A method and device for estimating vehicle roll angle based on point cloud in critical working conditions

Abstract

The application discloses a vehicle roll angle estimation method and device based on point cloud under critical working conditions and belongs to the technical field of intelligent automobile environment perception. Laser radar point cloud data and vehicle kinematics signals are acquired; a region of interest is dynamically adjusted according to the kinematics signals to adaptively focus on a ground plane; ground plane robust estimation is carried out in the dynamic region of interest based on a weighted total least squares method and a Huber loss function to obtain a ground plane normal vector; and the vehicle roll angle is solved according to the geometric relationship between the ground plane normal vector and a reference normal vector. The application solves the point cloud distortion problem caused by the violent movement of the vehicle body under critical working conditions, realizes stable and reliable estimation of the roll angle without increasing hardware, and effectively improves the adaptability of the perception system under extreme working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle environmental perception technology, and in particular to a method, apparatus, device, and storage medium for estimating vehicle roll angle based on point cloud under critical conditions. Background Technology

[0002] In autonomous driving and advanced driver assistance systems, LiDAR is a key sensor for acquiring the three-dimensional geometry of the environment. LiDAR is typically rigidly mounted on the vehicle body, with its coordinate system fixed to the vehicle. When a vehicle performs emergency braking, high-speed obstacle avoidance, or drives over bumpy roads in rain or snow, the vehicle's attitude changes drastically, causing a dynamic deflection of the radar coordinate system. This deflection directly distorts the spatial geometry of the point cloud, resulting in significant deviations in the input data of downstream perception modules such as obstacle detection and localization based on the point cloud. This can lead to system misjudgments and threaten driving safety.

[0003] In existing technologies, attitude estimation methods that directly rely on inertial measurement units are susceptible to sensor noise, temperature drift, and instantaneous errors caused by impact. Traditional point cloud-based attitude estimation methods are mostly designed for smooth driving conditions, and the estimation results are prone to inaccuracy when the vehicle body is tilted, the road surface is uneven, or there are obstacles. At the same time, the point cloud noise caused by violent vehicle movement and extreme environmental factors in critical driving conditions makes the algorithm even more difficult to cope with. Therefore, there is a need for a real-time vehicle roll angle estimation scheme that can adapt to critical driving conditions and has strong anti-interference capabilities. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] To address this issue, this invention discloses a point cloud-based method for estimating vehicle roll angle under critical conditions. The method acquires LiDAR point cloud data and vehicle kinematic signals; dynamically adjusts the extraction range of the region of interest (ROI) in the point cloud based on the vehicle kinematic signals to adaptively focus on the effective region including the ground plane during severe vehicle movement; estimates the ground plane normal vector within the dynamically dynamically adjusted ROI using a weighted robust optimization method; and calculates the vehicle roll angle based on the geometric relationship between the ground plane normal vector and the reference normal vector. This invention solves the point cloud distortion problem caused by severe vehicle movement under critical conditions, achieving stable and reliable roll angle estimation without increasing hardware requirements, and effectively improving the adaptability of the perception system under extreme conditions.

[0006] Another objective of this invention is to provide a point cloud-based vehicle roll angle estimation device for critical operating conditions.

[0007] The third objective of this invention is to provide a computer device.

[0008] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0009] To achieve the above objectives, this invention proposes a method for estimating vehicle roll angle based on point clouds under critical conditions, comprising: Acquire lidar point cloud data and vehicle kinematic signals; The extraction range of the region of interest in the point cloud is dynamically adjusted based on the vehicle's kinematic signals to adaptively focus on the effective region containing the ground plane when the vehicle body is in violent motion. Within the dynamically active region of interest, the ground plane is estimated using a weighted robust optimization method to obtain the ground plane normal vector; The vehicle roll angle is calculated based on the geometric relationship between the estimated ground plane normal vector and the reference normal vector.

[0010] The vehicle roll angle estimation method based on point cloud under critical conditions according to an embodiment of the present invention may also have the following additional technical features: In one embodiment of the present invention, acquiring lidar point cloud data and vehicle kinematic signals includes: Raw point cloud data is collected using a lidar rigidly mounted at the center of the vehicle roof. The vehicle kinematics signals, including vehicle speed, longitudinal acceleration, and lateral acceleration, are acquired in real time via the vehicle CAN bus, and the raw point cloud data and vehicle kinematics signals are transmitted to the domain controller for further processing.

[0011] In one embodiment of the present invention, the step of dynamically adjusting the extraction range of the region of interest in the point cloud based on vehicle kinematic signals to adaptively focus on an effective region including the ground plane during violent vehicle movement includes: The longitudinal range of the region of interest is determined based on vehicle speed to adaptively crop the near and far boundaries of the point cloud in the vehicle's direction of travel; the visual range of the region of interest is determined based on lateral acceleration to adaptively adjust the cropping angle of the point cloud in the horizontal direction; and the height range of the region of interest is determined based on longitudinal acceleration to adaptively compensate for changes in ground point height caused by vehicle pitch and suspension deformation. The spatial range of the dynamically interested region is determined by combining the longitudinal range, the viewing angle range, and the height range.

[0012] In one embodiment of the present invention, estimating the ground plane based on a weighted robust optimization method within a dynamically interested region to obtain the ground plane normal vector includes: Define the equation of the ground plane and the corresponding normal vector expression. Perform principal component analysis based on the covariance matrix of the point cloud in the dynamic region of interest to obtain the initial ground plane normal vector. Based on the initial ground plane normal vector, the Huber loss function is introduced to construct the residual weight coefficients from each point to the ground plane. The weighted full least squares method is used to construct the objective function. The optimized ground plane normal vector is obtained through iterative optimization. The final ground plane equation is determined based on the optimized ground plane normal vector.

[0013] In one embodiment of the present invention, the step of calculating the vehicle roll angle based on the estimated geometric relationship between the ground plane normal vector and the reference normal vector includes: Define the ground plane normal vector when the vehicle is stationary on a horizontal ground and the lidar is installed without tilt as the reference normal vector, and establish an optimized spatial rotation relationship model between the ground plane normal vector and the reference normal vector. Based on the aforementioned spatial rotation relationship model, the angular components of the rotation of the lidar coordinate system around the coordinate axis of the vehicle's forward direction are extracted and used as the vehicle's roll angle.

[0014] In one embodiment of the present invention, after acquiring the lidar point cloud data and vehicle kinematic signals, and before dynamically adjusting the extraction range of the region of interest in the point cloud based on the vehicle kinematic signals, the method further includes: Outlier removal is performed on the lidar point cloud data to eliminate noise points caused by lidar vibration and environmental interference.

[0015] In one embodiment of the present invention, after removing outliers from the lidar point cloud data, the method further includes: Voxel filtering is used to reduce the dimensionality of the point cloud data after outlier removal in order to reduce the amount of point cloud data.

[0016] To achieve the above objectives, another aspect of the present invention proposes a point cloud-based vehicle roll angle estimation device for critical operating conditions, comprising: The data acquisition and input module is used to acquire lidar point cloud data and vehicle kinematic signals; The dynamic region of interest extraction module is used to dynamically adjust the extraction range of the point cloud region of interest based on the vehicle's kinematic signals, so as to adaptively focus on the effective area containing the ground plane when the vehicle body is in violent motion. The robust ground plane estimation module is used to estimate the ground plane based on a weighted robust optimization method within a dynamically dynamic region of interest, so as to obtain the ground plane normal vector; The roll angle calculation module is used to calculate the vehicle roll angle based on the geometric relationship between the estimated ground plane normal vector and the reference normal vector.

[0017] This invention discloses a point cloud-based method and apparatus for estimating vehicle roll angle under critical conditions. By acquiring lidar point cloud data and vehicle kinematic signals, the method dynamically adjusts the region of interest (ROI) based on the kinematic signals to adaptively focus on the ground plane. A weighted robust optimization method is then used to estimate the ground plane normal vector, thereby calculating the vehicle roll angle. This effectively solves the problems of point cloud distortion caused by severe vehicle movement and inaccurate estimation using traditional methods in existing technologies. It achieves stable and reliable roll angle estimation under extreme conditions, significantly improving the adaptability and safety of intelligent driving perception systems. Furthermore, it requires no additional hardware, enhancing the method's engineering applicability.

[0018] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing a point cloud-based vehicle roll angle estimation method under critical conditions as described in the first aspect embodiment.

[0019] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a point cloud-based method for estimating vehicle roll angle under critical conditions as described in the first aspect embodiment.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for estimating vehicle roll angle based on point cloud under critical conditions according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall process of another method for estimating vehicle roll angle based on point cloud under critical conditions according to an embodiment of the present invention; Figure 3 This is a schematic diagram of a point cloud-based vehicle roll angle estimation device under critical conditions according to an embodiment of the present invention. Figure 4 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] The following description, with reference to the accompanying drawings, describes a method, apparatus, device, and storage medium for estimating vehicle roll angle based on point cloud under critical operating conditions, according to an embodiment of the present invention.

[0025] The core idea of ​​this invention is to construct a roll angle estimation framework based on the fusion of point cloud and vehicle kinematic signals. This framework acquires lidar point cloud data and vehicle kinematic signals and performs preprocessing to remove outlier interference and reduce data dimensionality. It dynamically adjusts the longitudinal range, viewing angle range, and height range of the region of interest based on vehicle speed, longitudinal acceleration, and lateral acceleration. During periods of intense vehicle movement, it adaptively focuses on the effective region including the ground plane, thereby suppressing environmental noise and redundant information under dynamic conditions. Furthermore, a weighted robust optimization method is introduced within the dynamically defined region of interest. Principal component analysis is used to obtain the initial ground plane normal vector, and based on… The Huber loss function is used to construct residual weight coefficients, and the weighted full least squares method is used to iteratively optimize the ground plane equation, achieving stable estimation of ground plane parameters under complex road conditions and vehicle tilt conditions. Finally, the vehicle roll angle is calculated based on the spatial geometric relationship between the ground plane normal vector and the reference normal vector. This transforms the traditional attitude estimation method that relies on a single sensor or static assumptions into an intelligent perception system that can fully exploit point cloud geometric features, integrate vehicle dynamic characteristics, and achieve multi-source information collaborative optimization. This significantly improves the accuracy, robustness, and real-time performance of roll angle estimation under extreme conditions, effectively enhancing the adaptability of the intelligent driving perception system in critical situations.

[0026] Example 1 To achieve the above invention, embodiments of the present invention provide a method for estimating vehicle roll angle based on point clouds under critical conditions, such as... Figure 1 As shown, it includes: S1 acquires lidar point cloud data and vehicle kinematic signals.

[0027] Specifically, the lidar emits a laser beam and receives the reflected signal from the target. Based on the time-of-flight, it calculates the distance between the target and the sensor, and combines this with the scanning angle to generate high-precision 3D point cloud data to characterize the geometry of the vehicle's surrounding environment. Vehicle kinematic signals are acquired in real-time via the CAN bus. This bus aggregates and broadcasts sensor data (such as wheel speed sensor and acceleration sensor signals) generated by various electronic control units (ECUs) in the vehicle. The domain controller listens for and parses relevant messages through the CAN interface to obtain dynamic parameters including vehicle speed, longitudinal acceleration, and lateral acceleration. The acquisition of both types of data is synchronized in time, providing spatiotemporally aligned input for subsequent fusion processing.

[0028] Specifically, in actual deployment, the LiDAR is rigidly mounted at the geometric center of the vehicle's roof, ensuring its coordinate system is fixedly connected to the vehicle's coordinate system, and the mounting plane remains horizontal and tilt-free to eliminate static installation errors. The LiDAR continuously scans at a sampling frequency of 10Hz, transmitting the raw point cloud data to the vehicle's domain controller in real time via a gigabit Ethernet interface. Simultaneously, the domain controller connects to the vehicle's CAN network via a CAN transceiver, acquiring signals such as vehicle speed, longitudinal acceleration, and lateral acceleration at a sampling rate of no less than 100Hz, and using hardware timestamps or software synchronization mechanisms to achieve time alignment between the point cloud data and kinematic signals. The domain controller incorporates a high-performance computing unit responsible for subsequent data preprocessing and algorithm execution.

[0029] Furthermore, to achieve accurate roll angle estimation, the lidar parameters must meet specific requirements: a horizontal field of view of 360° to ensure blind spots, a vertical field of view of 26° to cover the near ground and distant obstacles, a ranging accuracy better than ±2cm, and an angular resolution of no less than 0.2°, so as to clearly capture the details of the ground point cloud even in critical conditions. The accuracy of the vehicle kinematics signals directly affects the adjustment effect of the dynamic region of interest. The accuracy of the vehicle speed signal needs to reach 0.1m / s, and the accuracy of the longitudinal and lateral acceleration signals needs to be better than 0.05m / s² to ensure a sensitive response to changes in vehicle attitude. The acquisition and transmission delay of all signals should be controlled within 50ms to meet real-time requirements.

[0030] Specifically, this step is particularly suitable for the perception needs of intelligent driving systems in critical situations, such as emergency braking, high-speed obstacle avoidance, and driving on slippery roads. In these scenarios, the vehicle's attitude changes drastically, and traditional methods based on the assumption of fixed sensor installation are prone to failure. This step, however, provides a reliable data source for subsequent dynamic region of interest extraction and robust ground plane estimation by acquiring high-frequency, high-precision point cloud and kinematic signals in real time. Even if the vehicle tilts or pitches, the LiDAR can continue to scan the surrounding environment, and the CAN signal can reflect the vehicle's motion status in real time, thus ensuring the algorithm's adaptability under extreme conditions.

[0031] Specifically, point cloud data fully preserves the three-dimensional structure of the ground and obstacles, while kinematic signals provide real-time information on the vehicle's dynamic response. The fusion of these two data points enables the algorithm to adaptively adjust the processing area under critical conditions, effectively suppressing noise interference and improving the stability of ground plane estimation. Ultimately, this data acquisition method fully leverages the potential of existing onboard sensors without adding extra hardware, achieving accurate and robust estimation of the vehicle's roll angle and strongly supporting the safety decisions of the intelligent driving system.

[0032] Furthermore, S1 includes: The S11 uses a lidar rigidly mounted at the center of the roof to collect raw point cloud data.

[0033] Specifically, the lidar emits pulsed laser beams into the surrounding environment through a high-speed rotating transmitting unit and receives the echo signals reflected by the target. Based on the linear relationship between the speed of light and time of flight, it accurately calculates the spatial distance of each reflection point. Combined with the current scanning angle, it generates a three-dimensional point cloud coordinate system centered on the sensor. Rigidly mounting the lidar to the center of the vehicle roof establishes a fixed spatial transformation relationship between its coordinate system and the vehicle's coordinate system. This ensures that the point cloud data accurately reflects the geometry of the vehicle's surrounding environment and eliminates measurement errors caused by loose mounting or positional misalignment, providing a stable reference for subsequent attitude calculations.

[0034] Specifically, in actual deployment, the LiDAR is fixed to the geometric center of the vehicle's roof using a dedicated bracket. The bracket design ensures that the sensor base plane is parallel to the vehicle's design reference plane and is tilt-free in all directions. After installation, it is calibrated using a level or calibration tool. The LiDAR continuously scans at a preset frequency (e.g., 10Hz), generating raw point cloud data frames through its internal processor. Each frame contains tens of thousands to hundreds of thousands of three-dimensional spatial points. The data is output in real time via a gigabit Ethernet interface and transmitted to the high-performance computing unit of the domain controller via the vehicle network. To ensure reliable and low-latency data transmission, the physical layer uses automotive-grade wiring harnesses and is configured with a dedicated data receiving buffer to prevent data packet loss due to network congestion.

[0035] Furthermore, to achieve high-precision perception in critical situations, the parameters of the LiDAR must meet stringent requirements. A horizontal field of view of 360° ensures blind-spot-free coverage in all directions around the vehicle; a vertical field of view of no less than 26° to accommodate both near-field ground and distant obstacles; a ranging accuracy better than ±2cm to ensure the geometric accuracy of the ground point cloud; and an angular resolution of no less than 0.2° to clearly distinguish subtle ground undulations. In addition, the LiDAR's vibration resistance must meet automotive-grade standards (such as ISO 16750) to ensure stable operation even during periods of intense vehicle movement; and the data output rate (10Hz) must match the update frequency of the vehicle's kinematic signals to guarantee accurate time synchronization.

[0036] Specifically, this step applies to the environmental perception needs of intelligent driving systems under various extreme conditions, such as emergency obstacle avoidance, high-speed cornering, and driving on bumpy roads. In these scenarios, the vehicle's attitude changes dynamically, but the rigid mounting of the LiDAR ensures that the point cloud data is always based on the vehicle's coordinate system. This allows subsequent algorithms to infer the vehicle's attitude by analyzing changes in ground geometry within the point cloud. Simultaneously, the center position of the roof minimizes the obstruction of the scanning field of view by the vehicle's structure, ensuring the integrity of the point cloud over the ground area and providing ample data support for roll angle estimation.

[0037] Specifically, rigid mounting eliminates the relative motion between the sensor and the vehicle body, ensuring that geometric distortions in the point cloud are caused solely by changes in vehicle attitude. This allows for accurate calculation of the roll angle by analyzing changes in the ground normal vector. This data acquisition method fully leverages the potential of onboard LiDAR without adding extra sensor hardware, providing a stable and accurate input source for vehicle attitude perception under extreme conditions.

[0038] S12 acquires vehicle kinematic signals, including vehicle speed, longitudinal acceleration, and lateral acceleration, in real time via the vehicle CAN bus, and transmits the raw point cloud data and vehicle kinematic signals to the domain controller for further processing.

[0039] Specifically, the vehicle controller local area network bus (VLAN), as the standard communication protocol for data exchange between on-board electronic control units, employs differential signal transmission and a multi-master node arbitration mechanism to broadcast real-time data generated by sensors distributed throughout the vehicle (such as wheel speed sensors and inertial measurement units) to the bus. The domain controller monitors network traffic through the VLAN interface, parses signals such as vehicle speed, longitudinal acceleration, and lateral acceleration based on preset message identifiers, and performs time alignment and joint encapsulation of these kinematic signals with synchronously acquired LiDAR point cloud data, providing a unified data input for subsequent multimodal information fusion processing.

[0040] Specifically, in actual deployment, the domain controller has a built-in transceiver conforming to the Controller Area Network (CAN) bus physical layer standard, connected to the vehicle's main CAN bus via twisted-pair cables. The domain controller runs a real-time operating system, where the CAN bus driver receives bus messages via interrupts or polling. Based on the parsing rules defined in the database file (DBC file), it extracts physical quantities such as vehicle speed (typically from the anti-lock braking system or electronic stability program controller), longitudinal acceleration, and lateral acceleration (typically from the airbag controller or dedicated inertial measurement unit) from the raw messages. Simultaneously, LiDAR point cloud data is received via an Ethernet interface. The domain controller uses a high-precision system clock or GPS pulse-per-second signal to add hardware timestamps to both types of data, ensuring time alignment accuracy within milliseconds. Subsequently, point cloud frames and kinematic signals from nearby points at the same time are combined into data units and stored in shared memory for subsequent algorithm modules to read.

[0041] Furthermore, to ensure the real-time performance and accuracy of roll angle estimation under critical conditions, the acquisition of vehicle kinematic signals must meet stringent performance specifications. The update frequency of vehicle speed signals is typically no less than 100Hz, with a measurement accuracy better than 0.1m / s, to accurately reflect the instantaneous longitudinal motion of the vehicle. Similarly, the update frequency of longitudinal and lateral acceleration signals is also no less than 100Hz, with a measurement accuracy better than 0.05m / s², to ensure sensitive capture of dynamic changes in vehicle attitude. The latency of controller area network bus communication should be controlled within 20ms to ensure data freshness. The time synchronization error between point cloud data and kinematic signals must be less than 50ms to avoid ground misestimation due to data misalignment. In addition, the domain controller must have sufficient data throughput capacity to simultaneously process high-speed point cloud data streams (up to hundreds of megabytes per second) and multiple controller area network bus signals.

[0042] Specifically, this step is particularly applicable to the real-time perception needs of intelligent driving systems under dynamic extreme conditions, such as emergency braking, high-speed lane changes, and driving over bumpy roads. In these scenarios, the vehicle's longitudinal and lateral dynamics change drastically, and static point cloud data alone cannot accurately distinguish between vehicle motion and environmental structure. By synchronously acquiring vehicle kinematic signals, the domain controller can grasp the vehicle's motion state in real time, providing vehicle speed to determine the scanning distance, lateral acceleration to adjust the scanning angle, and longitudinal acceleration to compensate for suspension deformation for subsequent dynamic region of interest extraction. This allows the algorithm to accurately focus on the ground area even when the vehicle's attitude changes drastically, ensuring the stability of ground plane estimation.

[0043] Specifically, the introduction of controller area network bus data enables the algorithm to perceive the vehicle's motion status in real time, providing key control variables for dynamically adjusting the region of interest and significantly improving the accuracy of downstream robust ground plane estimation. Simultaneously, unified processing of both types of data within the domain controller avoids the data transmission delays and asynchrony issues common in multi-sensor systems, effectively enhancing the overall algorithm's real-time response capability and reliability under critical conditions, laying a solid foundation for subsequent accurate roll angle calculation.

[0044] S2 dynamically adjusts the extraction range of the region of interest in the point cloud based on the vehicle's kinematic signals, so as to adaptively focus on the effective region containing the ground plane when the vehicle body is in violent motion.

[0045] Specifically, in critical situations, drastic changes in vehicle attitude cause dynamic deflection of the lidar coordinate system, resulting in a shift in the projection position of ground points originally located within a fixed spatial range in the lidar coordinate system. Using a fixed region of interest (ROI) would not be able to stably capture a complete ground point cloud. This step utilizes kinematic signals such as vehicle speed, longitudinal acceleration, and lateral acceleration to establish a vehicle dynamic response model. By quantitatively analyzing the impact of vehicle pitch, roll, and longitudinal displacement on the spatial distribution of ground points, the boundaries of the ROI in the longitudinal, lateral, and height directions are adaptively adjusted to ensure that the extraction range always coincides with the projection area of ​​the real ground plane in the lidar coordinate system, thereby achieving precise focusing of the ground plane in complex dynamic environments.

[0046] Specifically, the domain controller first parses the vehicle kinematic signals acquired in step S1. Based on the current vehicle speed, it calculates the furthest longitudinal distance of the region of interest (ROI). The higher the vehicle speed, the farther the longitudinal extension, ensuring sufficient coverage of the road ahead during high-speed driving. It determines the horizontal viewing angle range of the ROI based on lateral acceleration. When the vehicle experiences lateral acceleration, the viewing angle shifts towards the inside of the curve to compensate for point cloud shift caused by vehicle roll. Based on longitudinal acceleration combined with vehicle mass, suspension stiffness, and wheelbase parameters, it calculates the change in center of gravity height caused by vehicle pitch, thereby determining the height range of the ROI to compensate for the impact of the vehicle's nose tilting up or down on the height of ground points during emergency braking or acceleration. The adjustment thresholds for these three dimensions are calculated and updated in real time, collectively forming a three-dimensional dynamic spatial mask. This mask iterates through and filters the original point cloud, retaining only point cloud data that simultaneously satisfies the longitudinal, lateral, and height constraints as input for subsequent processing.

[0047] Furthermore, the calculation coefficients for the furthest longitudinal distance need to be calibrated based on the effective ranging range of the lidar and the vehicle's braking performance to ensure that effective surface points are completely included. The dynamic adjustment range of the horizontal viewing angle should be controlled within the horizontal field of view of the lidar, and the adjustment step size should match the resolution of the lateral acceleration, typically not exceeding 1 degree. The calculation accuracy of the height compensation amount needs to reach the centimeter level to ensure that ground points are not mistakenly excluded. The complete process of region of interest extraction should be completed within a single frame point cloud processing cycle (e.g., within 100 milliseconds) to ensure the real-time performance of the algorithm. In addition, the adjustments to the longitudinal range, viewing angle range, and height range should have smooth transition characteristics to avoid severe jittering of the extraction boundary due to signal noise.

[0048] Specifically, this step is particularly applicable to the environmental perception needs of intelligent driving systems under dynamic extreme conditions, such as scenarios involving emergency braking, high-speed cornering, emergency obstacle avoidance, or driving on bumpy roads. In these scenarios, instantaneous changes in vehicle attitude can cause the traditional fixed region of interest to partially or completely detach from the actual ground area, leading to the failure of ground plane estimation. This step dynamically adjusts the region of interest by fusing vehicle kinematic signals, ensuring that regardless of the vehicle's tilt, pitch, or displacement, the algorithm always focuses on the real ground point cloud. This provides a stable and clean data source for subsequent robust ground plane estimation, thereby ensuring the continuous and reliable operation of the perception system under critical conditions.

[0049] Specifically, compared to the fixed-region method, this step significantly reduces interference from non-ground points (such as vehicles, buildings, and trees), improving the input signal-to-noise ratio of ground plane estimation. Simultaneously, by eliminating a large number of irrelevant point clouds, the amount of data for subsequent processing is effectively reduced, improving the algorithm's computational efficiency. More importantly, this dynamic adjustment mechanism makes the ground plane estimation algorithm naturally adaptable to drastic vehicle movements, ensuring complete capture of ground point clouds under dynamic conditions such as roll and pitch, laying a solid foundation for subsequent high-precision and robust ground plane normal vector calculation.

[0050] Furthermore, S2 includes: S21. Determine the longitudinal range of the region of interest based on the vehicle speed to adaptively crop the near and far boundaries of the point cloud in the vehicle's driving direction; determine the viewing angle range of the region of interest based on the lateral acceleration to adaptively adjust the cropping angle of the point cloud in the horizontal direction; determine the height range of the region of interest based on the longitudinal acceleration to adaptively compensate for changes in ground point height caused by vehicle pitch and suspension deformation.

[0051] Specifically, vehicle speed determines the distance the vehicle displaces per unit time. According to the uniformly accelerated motion model, the higher the vehicle speed, the farther the projection position of the ground point cloud in the radar coordinate system. Therefore, the longitudinal range needs to be extended accordingly to ensure effective coverage. Lateral acceleration characterizes the degree of body roll when the vehicle is turning. According to the relationship between centrifugal force and body roll angle, the greater the lateral acceleration, the more significant the body roll, causing the horizontal distribution of the ground point cloud to deflect. Therefore, the viewing angle needs to be adjusted to compensate for the deflection effect. Longitudinal acceleration reflects the vehicle's pitch dynamics. According to the pitch moment balance in vehicle dynamics, longitudinal acceleration causes suspension compression or extension, resulting in changes in the installation height of the lidar. Therefore, the height range needs to be dynamically adjusted to match the actual vertical distribution of ground points.

[0052] Specifically, the longitudinal range is determined based on the product of the vehicle speed signal and a preset speed coefficient. The calculation formula is: the furthest distance equals the nearest distance plus the vehicle speed multiplied by the speed coefficient, where the speed coefficient is calibrated based on the effective range of the LiDAR and the road curvature. The viewing angle range is determined based on the lateral acceleration signal and the lateral acceleration coefficient. The tilt direction is determined by a sign function, and half of the LiDAR's horizontal field of view is multiplied by the lateral acceleration coefficient to obtain the deflection angle of the current frame, thus achieving adaptive offset of the horizontal truncation boundary. The height range is determined based on the longitudinal acceleration signal and the vehicle's inherent parameters. The pitch displacement at the center of gravity is calculated by multiplying the longitudinal acceleration by the vehicle's mass and center of gravity height, and dividing by the suspension stiffness and wheelbase, thus obtaining the change in ground point height. Finally, with the LiDAR installation height as the center, this change and a preset tolerance are combined to form a dynamic height range. The boundaries of these three dimensions are updated in real time, jointly constructing a three-dimensional dynamic mask to spatially filter the original point cloud.

[0053] Furthermore, in the calculation of the longitudinal range, the velocity coefficient is typically between 0.1 and 0.5 seconds, ensuring that the farthest distance covers an effective ground area of ​​30 to 80 meters ahead. The adjustment step size of the viewing angle range needs to match the resolution of the lateral acceleration. After the lateral acceleration coefficient is calibrated, the viewing angle adjustment accuracy should be within 1 degree, and the maximum deflection angle should not exceed one-quarter of the horizontal field of view of the lidar to avoid excessive truncation. The calculation accuracy of the compensation amount in the height range needs to reach the centimeter level. The estimation error of the centroid height change caused by longitudinal acceleration should be less than 2 centimeters, and the height tolerance is usually set to 10 to 20 centimeters to accommodate minor road surface undulations. The adjustment algorithms for all three dimensions need to be calculated and updated within a single frame point cloud processing cycle (e.g., 100 milliseconds).

[0054] Specifically, in high-speed straight-line driving scenarios, the extended longitudinal range ensures that distant ground points are captured in advance, providing data support for pre-aiming control. In sharp turns or emergency obstacle avoidance scenarios, the dynamic offset of the field of view compensates for the point cloud rotation caused by vehicle roll, ensuring complete coverage of the ground inside the curve. In emergency braking or rapid acceleration scenarios, adaptive compensation of the height range eliminates the vertical deviation caused by vehicle pitch, preventing ground points from being mistakenly removed due to height changes. Through these multi-dimensional coordinated adjustments, the algorithm can maintain precise focus on the ground area under various critical conditions.

[0055] Specifically, the dynamic extension of the longitudinal range effectively balances the integrity and data volume of ground points at both near and far distances, avoiding computational waste for invalid distant points. The adjustment of the viewpoint range significantly reduces the false inclusion rate of non-ground points (such as roadside guardrails and trees) under tilt conditions, improving the purity of the ground point cloud. Dynamic compensation within the height range successfully solves the problem of vertical drift of ground points caused by suspension deformation, ensuring that ground points always fall within the selection range. The synergistic effect of these three dimensions significantly improves the input signal-to-noise ratio of subsequent ground plane estimation algorithms, laying a reliable data foundation for high-precision and highly robust ground plane normal vector calculation.

[0056] S22, combining the longitudinal range, the viewing angle range, and the height range, the spatial range of the dynamic region of interest is determined.

[0057] Specifically, the construction of a dynamic region of interest essentially involves defining a three-dimensional subspace within the LiDAR coordinate system. This subspace is constrained by constraints in three orthogonal dimensions: longitudinal, lateral (viewpoint), and vertical (height). The longitudinal range defines the near and far boundaries along the vehicle's direction of travel; the viewpoint range delineates the left and right scanning intervals in the horizontal plane using polar angles; and the height range defines the upper and lower boundaries in the vertical direction. These three dimensional constraints are independent yet coupled, and their intersection forms an irregular hexahedral or prism-shaped spatial region. Any point cloud located within this intersection simultaneously satisfies the dynamic adjustment requirements of all three dimensions, thus ensuring that the extracted region geometrically and completely covers the actual ground area affected by vehicle motion.

[0058] Specifically, the domain controller first obtains the longitudinal nearest and farthest distances, the left and right boundary angles of the viewing angle range (with the area directly in front of the LiDAR as the zero-degree reference, left being positive and right being negative), and the lower and upper limits of the height range, based on the calculation results of step S21. Then, it iterates through each 3D point in the original point cloud: first, it checks whether the point's x-axis coordinate (vehicle's forward direction) is within the closed interval of the longitudinal range; if so, it calculates the angle between the point's projection on the horizontal plane and the positive x-axis, and determines whether this angle falls between the left and right boundaries of the viewing angle range; if still satisfied, it checks whether the point's z-axis coordinate (vertical direction) is within the closed interval of the height range. Only when all three conditions are met is the point of interest marked and retained in the output point cloud set; if any one condition is not met, it is discarded. After the traversal is complete, the output point cloud is the effective point set within the dynamically defined region of interest.

[0059] Furthermore, to achieve accurate joint determination of three-dimensional constraints, the boundary conditions of each dimension must meet strict quantification and logical consistency requirements. The nearest point in the longitudinal range is usually taken as the edge of the near-field blind zone of the lidar, for example, 0.5 meters to 1 meter, while the farthest point is limited by the effective ranging and velocity coefficient calculation results of the lidar, generally not exceeding 80 meters. The left and right boundaries of the viewing range should maintain symmetry, with the left boundary angle not exceeding -30 degrees and the right boundary not exceeding +30 degrees after adjustment, to avoid excessive truncation that could lead to the loss of ground points. The lower and upper limits of the height range must completely encompass the road surface undulations, typically centered on the lidar installation height, extending 20 to 30 centimeters above and below. The determination conditions for the three dimensions must be applied synchronously at the same timestamp to ensure spatial consistency, and the computational complexity of the entire screening process should be controlled within one million points per second to meet real-time requirements.

[0060] Specifically, in complex scenarios involving simultaneous longitudinal acceleration / deceleration, lateral steering, and vertical pitch, such as emergency braking and obstacle avoidance maneuvers, a single dimensional adjustment cannot fully capture the ground area. The longitudinal range ensures that ground points in the area of ​​emergency braking ahead are included; the viewing angle range compensates for roll deviation caused by steering; and the height range corrects for height changes caused by braking pitch. These three factors work together to enable the algorithm to extract a complete and clean ground point cloud even under complex dynamic disturbances. This mechanism provides high-quality input for subsequent robust ground plane estimation, significantly improving the algorithm's adaptability under extreme conditions.

[0061] Specifically, by jointly constraining the longitudinal range, viewing angle range, and height range, the spatial range of the dynamically defined region of interest achieves a precise three-dimensional wrapping of ground points. Compared to single-dimensional adjustment, multi-dimensional joint determination effectively eliminates spatial gaps or overlapping redundancies that may arise from independent adjustments of each dimension, ensuring the integrity of the ground point cloud. Simultaneously, through collaborative filtering across the three dimensions, a large number of non-ground points (such as overpasses, roadside trees, and distant vehicles) are more effectively removed, further improving the purity and signal-to-noise ratio of the point cloud. This step, acting as a bridge connecting dynamic parameter adjustment and subsequent geometric estimation, lays a solid data foundation for the calculation of high-precision ground plane normal vectors and is a crucial link in ensuring the stable and reliable operation of the entire tilt angle estimation method under critical conditions.

[0062] S3. Within the dynamically interested region, the ground plane is estimated using a weighted robust optimization method to obtain the ground plane normal vector.

[0063] Specifically, within the dynamically defined region of interest (GROUP), the ground can be approximated as a spatial plane, mathematically expressed as a three-dimensional plane equation. The normal vector, perpendicular to the ground and pointing upwards, is the key geometric parameter of this equation. The essence of ground plane estimation is to find the optimal plane parameters from point cloud data containing noise and outliers, minimizing the weighted sum of squared distances from all interior points (real ground points) to the plane. Traditional least squares methods are susceptible to outlier interference. This step introduces a weighted robust optimization approach, iteratively updating the weight coefficients of each point to reduce the influence of non-ground points (such as curbs, vehicle undercarriages, and low obstacles) on plane fitting. This allows for the stable and accurate extraction of the ground plane normal vector within the dynamically defined GROUP, providing a crucial geometric reference for subsequent roll angle calculations.

[0064] Specifically, the domain controller first takes the point cloud within the dynamically defined region of interest as input. The ground plane equation is defined as a normal vector, meaning the signed distance from all points satisfying the plane equation to the origin is constant. The first step involves calculating the covariance matrix of the point set and performing eigenvalue decomposition. The eigenvector corresponding to the smallest eigenvalue is used as the initial ground plane normal vector. This step, based on principal component analysis, quickly obtains a rough estimate. The second step calculates the residual distance from each point to the plane based on the current normal vector. A robust weighting coefficient is constructed using the Huber loss function. This function assigns quadratic weights when the residual is less than a threshold and linear weights when it is greater than the threshold, effectively suppressing the influence of outliers. The third step uses weighted full least squares to construct a weighted covariance matrix and performs eigenvalue decomposition again. The eigenvector corresponding to the smallest eigenvalue is the optimized ground plane normal vector. This process can be iterated until the change in the normal vector is less than a preset threshold, at which point the final ground plane normal vector is output.

[0065] Furthermore, in obtaining the initial normal vector, the calculation of the covariance matrix needs to be based on a sufficient number of point clouds, typically no less than 500 point clouds within the dynamically relevant region of interest, to ensure statistical significance. The threshold parameter of the Huber loss function needs to be calibrated according to the dynamic characteristics of the vehicle suspension, with a value typically ranging from 5 cm to 15 cm to distinguish between minor undulations in the real ground and non-ground obstacles. The number of iterations of the weighted full least squares method generally does not exceed 5, and the convergence threshold for the change in normal vector in each iteration is set to within 0.001 radians to ensure the stability of the estimation results. The computation time of the entire ground plane estimation process should be controlled within 50 milliseconds to meet real-time requirements. The accuracy of the final output ground plane normal vector needs to achieve an angle error of less than 0.5 degrees to ensure the accuracy of subsequent roll angle calculations.

[0066] Specifically, in emergency obstacle avoidance scenarios, the dynamically generated region of interest may contain point clouds of the underside of a vehicle that has come to an abrupt stop or debris from the road shoulder. Using ordinary least squares for these non-ground points would severely distort the ground plane estimation results. The weighted robust optimization method, by dynamically reducing their weights, effectively suppresses this interference. In wet or snow-covered scenarios, the ground point cloud is sparse and contains significant noise. The introduction of the Huber loss function ensures that the ground plane estimation remains stable even when data quality deteriorates. When driving on bumpy roads, minor undulations are reasonably incorporated, preventing road surface fluctuations from being misinterpreted as vehicle roll, thus providing a clean ground geometric reference for roll angle calculation.

[0067] Specifically, compared to traditional least-squares fitting methods, this step significantly reduces the interference of non-ground points (such as vehicles, obstacles, and curbs) on plane fitting, improving the accuracy and stability of ground plane estimation. The introduction of the Huber loss function enables the algorithm to effectively eliminate outliers while accommodating reasonable ground undulations, achieving a good balance between accuracy and robustness. As the core link connecting dynamic region of interest extraction and roll angle calculation, this step provides reliable and accurate input for subsequent attitude calculation based on the geometric relationship of the normal vector space, and is a key technical support for ensuring the performance of the entire roll angle estimation method under critical conditions.

[0068] Furthermore, S3 includes: S31, define the ground plane equation and the corresponding normal vector expression, perform principal component analysis based on the covariance matrix of the point cloud in the dynamically interested region, and obtain the initial ground plane normal vector.

[0069] Specifically, the ground plane in three-dimensional space can be represented as a set of points satisfying specific geometric constraints. Its mathematical form is a plane equation, meaning that the inner product of the normal vector and the point's coordinates for all points on the plane is a constant. The normal vector, as a direction vector perpendicular to the plane, is a core geometric parameter describing the plane's orientation. Principal component analysis (PCA) states that the eigenvectors of the covariance matrix of a 3D point cloud correspond to the directions with the largest data variance, and the eigenvector corresponding to the smallest eigenvalue is the direction in which the point cloud distribution is most concentrated. For ground point clouds, the variance in the normal direction is the smallest. Therefore, by analyzing the covariance matrix of the point cloud within a dynamically defined region of interest and extracting the eigenvector corresponding to the smallest eigenvalue, an initial estimate of the ground plane normal vector can be quickly obtained.

[0070] Specifically, the domain controller first defines the mathematical expression of the ground plane equation. Let the ground plane normal vector be a three-dimensional unit vector, and the plane equation be that the inner product of this vector and the coordinates of a spatial point is equal to a constant. Then, using the point cloud set within the dynamic region of interest as input, the mean of the three-dimensional coordinates of all points is calculated to obtain the center point of the point cloud. Based on this, a covariance matrix is ​​constructed, where each element is the second moment of the coordinate dimension relative to the mean. The covariance matrix is ​​constructed using an unbiased estimation form, i.e., divided by the sample size minus one. After calculating the covariance matrix, numerical methods such as Jacobi iteration or QR decomposition are used to decompose its eigenvalues, obtaining three eigenvalues ​​and their corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue is selected as the initial estimate of the ground plane normal vector. This initial normal vector, along with the center point of the point cloud, constitutes the initial ground plane equation.

[0071] Furthermore, to ensure the accuracy and numerical stability of the initial normal vector estimation, the construction of the covariance matrix must be based on a sufficient number of point cloud samples. The number of effective point clouds within the dynamically relevant region of interest should ideally be no less than 300 to guarantee statistical significance. The computational precision of the covariance matrix must meet single-precision floating-point standards, and the numerical error of eigenvalue decomposition should be controlled within 1e-6 to avoid deviations in the normal vector direction due to numerical instability. The precision of the initial normal vector must meet the requirement of an angle error of less than 2 degrees to provide a good iterative starting point for subsequent robust optimization. The time cost of the entire computation process should be controlled within 20 milliseconds to ensure it does not become a performance bottleneck for the overall algorithm. In addition, the construction of the covariance matrix must consider the uniformity of the units of the point cloud coordinates to avoid numerical distortion introduced by inconsistent units.

[0072] Specifically, in scenarios where the vehicle travels in a straight line and the road surface is flat, the point cloud distribution within the dynamically defined region of interest (GROUP) is relatively ideal, and Principal Component Analysis (PCA) can directly obtain normal vectors close to the true values. In scenarios where the vehicle is slightly tilted or the road surface has a small slope, PCA can still effectively capture the main direction of the point cloud distribution, providing a reasonable initial estimate. Even if a small number of non-ground points are mixed into the GROUP, their impact on the covariance matrix is ​​limited because the number of non-ground points is usually much smaller than that of ground points, and the initial normal vector can still maintain the correct overall direction. This fast initial estimate reduces the search space for subsequent weighted robust optimization, significantly improving the overall convergence speed and stability of the algorithm.

[0073] Specifically, compared to direct iterative optimization, this step obtains preliminary solutions for the ground plane parameters at a lower computational cost, providing good initial values ​​for subsequent refined estimation. The mathematical rigor of the covariance matrix eigenvalue decomposition ensures the theoretical correctness of the initial normal vector direction. As the starting point of the entire robust ground plane estimation process, this step effectively narrows the optimization search space of the subsequent weighted full least squares method, reduces the number of iterations, lowers the overall computational burden, and lays a solid foundation for maintaining the algorithm's real-time performance and convergence in complex dynamic environments.

[0074] S32, based on the initial ground plane normal vector, introduces the Huber loss function to construct the residual weight coefficients from each point to the ground plane, uses the weighted full least squares method to construct the objective function, obtains the optimized ground plane normal vector through iterative optimization, and determines the final ground plane equation based on the optimized ground plane normal vector.

[0075] Specifically, although the initial ground plane normal vector reflects the main distribution direction of the point cloud, under critical conditions, non-ground points (such as vehicle undersides, curbs, and low obstacles) inevitably mix into the dynamically dynamic region of interest. These outliers can cause bias in the plane fitting. The Huber loss function, as a classic robust estimation criterion, can use squared loss to maintain efficiency when the residuals are small, and switch to linear loss to suppress the influence of outliers when the residuals are large. The weighting coefficients constructed based on this can dynamically reduce the contribution of non-ground points to the plane fitting. The weighted full least squares method further incorporates the noise characteristics of the point cloud anisotropy by constructing a weighted covariance matrix and solving its eigenvectors, iteratively optimizing the ground plane normal vector until it converges to a robust estimation result.

[0076] Specifically, the domain controller first uses the initial ground plane normal vector obtained in step S31 as the starting point for iteration. For each 3D point in the dynamic region of interest, the signed distance from it to the current ground plane is calculated to obtain the residual. Based on the magnitude of the residual, the weight coefficient of each point is defined using the Huber loss function: when the absolute value of the residual is less than a preset threshold, the weight is assigned a value of 1; when the residual is greater than the threshold, the weight decreases inversely proportional to the residual. Subsequently, the weighted center point is calculated based on the weighted coordinates of all points, and a weighted covariance matrix is ​​constructed. Eigenvalue decomposition is performed on this matrix, and the eigenvector corresponding to the smallest eigenvalue is extracted as the updated normal vector. The updated normal vector is substituted into the plane equation, and the residuals and weights of each point are recalculated, repeating the above process. Iteration continues until the change in the angle between two adjacent normal vectors is less than a preset convergence threshold, or the maximum number of iterations is reached. The final output normal vector and the weighted center point together constitute the final ground plane equation.

[0077] Furthermore, the threshold parameter of the Huber loss function is crucial for distinguishing ground-based points from non-ground outliers. Its value needs to be determined in conjunction with the vehicle's suspension dynamics and road surface smoothness calibration, typically ranging from 5 cm to 15 cm. The calculation of weighting coefficients must ensure numerical stability to avoid numerical overflow caused by excessively small or large residuals. The construction of the weighted covariance matrix uses double-precision floating-point arithmetic, and the numerical error of eigenvalue decomposition is controlled within 1e-8. The iteration convergence threshold is set to a normal vector angle change of less than 0.001 radians to ensure the stability of the estimation results. The maximum number of iterations is typically set to 5 to 10 to balance computational efficiency and accuracy. The computation time of the entire optimization process should be controlled within 30 milliseconds to meet real-time requirements.

[0078] Specifically, in emergency obstacle avoidance scenarios, the dynamic region of interest may contain point clouds from the bottom of the preceding vehicle or scattered obstacles. The Huber loss function effectively suppresses the weights of these outliers, preventing the ground plane from being raised or distorted. When driving over speed bumps or potholes, local undulations in the ground increase the residual, but this does not exceed the threshold range. The algorithm retains these as reasonable interior points, avoiding excessive removal of real ground points. When the vehicle tilts during high-speed cornering, the ground point cloud distribution becomes tilted. The weighted full least squares method, through iterative optimization, accurately tracks the dynamic changes in the ground plane, ensuring the real-time accuracy of the normal vector estimation.

[0079] Specifically, compared to the initial estimation relying solely on principal component analysis, this step significantly reduces sensitivity to non-ground point noise and effectively suppresses the interference of outliers on plane fitting. The introduction of the Huber loss function enables the algorithm to decisively eliminate abnormal outliers while encompassing reasonable ground undulations, achieving an ideal balance between estimation accuracy and robustness. The convergence of iterative optimization ensures that the ground plane estimation results remain stable even under complex dynamic conditions. As the core component of robust ground plane estimation, this step provides a highly reliable geometric benchmark for the subsequent accurate calculation of the dip angle.

[0080] S4. Based on the geometric relationship between the estimated ground plane normal vector and the reference normal vector, the vehicle roll angle is calculated.

[0081] Specifically, in the initial state where the vehicle is stationary on a level surface and the lidar is installed without tilt, the ground plane normal vector coincides with the vertical axis of the vehicle coordinate system and is defined as the reference normal vector. When the vehicle tilts, the vehicle body rotates around its longitudinal axis, and the lidar coordinate system fixed to the vehicle body deflects by the same angle. At this time, the ground plane normal vector observed by the lidar undergoes spatial rotation relative to the reference normal vector. According to the theory of rigid body rotation, the angle between two unit vectors from the same starting point can be solved by vector dot product and cross product operations. Specifically, the tilt angle corresponds to the rotation component of the lidar coordinate system around the coordinate axis of the vehicle's forward direction, which can be calculated by analyzing the geometric relationship between the current ground plane normal vector and the reference normal vector.

[0082] Specifically, the domain controller first defines a reference normal vector, which is acquired when the vehicle is stationary on a level surface and the LiDAR is not tilted. This vector is typically a unit vector pointing vertically upwards. Then, the ground plane normal vector obtained in step S3 is used as the current observation value. A transformation relationship is established between the LiDAR coordinate system and the vehicle coordinate system, decomposing the attitude deviation into a roll angle (rotation around the longitudinal axis) and a pitch angle (rotation around the transverse axis). Based on the spatial rotation matrix, the current normal vector equals the rotation matrix multiplied by the reference normal vector. By solving this matrix equation, an analytical expression for the roll angle can be derived. In the specific calculation, the arctangent of the ratio of the y-axis component to the z-axis component of the current normal vector is used to obtain the numerical solution for the roll angle. This calculation result is directly output as the vehicle roll angle at the current moment for use by the downstream perception and control modules.

[0083] Furthermore, the accuracy of the ground plane normal vector must achieve an angle error of less than 0.5 degrees to ensure the accuracy of the roll angle output. The calibration of the reference normal vector should be performed under unloaded vehicle and level road conditions, with the calibration error controlled within 0.1 degrees. The numerical calculation of the roll angle uses double-precision floating-point arithmetic, and the numerical error of the arctangent function should be controlled within 1e-6 radians. The output frequency of the calculation results must be consistent with the lidar acquisition frequency, typically 10Hz to 20Hz, to ensure the real-time performance of attitude estimation. The resolution of the roll angle should reach 0.1 degrees to sensitively reflect minute changes in vehicle attitude.

[0084] Specifically, in emergency obstacle avoidance scenarios, the vehicle experiences significant body roll during high-speed steering. The real-time calculated roll angle can be used to dynamically adjust the projected position of the obstacle bounding box, compensating for the impact of point cloud distortion on target detection. In scenarios involving driving on bumpy roads, the continuous output of the roll angle provides feedback input to the active suspension control system, assisting in real-time adjustment of the vehicle's attitude. In scenarios involving driving on slopes, the roll angle, combined with the pitch angle, can comprehensively describe the vehicle's attitude relative to the horizontal plane, providing accurate vehicle state information for path planning and motion control. The roll angle output in this step can be cross-checked with the attitude signal acquired by the vehicle's CAN bus, improving the system's fault tolerance.

[0085] Specifically, compared to attitude calculations directly relying on inertial measurement units, this method utilizes the ground plane geometry features in the lidar point cloud, fundamentally avoiding the influence of accelerometer noise, temperature drift, and impact errors. The analytical calculation of the roll angle avoids complex nonlinear optimization processes, significantly reducing computational overhead and ensuring real-time performance. This step transforms the three-dimensional spatial geometry into a single angle output, providing clear and intuitive vehicle attitude information to downstream perception, decision-making, and control modules, significantly improving the intelligent driving system's environmental understanding and safety under critical conditions.

[0086] Furthermore, S4 includes: S41, define the ground plane normal vector when the vehicle is stationary on a horizontal ground and the lidar is installed without tilt as the reference normal vector, and establish an optimized spatial rotation relationship model between the ground plane normal vector and the reference normal vector.

[0087] Specifically, when the vehicle is stationary on a level surface and the lidar is installed without tilt, there is a fixed translational relationship between the lidar coordinate system and the vehicle coordinate system, and the coordinate axes of the two coordinate systems are parallel to each other. At this time, the normal vector of the ground plane in the lidar coordinate system coincides completely with the vertical axis of the vehicle coordinate system. This vector is defined as the reference normal vector and serves as the reference for subsequent attitude calculation. When the vehicle tilts, the vehicle body rotates around its longitudinal axis, and the lidar coordinate system fixed to the vehicle body deflects by the same angle, causing the ground plane normal vector observed by the lidar to rotate spatially relative to the reference normal vector. According to rigid body rotation theory, the rotational relationship between two vectors from the same starting point can be completely described by a rotation matrix, which contains rotation angle information about the three coordinate axes.

[0088] Specifically, the domain controller first acquires the reference normal vector during the vehicle's factory calibration phase or during a stationary period after each startup. This is achieved by collecting at least 100 frames of LiDAR point cloud data under conditions of vehicle idling and parking on a level, hardened surface. The ground plane normal vector is extracted from each frame, and the average value is calculated as the final value of the reference normal vector to eliminate the influence of random noise. Subsequently, a spatial rotation relationship model is established, representing the transformation relationship between the optimized ground plane normal vector and the reference normal vector at the current moment as a three-dimensional rotation matrix. Based on the Euler angle rotation order definition, the vehicle attitude is decomposed into roll angle about the longitudinal axis (x-axis), pitch angle about the transverse axis (y-axis), and yaw angle about the vertical axis (z-axis). Since this method focuses on roll angle estimation, the rotation relationship model can be simplified to consider only rotations about the x-axis and y-axis, where the rotation component about the x-axis is the roll angle to be determined. This mathematical model provides the theoretical basis for the subsequent analytical solution of the roll angle.

[0089] Furthermore, the calibration of the reference normal vector should be carried out on a level field with a road slope of less than 0.5 degrees. During the calibration process, the vehicle tire pressure should meet the standard value, and the load distribution inside the vehicle should be uniform. The calibration accuracy of the reference normal vector must reach an angle error of less than 0.1 degrees to ensure the accuracy of the reference for subsequent attitude calculations. The establishment of the rotation relationship model requires a clear definition of the Euler angle rotation order, usually using the xyz order, to avoid ambiguity in angle calculations due to unclear definition of the rotation order. The numerical expression of the rotation matrix uses double-precision floating-point format, and the numerical error of the matrix elements is controlled within 1e-8. Once the reference normal vector is calibrated, it is stored in the non-volatile memory of the domain controller and is automatically loaded after each power-on. Unless the installation position or angle of the LiDAR changes, no recalibration is required.

[0090] Specifically, during the calibration phase before the vehicle leaves the factory, the accurate acquisition of the baseline normal vector lays a unified reference benchmark for the calculation of roll angles under all subsequent operating conditions. During daily vehicle use, the baseline normal vector can be quickly verified during the stationary period after each start-up. If the deviation from the stored value exceeds a preset threshold, it indicates that the LiDAR installation status may have changed, requiring recalibration. When the vehicle is traveling on a road surface with an unknown slope, the rotation relationship model can decompose the currently observed ground plane normal vector into the part caused by the road slope and the part caused by changes in vehicle attitude, providing theoretical support for the accurate extraction of the vehicle roll angle. This benchmark model is also applicable to the calculation of pitch angles and has good scalability.

[0091] Specifically, the precise calibration of the reference normal vector eliminates the influence of sensor installation errors and vehicle initial state uncertainties on attitude calculation. The rotation relationship model decomposes the complex three-dimensional rotation of the vehicle body into component rotations about each coordinate axis, allowing the roll angle to be solved as an independent degree of freedom. This model provides a clear mathematical path for subsequent analytical calculations of the roll angle, avoiding complex nonlinear optimization processes and significantly reducing computational overhead. Simultaneously, the reference normal vector and the rotation relationship model together constitute the geometric theoretical foundation of the entire roll angle estimation method, ensuring the universality and consistency of the algorithm under different vehicles and installation conditions.

[0092] S42, Based on the aforementioned spatial rotation relationship model, extract the angular components of the rotation of the lidar coordinate system around the coordinate axis of the vehicle's forward direction, and use them as the vehicle's roll angle.

[0093] Specifically, in rigid body rotation, the three-dimensional rotation matrix can be decomposed into sequential rotations around different coordinate axes. Based on the spatial rotation relationship model established in step S41, the transformation matrix between the current ground plane normal vector and the reference normal vector contains rotation information around the three coordinate axes of the lidar coordinate system. Since the vehicle's forward direction coincides with the x-axis of the lidar coordinate system, the roll angle is defined as the rotation component around this axis. By analyzing the projection component of the current normal vector in the plane perpendicular to the x-axis, and utilizing the correspondence between vector geometry and trigonometric functions, the rotation angle around the x-axis can be extracted from the rotation matrix. This angle directly reflects the degree of lateral tilt of the vehicle body relative to the horizontal plane, i.e., the vehicle roll angle.

[0094] Specifically, the domain controller first obtains the current ground plane normal vector obtained from step S3 optimization, which has been normalized to a unit vector. Let the reference normal vector be a vertically upward unit vector, denoted as (0,0,1). According to the Euler angle rotation order definition (e.g., xyz order), the relationship between the coordinate components and the rotation angle of the current normal vector after rotation matrix transformation can be analytically expressed. Specifically, the roll angle φ satisfies tan(φ) = -n_y / n_z, where n_y is the component of the current normal vector along the y-axis (transverse direction of the vehicle), and n_z is the vertical component. The domain controller calls the arctangent function atan2(n_y, n_z) to calculate the roll angle and outputs a signed angle value according to the sign convention: a positive value indicates a rightward roll, and a negative value indicates a leftward roll. This calculation process requires no iteration and is directly solved based on the component relationship between the current normal vector and the reference normal vector.

[0095] Furthermore, the accuracy of the current normal vector needs to achieve an angle error of less than 0.5 degrees to ensure the accuracy of the roll angle output. The numerical implementation of the arctangent function atan2 should support double-precision floating-point operations, with numerical errors controlled within 1e-6 radians. The resolution of the roll angle output needs to reach 0.1 degrees to sensitively reflect minute changes in vehicle attitude. The calculation process should avoid introducing additional latency, with the time cost of a single calculation being less than 1 millisecond, ensuring matching with the 10Hz to 20Hz output frequency of the LiDAR. In addition, the singular case where the vertical component of the normal vector is zero needs to be handled. In this case, the roll angle is directly taken as ±90 degrees, and the direction is determined based on the sign of the y-component.

[0096] Specifically, in emergency obstacle avoidance scenarios, the instantaneous value of the roll angle can be used to dynamically correct the projection distortion of the LiDAR point cloud, ensuring the accuracy of obstacle detection. In active suspension control scenarios, the continuous output of the roll angle serves as a feedback signal, driving the suspension actuators to adjust the vehicle's attitude, improving ride comfort and handling stability. In hill start or off-road driving scenarios, the roll angle information can assist in calculating the vehicle's center of gravity shift and optimize drive force distribution. The roll angle output in this step can also be fused with inertial measurement unit data to construct a redundant attitude estimation system, improving the system's reliability and fault tolerance.

[0097] Specifically, compared to the complex numerical optimization of directly solving the rotation matrix, this step uses analytical formulas, significantly reducing the computational burden and ensuring real-time performance. By utilizing the geometric relationship between the current normal vector and the reference normal vector, accumulated integral errors are avoided, improving long-term stability. This step transforms three-dimensional spatial geometric information into intuitive roll angle values, providing clear and reliable vehicle attitude input for downstream perception, decision-making, and control modules, significantly enhancing the intelligent driving system's environmental adaptability and safety under critical conditions.

[0098] S5 removes outliers from the lidar point cloud data to eliminate noise points caused by lidar vibration and environmental interference.

[0099] Specifically, in critical operating conditions, LiDAR systems generate numerous outliers that deviate from the actual object surface due to factors such as severe vehicle vibration, dust kicked up by tires, and scattering from rain and fog. These outliers exhibit local sparsity in their spatial distribution, meaning their average distance to their neighbors is significantly greater than that of normal points. Simultaneously, the vehicle's own structure (such as the front and hood) may strongly reflect the LiDAR, creating false point clouds within the vehicle's interior. Based on these characteristics, outlier removal employs a dual mechanism: firstly, by statistically analyzing the average distance between each point and its nearest neighbors, locally sparse noise points are identified and filtered out; secondly, by establishing precise three-dimensional forbidden regions, false reflection points located inside the vehicle body are directly removed, thereby restoring the purity of the point cloud data.

[0100] Specifically, the domain controller first constructs a 3D spatial index structure (such as a KD-tree) on the original point cloud to improve the efficiency of nearest neighbor search. For each LiDAR point, it searches for its k nearest neighbors, calculates the Euclidean distance between these neighbors and the point, and takes the average. Assuming that the average distance of all points follows a Gaussian distribution, it calculates the global mean and standard deviation and sets a dynamic threshold. When the average distance of a point exceeds the threshold, it is identified as an outlier and removed. Simultaneously, to address potential self-reflection of laser light at the vehicle's front end, a precise 3D forbidden region is established in the LiDAR coordinate system. This region is obtained by performing polynomial surface fitting on the point cloud within the vehicle's front end while the vehicle is stationary, and is described as a quadratic polynomial function of the x-coordinate. For any point, if its x-coordinate is within the vehicle's front end area, its y-coordinate is within the vehicle's width, and its z-coordinate is below the corresponding height of the fitted surface, it is identified as a false point on the vehicle body and removed.

[0101] Furthermore, the k-value in the nearest neighbor search is typically between 20 and 50, and needs to be dynamically adjusted according to the point cloud density to ensure statistical significance. The threshold coefficient for outlier identification needs to be dynamically adjusted according to the severity of the problem, with a typical range of 1.0 to 3.0. The higher the severity, the more lenient the threshold should be to avoid excessive removal of valid points. The establishment of the 3D forbidden region requires accurate fitting of the vehicle front surface, with the fitting error controlled within 5 cm. The vehicle front range is defined as extending forward from the LiDAR installation position to the foremost point of the vehicle, with a width equal to the width of the vehicle body. The processing time of the entire outlier removal process should be controlled within 20 milliseconds to meet real-time requirements. The removal operation should preserve the integrity of the ground point cloud to ensure that ground points are not misidentified as outliers.

[0102] Specifically, in emergency braking scenarios, the intense friction between tires and the ground generates a large amount of dust, resulting in suspended noise points in the point cloud. Statistical filtering can effectively identify and remove this non-ground noise. When driving on slippery roads, rain and fog scatter the laser, creating fog-like noise that affects the accuracy of ground plane extraction. Outlier removal purifies the data source for subsequent processing. When a vehicle travels over bumpy roads, severe vibrations of the LiDAR may cause some point cloud positions to shift. Statistical filtering can identify and remove these abnormal points caused by vibrations. The establishment of a no-go zone in front of the vehicle effectively eliminates false targets caused by reflections from the vehicle itself, avoiding the risk of misidentifying the vehicle as an obstacle.

[0103] Specifically, the statistical filtering method identifies and removes random noise generated by environmental interference based on local sparsity, while the 3D forbidden region method accurately filters out deterministic false points formed by vehicle self-reflection. These two methods complement each other, significantly improving the signal-to-noise ratio of the point cloud. Early removal of outliers avoids interference from noise points in subsequent region of interest extraction and ground plane estimation, providing a clean data foundation for the stable operation of the entire roll angle estimation method. This step, without increasing hardware costs, enhances the algorithm's adaptability to harsh environments and is a key preprocessing step to ensure the reliability of perception under extreme conditions.

[0104] Furthermore, S5 includes: S51 performs voxel filtering dimensionality reduction on the point cloud data after removing outliers to reduce the amount of point cloud data.

[0105] Specifically, the raw LiDAR point cloud data is massive, with a single frame containing tens of thousands to hundreds of thousands of 3D points. Direct processing would consume significant computational resources and memory bandwidth, impacting the algorithm's real-time performance. Voxel filtering, a classic point cloud dimensionality reduction method, works by dividing the 3D space into regular cubic mesh units, each called a voxel. For all points falling within the same voxel, their geometric center or centroid replaces the original point set, thus significantly reducing the number of points while maintaining the macroscopic geometric structure of the point cloud. This method reduces the computational complexity of subsequent processing and, through local averaging effects, smooths out minor noise to some extent, achieving a balance between data compression and feature preservation.

[0106] Specifically, the domain controller first reads the point cloud data after outlier removal, calculates the maximum and minimum values ​​of the point cloud along the three coordinate axes of the Cartesian coordinate system, and determines the overall bounding box of the point cloud. Based on preset voxel side lengths, the bounding box is uniformly divided into multiple voxel grids along the x, y, and z axes, with the total number of voxels determined by the product of the number of grids in each direction. Then, for each original point cloud, the index number of its corresponding voxel is calculated based on its 3D coordinates, and all points are categorized according to their voxel indices. For each voxel containing at least one point, the average coordinates of all points within that voxel are calculated, and this average point is used as the representative point of that voxel and added to the dimensionality-reduced point cloud set. Voxels containing no points are directly ignored. After processing, the original dense point cloud is sparsified, significantly reducing the number of output point clouds, while preserving the overall spatial distribution.

[0107] Furthermore, the setting of voxel side length directly affects the density of the point cloud after dimensionality reduction. A larger side length results in a more significant dimensionality reduction, but also leads to greater loss of geometric details. In typical applications, the voxel side length ranges from 0.1 meters to 0.5 meters, and needs to be comprehensively calibrated based on the lidar angular resolution, target distance, and the accuracy requirements of subsequent algorithms. For ground plane estimation tasks, the voxel side length is typically around 0.2 meters, which effectively reduces the amount of data while preserving the macroscopic features of ground undulations. The number of points in the dimensionality-reduced point cloud should be controlled between 10% and 30% of the original number to significantly reduce the computational burden. The processing time of the entire voxel filtering process should be controlled within 15 milliseconds to ensure it does not become a performance bottleneck for the overall algorithm. The boundary processing of voxel partitioning must ensure numerical stability to avoid point cloud misalignment caused by floating-point errors.

[0108] Specifically, in emergency obstacle avoidance scenarios, the domain controller needs to complete point cloud processing within milliseconds to support rapid decision-making. Voxel filtering significantly reduces the computational load of subsequent dynamic region of interest extraction and ground plane estimation, ensuring real-time response. In high-speed driving scenarios, the amount of point cloud data increases with scanning distance. Voxel filtering effectively controls the data size, avoiding excessive consumption of computing resources. In embedded platforms with limited storage resources, dimensionality reduction processing enables the algorithm to run stably under limited computing power. Furthermore, for highway scenarios with relatively flat ground features, the voxel side length can be appropriately increased to further compress the data volume without losing key information.

[0109] Specifically, the dimensionality-reduced point cloud, while preserving the macroscopic geometric features of the ground, significantly reduces redundant points, resulting in a reduction of over 50% in the runtime of subsequent algorithms for dynamic region of interest extraction and ground plane estimation. The local averaging effect of voxel filtering also plays a secondary noise reduction role, further smoothing out minor noise and improving the stability of ground plane estimation. This step, as a crucial preprocessing step after outlier removal, provides key support for the real-time performance of the entire roll angle estimation method without sacrificing estimation accuracy, ensuring that the algorithm can run efficiently on computationally limited vehicle platforms.

[0110] This invention discloses a point cloud-based vehicle roll angle estimation method for critical conditions. It acquires LiDAR point cloud data and vehicle kinematic signals, dynamically adjusts the region of interest (ROI) based on the kinematic signals to adaptively focus on the ground plane, and employs a weighted robust optimization method for ground plane estimation. The roll angle is then calculated based on the geometric relationship of the normal vectors. Simultaneously, outlier removal and voxel filtering dimensionality reduction preprocessing are introduced, effectively solving the problems of point cloud distortion caused by violent vehicle movement and inaccurate estimation by traditional methods under critical conditions. This method achieves integrated modeling from data preprocessing, dynamic region extraction, robust ground plane estimation to roll angle calculation, significantly improving the accuracy, robustness, and real-time performance of roll angle estimation, and enhancing the adaptability and engineering applicability of intelligent driving perception systems under extreme conditions.

[0111] Example 2 To achieve the above invention, embodiments of the present invention also provide another method for estimating vehicle roll angle based on point clouds under critical conditions, such as... Figure 2 As shown, it includes: S101, LiDAR data acquisition.

[0112] Specifically, point cloud data is collected by an onboard LiDAR with a horizontal viewing angle of 360 degrees and a vertical viewing angle of 26 degrees, at a frequency of 10Hz. The LiDAR is positioned at the center of the roof of the vehicle without any tilt. Vehicle signals such as vehicle speed, longitudinal acceleration, and lateral acceleration are acquired in real time via the vehicle's CAN bus. The collected point cloud data is transmitted to the domain controller via the network port, and the vehicle signals are transmitted to the domain controller via the CAN interface for further processing.

[0113] S102, Outlier removal from point cloud.

[0114] Specifically, when vehicles perform extreme maneuvers such as emergency braking and obstacle avoidance, or travel on slippery or uneven surfaces, the point cloud will generate numerous outliers that deviate from the actual objects due to severe vibrations, dust, or rain / fog interference from the lidar. These outliers need to be removed. Calculate the average distance from its k nearest neighbors. .

[0115] Furthermore, assume that the average distance between all points follows a Gaussian distribution. Then, eliminate the lidar points that satisfy the following formula: (1) in, These are parameters that are dynamically adjusted based on the severity of the emergency.

[0116] Furthermore, to address potential laser self-reflection at the front of the vehicle, a precise three-dimensional forbidden region is established in the radar coordinate system L to directly filter out false points within the vehicle's range. This is described using the following polynomial: (2) Among them, parameters a , b , c , d It is obtained by fitting the point cloud of the front of the vehicle when it is stationary, assuming To position the laser radar at the very front of the vehicle, Location for lidar installation. W Given the vehicle width, the front end range is... .

[0117] When a certain point is a lidar point ,satisfy If the point is within the area of ​​the front of the vehicle, it is also excluded.

[0118] S103, dimensionality reduction of point cloud data.

[0119] Specifically, due to the large volume of raw point cloud data, direct processing would waste computational resources. Therefore, voxel filtering is used for preprocessing the raw point cloud. The maximum and minimum values ​​of the point cloud data set are then taken along the three coordinate axes of the Cartesian coordinate system. , , , , , .

[0120] Furthermore, set the side length dimensions of the voxels. The three-dimensional space occupied by the point cloud is uniformly divided into Each voxel is of equal size, namely: (3) in, This indicates a round-down operation.

[0121] The point cloud is numbered according to its voxel, and the point cloud in each voxel is divided into parts, i.e.: (4) Then, the point cloud in each voxel is centrally sampled for dimensionality reduction.

[0122] S104, Dynamic Region of Interest Extraction.

[0123] Specifically, in critical operating conditions, fixing the region of interest has certain limitations. In order to ensure the real-time performance and accuracy of roll angle compensation, the point cloud capture boundary is dynamically adjusted by using vehicle kinematic signals.

[0124] Furthermore, the farthest point in the longitudinal range of the point cloud is determined based on the vehicle speed. The following formula can be used for calculation: (5) in, This is the closest point within the intersection area between the radar point cloud and the ground. This is the velocity coefficient.

[0125] Furthermore, based on lateral acceleration Determine the region of interest from the perspective The following formula can be used for calculation: (6) in, This refers to the horizontal field of view of the lidar. The lateral acceleration coefficient, This is the operator for taking the sign.

[0126] Furthermore, based on longitudinal acceleration To determine the appropriate height for the point cloud, use the following formula for calculation: (7) Where m is the vehicle mass. For the height of the vehicle's center of gravity, For suspension stiffness, This refers to the vehicle's wheelbase.

[0127] The height range extracted from the ground points is: (8) in, For the installation height of the lidar, For tolerance.

[0128] Specifically, the lidar point cloud is traversed and judged according to the discrimination threshold calculated above. Points that meet the following conditions are considered to be within the region of interest: (9) S105, Ground plane optimization calculation.

[0129] Specifically, the equation of the ground plane is defined as follows: (10) Then the ground plane normal vector is ; Furthermore, principal component analysis is performed using the covariance matrix of the point set within the region of interest to quickly obtain a coarse normal vector, thereby narrowing the search space. C The calculation formula is: (11) in (12) Specifically, for the covariance matrix C Perform eigenvalue decomposition; the eigenvector corresponding to the smallest eigenvalue is the initial normal vector. Thus, the initial ground plane is obtained; under critical conditions, the initial ground plane often contains noise, and the weighted full least squares method is used to iteratively optimize the area near the initial ground plane to obtain the final ground plane.

[0130] Furthermore, to suppress interference from non-ground points during tilting, the Huber loss function is introduced, defining the residual from the i-th point to the plane as: (13) The corresponding weighting coefficients are: (14) Among them, the threshold is dynamically adjusted based on the vehicle's suspension travel.

[0131] The objective function is to minimize the sum of the weighted squared distances from all points to the plane: (15) Furthermore, to solve the above constrained optimization problem, an auxiliary function is constructed: (16) make ,get: (17) make and will d Substituting into the objective function, then (18) in, For the weighted covariance matrix, use Therefore, the optimization problem (15) becomes: (19) Specifically, the above optimization problem is modeled as a Rayleigh quotient optimization problem. Perform eigenvalue decomposition; the eigenvector corresponding to the smallest eigenvalue is the optimized ground plane normal vector. .

[0132] S106, roll angle calculation.

[0133] Specifically, when the vehicle is stationary on a level surface, assuming the lidar is installed without tilt, the ground plane normal vector is defined as follows: We decompose the attitude deviation of the lidar relative to the ground into orbital deviations. x Side tilt angle of axis rotation and around y Pitch angle of axis rotation When the lidar rotates, the observed normal vector... With reference normal vector The relationship is: (20) in, , After simplification, we get: (twenty one) The formula for calculating the roll angle is: (twenty two) Compared with existing technologies, this invention adopts a region of interest extraction method that integrates vehicle motion characteristics, adaptively focuses on key road areas, and balances processing efficiency and information integrity. It employs a weighted robust ground plane optimization algorithm to improve the accuracy and stability of ground plane estimation under vehicle tilt and complex road conditions. This provides a reliable and independent vehicle roll angle estimation method for autonomous driving systems in crisis situations, thereby improving overall vehicle safety.

[0134] Another method for estimating vehicle roll angle under critical conditions based on point clouds, as described in this invention, collects point cloud data using an onboard LiDAR and combines it with vehicle kinematic signals. Outlier removal is achieved using dynamic thresholding and a 3D prohibited region. Voxel filtering is used for data dimensionality reduction. The longitudinal range, viewing angle range, and height interval of the region of interest are dynamically adjusted based on vehicle speed, lateral acceleration, and longitudinal acceleration. Then, robust optimization estimation of the ground plane is performed based on weighted full least squares and the Huber loss function. Finally, the roll angle is calculated using the geometric relationship between the ground plane normal vector and the reference normal vector. This method effectively solves the problems of point cloud distortion caused by violent vehicle movement under critical conditions and the inaccuracy of traditional estimation methods, achieving stable and reliable roll angle estimation and significantly improving the adaptability and safety of intelligent driving perception systems under extreme conditions.

[0135] Example 3 To achieve the above invention, such as Figure 3As shown, this embodiment also provides a point cloud-based vehicle roll angle estimation device 10 for critical operating conditions. The device 10 includes: The data acquisition and input module 100 is used to acquire lidar point cloud data and vehicle kinematic signals.

[0136] The dynamic region of interest extraction module 200 is used to dynamically adjust the extraction range of the point cloud region of interest based on the vehicle kinematic signals, so as to adaptively focus on the effective area containing the ground plane when the vehicle body is in violent motion.

[0137] The robust ground plane estimation module 300 is used to estimate the ground plane based on a weighted robust optimization method within a dynamic region of interest to obtain the ground plane normal vector.

[0138] The roll angle calculation module 400 is used to calculate the vehicle roll angle based on the geometric relationship between the estimated ground plane normal vector and the reference normal vector.

[0139] This invention discloses a point cloud-based vehicle roll angle estimation device for critical conditions. It employs a data acquisition and input module to obtain LiDAR point cloud data and vehicle kinematic signals; a dynamic region of interest extraction module to adaptively focus on an effective region containing the ground plane based on the kinematic signals; a robust ground plane estimation module to estimate the ground plane normal vector within the dynamic region using a weighted robust optimization method; and a roll angle calculation module to calculate the roll angle based on the geometric relationship of the normal vector. This effectively solves the problems of point cloud distortion caused by violent vehicle movement and inaccurate estimation using traditional methods under critical conditions. It achieves integrated processing from data input, dynamic region extraction, robust ground plane estimation to roll angle calculation, significantly improving the accuracy and stability of vehicle roll angle estimation and enhancing the adaptability and reliability of the intelligent driving perception system under extreme conditions.

[0140] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 4 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads the executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the point cloud-based vehicle roll angle estimation method under critical conditions described above.

[0141] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a point cloud-based vehicle roll angle estimation method under critical conditions as described in the foregoing embodiments.

[0142] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0143] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for estimating vehicle roll angle based on point cloud under critical operating conditions, characterized in that, include: Acquire lidar point cloud data and vehicle kinematic signals; The extraction range of the region of interest in the point cloud is dynamically adjusted based on the vehicle's kinematic signals to adaptively focus on the effective region containing the ground plane when the vehicle body is in violent motion. Within the dynamically active region of interest, the ground plane is estimated using a weighted robust optimization method to obtain the ground plane normal vector; The vehicle roll angle is calculated based on the geometric relationship between the estimated ground plane normal vector and the reference normal vector.

2. The method as described in claim 1, characterized in that, The acquisition of lidar point cloud data and vehicle kinematic signals includes: Raw point cloud data is collected using a lidar rigidly mounted at the center of the vehicle roof. The vehicle kinematics signals, including vehicle speed, longitudinal acceleration, and lateral acceleration, are acquired in real time via the vehicle CAN bus, and the raw point cloud data and vehicle kinematics signals are transmitted to the domain controller for further processing.

3. The method as described in claim 1, characterized in that, The step of dynamically adjusting the extraction range of the point cloud region of interest based on vehicle kinematic signals to adaptively focus on the effective region including the ground plane during violent vehicle movement includes: The longitudinal range of the region of interest is determined based on vehicle speed to adaptively crop the near and far boundaries of the point cloud in the vehicle's direction of travel; the visual range of the region of interest is determined based on lateral acceleration to adaptively adjust the cropping angle of the point cloud in the horizontal direction; and the height range of the region of interest is determined based on longitudinal acceleration to adaptively compensate for changes in ground point height caused by vehicle pitch and suspension deformation. The spatial range of the dynamically interested region is determined by combining the longitudinal range, the viewing angle range, and the height range.

4. The method as described in claim 1, characterized in that, Within the dynamically interested region, the ground plane is estimated using a weighted robust optimization method to obtain the ground plane normal vector, including: Define the equation of the ground plane and the corresponding normal vector expression. Perform principal component analysis based on the covariance matrix of the point cloud in the dynamic region of interest to obtain the initial ground plane normal vector. Based on the initial ground plane normal vector, the Huber loss function is introduced to construct the residual weight coefficients from each point to the ground plane. The weighted full least squares method is used to construct the objective function. The optimized ground plane normal vector is obtained through iterative optimization. The final ground plane equation is determined based on the optimized ground plane normal vector.

5. The method as described in claim 1, characterized in that, The calculation of the vehicle roll angle based on the estimated geometric relationship between the ground plane normal vector and the reference normal vector includes: Define the ground plane normal vector when the vehicle is stationary on a horizontal ground and the lidar is installed without tilt as the reference normal vector, and establish an optimized spatial rotation relationship model between the ground plane normal vector and the reference normal vector. Based on the aforementioned spatial rotation relationship model, the angular components of the rotation of the lidar coordinate system around the coordinate axis of the vehicle's forward direction are extracted and used as the vehicle's roll angle.

6. The method as described in claim 1, characterized in that, After acquiring the lidar point cloud data and vehicle kinematic signals, and before dynamically adjusting the extraction range of the region of interest in the point cloud based on the vehicle kinematic signals, the method further includes: Outlier removal is performed on the lidar point cloud data to eliminate noise points caused by lidar vibration and environmental interference.

7. The method as described in claim 6, characterized in that, After removing outliers from the lidar point cloud data, the process also includes: Voxel filtering is used to reduce the dimensionality of the point cloud data after outlier removal in order to reduce the amount of point cloud data.

8. A vehicle roll angle estimation device based on point cloud under critical operating conditions, characterized in that, include: The data acquisition and input module is used to acquire lidar point cloud data and vehicle kinematic signals; The dynamic region of interest extraction module is used to dynamically adjust the extraction range of the point cloud region of interest based on the vehicle's kinematic signals, so as to adaptively focus on the effective area containing the ground plane when the vehicle body is in violent motion. The robust ground plane estimation module is used to estimate the ground plane based on a weighted robust optimization method within a dynamically dynamic region of interest, so as to obtain the ground plane normal vector; The roll angle calculation module is used to calculate the vehicle roll angle based on the geometric relationship between the estimated ground plane normal vector and the reference normal vector.

9. An electronic device, comprising: processor; The memory stores executable instructions; when the processor executes the instructions, it implements the point cloud-based vehicle roll angle estimation method for critical conditions as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, which, when executed by a processor, implements a point cloud-based method for estimating vehicle roll angle under critical conditions as claimed in any one of claims 1-7.

Images