensors, such as the triaxial angular velocity (ωx, ωy, ωz) measured by the IMU, the vehicle speed v provided by a wheel speedometer, and environmental data collected by temperature and humidity sensors. There are two common forms of the state transfer function f(·). One is a nonlinear physical model. For example, using the variation of the lidar pitch angle with vehicle turbulence as an example, its expression is:

[0044] ω y Represents the pitch angular velocity of the IMU, a vertical is the vertical acceleration, and β is the stiffness coefficient of the suspension system. The other is a data-driven model that uses a sparse autoencoder to learn the evolution law of the state, thereby more accurately capturing the state changes in complex environments. Process noise wt: It is used to represent external interference or uncertainty not considered by the model. It is usually assumed that the process noise obeys the Gaussian white noise distribution, that is, w t -N(0,Q t ), where the covariance matrix Q tDynamically adjustable. In this embodiment, process noise comes from a wide range of sources, including random vibrations between vehicle tires and the road surface, slight deviations in multi-sensor clock synchronization, and sudden environmental changes (such as vehicle body vibration caused by strong winds). These noise factors may affect sensor states, so accounting for them in the model helps improve the accuracy of state estimation. Using Extended Kalman Filter (EKF) to achieve X t Specifically, in the prediction stage, the state transfer function f(·) is used to calculate the prior state estimate:

[0045] In the update phase, the multimodal sensor observation data Z is fused t (such as the error between lidar and visual feature matching), calculate the posterior estimate:

[0046] Where h(・) is the observation model, K t This is the Kalman gain. In this way, the sensor parameters can be updated in time, the parameter offset caused by various factors can be compensated, and the accuracy of the sensor data can be ensured. By using FPGA hardware to accelerate the calculation of θ, the calibration network parameters can be updated in real time, effectively compensating for parallax errors caused by differences in the installation positions of the IMU and camera when the vehicle is cornering, thereby improving the overall performance and stability of the system. The goal is to optimize the registration and calibration of sensor data. The objective function is:

[0047] The first is to ensure that the sensor data With the observation value Z i The first item maintains consistency, for example, by reducing errors in the alignment of point clouds and image features. The second item imposes a sparsity constraint (L1 regularization) to filter key calibration parameters and prevent overfitting. In terms of implementation, FPGA hardware-accelerated calculation of θ enables real-time updates of calibration network parameters, effectively compensating for parallax errors caused by differences in the mounting positions of the IMU and camera when the vehicle is cornering, improving overall system performance and stability. When a vehicle is driving on a sharp curve (with a curvature radius less than 300m), the body roll will significantly increase the spatial registration error between the lidar point cloud and the camera image. In this case, the model works as follows: Select the lateral acceleration a measured by IMU in real time y and yaw angular velocity ω zAs state input, these data can reflect the dynamic posture changes of the vehicle during the curve. According to the dynamic characteristics of the vehicle body, the state transfer equation of the radar installation yaw angle ϕlidar is established. That is, the radar installation yaw angle is dynamically adjusted according to the vehicle body posture. k is the correlation coefficient, which is used to characterize the relationship between the lateral acceleration and the yaw angle change. Observation data Z t ,Using the feature matching algorithm, the point cloud - image matching error is extracted as ,observation data, such as the lane line curvature difference, etc. ,These error information intuitively reflects the registration accuracy problem between ,sensor data; The final output is the dynamically corrected radar installation angle ϕ lidar , ensuring that in the enhanced scanning mode of curves (0.1°×0.1° point cloud density), the lidar point cloud and camera image data can be accurately aligned, providing a reliable data foundation for subsequent autonomous driving decision-making and other tasks; The sparse autoencoder contains a 5-layer bottleneck structure (1024-512-256-512-1024), and the sparsity constraint is implemented by KL divergence:

[0048] Among them, ρ=0.05 is a sparse target, is the average activation rate of the hidden layer. The network weights are updated online every 30 seconds through FPGA, using mixed precision training (activation value FP16, gradient FP32); This embodiment uses FPGA hardware acceleration to efficiently complete state prediction and parameter update, meet the 200Hz IMU data fusion requirements, and ensure the system operates stably in scenarios with high real-time requirements such as high-speed driving. The process noise w t It works together with the sparse autoencoder to suppress abnormal disturbances such as sudden pothole impacts, enhance the stability and reliability of the system in complex and harsh environments, and prevent the calibration parameters from deviating from the reasonable range due to interference factors. The model parameter θ supports online learning and can automatically adjust according to the differences in sensor installation on different vehicle platforms (such as trucks, cars, etc.), achieving wide adaptation to various types of vehicles, improving the versatility and practicality of the model, and reducing the development and debugging costs when applying it to different vehicle models.

[0049] Example 3: In the S2 curve recognition process of Example 1 above, an improved region growing algorithm is used. To improve its performance and accuracy, a multi-scale Bessel-Transformer network (MBT-Net) can be optionally used for optimization. This algorithm can divide the curve into differentiable piecewise Bessel curves, and its mathematical expression is:

[0050] In this process, the control point P i It is obtained through real-time fitting of the lidar point cloud, and the fitting error is strictly controlled within the range of less than 0.05 meters to ensure high accuracy of curve fitting. At the same time, in order to accurately capture the curvature changes of complex curves, the proximal strategy optimization algorithm is used to dynamically adjust the distribution of control points. The algorithm aims to minimize the projection error, and its loss function is defined as:

[0051] L ppo The objective function of the proximal policy optimization algorithm is used to measure the difference between the current policy and the old policy. The policy is updated by optimizing this function to improve the performance of the policy. E t is the empirical expectation at time step t, which is an estimate of the expected value of future rewards and is used to measure the expectation of long-term returns under the current strategy; min takes the minimum function, which is used to select the smaller one between two values to limit the amplitude of the policy update; r t (θ) is the policy update ratio, which is defined as the ratio of the probability of the current policy selecting an action at time step t to the probability of the old policy selecting an action at time step t, that is:

[0052] Among them, πθ(a t ∣s t ), is the current strategy in state s t Next select action a t probability; πθ old (a t ∣s t ) is the old policy in state s t Next select action a t probability; clip(r t (θ),1−ϵ,1+ϵ) will update the policy ratio r t (θ) is limited to the range [1−ϵ, 1+ϵ]. This is the key mechanism used in the PPO algorithm to control the amplitude of policy updates. Through the clip operation, it ensures that the policy update is not too large, thereby maintaining the stability of training. ϵ is a hyperparameter that controls the range of policy updates and is usually between 0.1 and 0.3. θ is the parameter of the policy network, which is used to define the policy πθ, that is, the probability distribution of choosing action a under a given state s. In reinforcement learning, the policy network is usually a neural network, whose parameters θ are obtained by optimizing the objective function (such as L ppo) is continuously updated and improved. In curve recognition scenarios, this formula is used to optimize the distribution of control points, enabling sub-pixel modeling of complex curve curvature variations. By minimizing projection errors, it can more accurately fit the geometry of the curve, thereby improving the accuracy and reliability of curve recognition. The multi-scale Bessel-Transformer network contains: Feature extraction layer, PointNet++ downsamples to 1 / 16 resolution; Multi-scale fusion module, 3 sets of dilated convolutions with different dilation rates (rates=2, 4, 8); Bessel control point prediction head, the self-attention mechanism calculates the control point weight loss function and adds curvature continuity constraints:

[0053] L total is the total loss function, L pp0 is the point projection loss, is the second-order derivative of the Bezier curve B(t)B, which represents the rate of change of curvature, and the coefficient 0.5 is the regularization weight.

[0054] Example 4: In the S4 warning of Example 1 above, blockchain verification deploys a dynamic heterogeneous computing engine to improve the overall performance and efficiency of the system. This module uses the Xilinx Zynq UltraScale+ MPSoC platform to achieve mixed-precision acceleration, thereby meeting the high real-time and accuracy requirements of the warning system; At the hardware level, tasks are optimally allocated using a genetic algorithm. The goal is to minimize the overall execution time of the system. The optimization model is as follows: The optimization objective is to minimize:

[0055] Constraints:

[0056] Where T i,cpu represents the execution time of task i on the CPU, T i,fpga represents the execution time of task i on FPGA, x i Is a binary variable, if x i =1 indicates that task i is assigned to CPU, otherwise it is assigned to FPGA, N core Is the number of CPU cores In this way, the model inference delay can be reduced, and the real-time performance of the system can be greatly improved. At the algorithm level, a channel-level dynamic pruning strategy is implemented to further improve system performance. Its pruning rules are defined as:

[0057] W represents the weight matrix of the neural network, l is the number of layers of the network, C l is the number of channels in layer l, W l,i represents the weight of the i-th channel in the l-th layer, I(·) is the indicator function, when |W l,i The value is 1 when |<τ, otherwise it is 0, and τ is the set pruning threshold; The system monitors the FPGA resource usage in real time. When the FPGA resource usage exceeds 85%, the pruning mechanism is automatically activated to prune some channels that have less impact on model accuracy, thereby reducing the amount of calculation and ensuring that the system can always maintain real-time performance and issue early warning information in a timely manner.

[0058] Through the above method, sub-pixel modeling of complex curve curvature changes can be achieved, thereby providing high-precision curve information for subsequent tasks such as vehicle control and path planning, and improving the safety and reliability of the autonomous driving system in curved driving scenarios.

[0059] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for automatically measuring sight distance on a curved highway, characterized in that: The following steps are involved: S1: Collecting three-dimensional spatial data and vehicle motion parameters of the curve area through a multimodal sensor, wherein the multimodal sensor includes a laser radar, a camera device, and an inertial measurement unit; S2: Curve recognition, performing multi-source data fusion on the three-dimensional spatial data and vehicle motion parameters to extract curve curvature, obstacle distribution, and lane line geometry features; S3: Constructing a safe driving sight distance calculation model based on a dynamic vehicle kinematic model, which integrates curve geometry parameters, vehicle real-time speed, and braking response time; S4: Dynamically correct the safe sight distance calculation result according to the real-time environmental parameters, generate a driving warning signal and output it to the vehicle control system.

2. The method for automatically measuring the sight distance on a highway curve according to claim 1, characterized in that: The dynamic configuration of the multimodal sensor includes: the laser radar adaptively adjusting the scanning density according to the curvature of the curve; The camera device uses image optimization algorithms to enhance lane edge feature capture; The sampling frequency of the inertial measurement unit is dynamically matched to the vehicle acceleration.

3. The method for automatically measuring the sight distance on a highway curve according to claim 2, characterized in that: The dynamic configuration method also includes triggering the curve enhancement scanning mode of the laser radar when it is detected that the curvature of the curve exceeds a set threshold, and dynamically adjusting the color space conversion parameters of the camera device according to the ambient light intensity.

4. The method for automatically measuring the sight distance on a curved highway according to claim 1, characterized in that: The multi-source data fusion in S2 is to align the lidar point cloud and visual image data using a spatiotemporal registration algorithm, fuse the inertial measurement data with the lane line geometric features through a filtering algorithm to compensate for measurement errors, establish an obstacle motion trajectory prediction model and dynamically update the threat level.

5. The method for automatically measuring the sight distance on a curved highway according to claim 1, characterized in that: The safety sight distance calculation model formula in S3 is: Among them, v is the real-time speed of the vehicle, μ is the road adhesion coefficient, θ is the curve slope angle, t reaction is the driver's reaction time, L vehicle For the safety margin distance, the functional relationship is determined by the vehicle dynamics constraints.

6. The method for automatically measuring the sight distance on a curved highway according to claim 5, characterized in that: Through vehicle-road cooperative communication, the position information of traffic participants within the set range ahead is obtained, the road adhesion coefficient is updated in real time by integrating meteorological sensor data, and the calculation weight of the reaction time is adjusted according to the driver's gaze direction.

7. The method for automatically measuring sight distance on a highway curve according to claim 1, characterized in that: The driving warning signal logic in S4 is to trigger a multi-level progressive warning when the measured sight distance value is lower than the safety threshold: Primary warning, showing the safe curve area boundary through the head-up display; Intermediate warning, applying tactile feedback; Advanced warning, linked to the vehicle control system to implement active braking.

8. The method for automatically measuring sight distance on a highway curve according to claim 1, characterized in that: It also includes data verification, cross-modal consistency checking of multi-sensor data and elimination of abnormal data. When a single sensor fails, the line-of-sight prediction mode based on historical data is activated.

9. The method for automatically measuring sight distance on a highway curve according to claim 1, characterized in that: The curve recognition also includes using a region growing algorithm to segment the curve area point cloud data, defining seed points based on the lane line curvature characteristics, and reconstructing broken lane lines and eliminating interference features through an interpolation algorithm.

10. The method for automatically measuring sight distance on a curved highway according to claim 5, characterized in that: The road adhesion coefficient μ is dynamically updated through the tire-road friction model, and its calculation formula is: 0.8 is a constant term, which represents the reference value of the road adhesion coefficient under standard conditions; -0.003T is a term that is proportional to the tire temperature T, where T is the tire temperature, and the data comes from the tire pressure monitoring sensor.

Images