Large new energy vehicle self-adaptive auxiliary turn method and system based on multiple sensors

Abstract

The application discloses a large new energy vehicle self-adaptive auxiliary turning method and system based on multiple sensors, acquires vehicle surrounding environment information and vehicle state information through a multi-source sensor fusion technology, adopts a self-adaptive fusion algorithm to perform weighted processing on sensor data, obtains accurate and reliable environment sensing results, performs intelligent mode selection and path planning based on a parameter database and an underground garage scene feature database, and performs a turning operation through an optimized four-wheel steering coordination strategy and a self-adaptive torque distribution algorithm. Through real-time monitoring of parameters such as a road surface adhesion coefficient, a vehicle gravity center height and an axle load distribution, the torque output of each driving wheel is dynamically adjusted, and the tire slip problem under the condition of a low adhesion road surface is effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of automotive driver assistance technology. It relates to a method and system for adaptive assisted U-turn of large new energy vehicles based on multiple sensors, which is particularly suitable for underground parking garage scenarios. Background Technology

[0002] In existing technological solutions, traditional vehicle U-turn assistance systems mainly rely on single-sensor perception methods, such as using ultrasonic radar or a monocular camera for obstacle detection and GPS satellite positioning to determine the vehicle's location. With the continued increase in the market penetration of large new energy SUVs, U-turn operations in underground parking garages have become a prominent pain point in the driving experience, mainly in the following aspects: First, existing technologies suffer from significant shortcomings in the reliability of perception in underground parking garage scenarios. Underground parking garage environments are characterized by satellite signal interruptions, dim lighting, dense pillars, and strong wall reflections, making traditional single-sensor perception solutions highly susceptible to interference. For example, ultrasonic radar has blind spots when detecting slender obstacles such as pillars, and its detection distance is typically no more than 2.5 meters, failing to meet the comprehensive environmental perception requirements of large SUVs making U-turns. Monocular cameras experience a significant drop in image quality in dim lighting conditions, and strong wall reflections can lead to overexposure or false obstacles. Furthermore, GPS satellite positioning systems are completely ineffective in underground parking garages, and existing systems lack effective alternative positioning solutions. According to relevant research data, the obstacle detection accuracy of single sensors in underground parking garage environments is typically below 85%, with positioning errors reaching several meters, far from meeting the requirements for safe U-turns.

[0003] Secondly, existing technologies are not optimized for the steering characteristics and body inertia of large SUVs. Large SUVs are characterized by long wheelbases, high centers of gravity, and large masses, resulting in steering response characteristics that differ significantly from compact vehicles. For example, a typical large SUV has a wheelbase of 3.1 meters, a center of gravity height of approximately 0.9 meters, and a curb weight exceeding 2.5 tons. During low-speed U-turns, its body roll angle can reach 3-5 degrees, far exceeding that of ordinary passenger cars. Existing U-turn assistance systems typically employ a fixed steering angular velocity control strategy, failing to dynamically adjust based on vehicle posture. This easily leads to problems such as body swaying and tire slippage during U-turns in large SUVs. Furthermore, the minimum turning radius of large SUVs is typically between 5.5 and 6.5 meters, while the existing systems' four-wheel steering coordination strategy is not sufficiently optimized, making it difficult to effectively reduce the turning radius to within 5 meters.

[0004] Secondly, existing torque distribution strategies for low-adhesion road surfaces have shortcomings. Underground parking garage floors often use epoxy flooring materials, and in wet environments or winter conditions, the coefficient of friction can drop to 0.3-0.5, far lower than the 0.7-0.9 of dry asphalt surfaces. Existing torque distribution algorithms for auxiliary U-turn systems are typically based on a preset fixed ratio, such as a 50:50 front-to-rear axle distribution, without considering changes in the coefficient of friction. When a large SUV makes a sharp turn on a low-adhesion surface, the outer drive wheel is prone to exceeding its traction limit and slipping, leading to vehicle instability. Related test data shows that on epoxy floors with a coefficient of friction of 0.4, the probability of tire slippage during a U-turn for a large SUV using a fixed torque distribution strategy is as high as 35%, seriously affecting driving safety.

[0005] Finally, existing technologies lack spatial adaptability, making it difficult to meet the U-turn requirements in the narrow spaces of underground parking garages. Standard underground parking garage lanes are 5.0-5.5 meters wide, while large SUVs are nearly 2.0 meters wide, leaving only 1.5-1.75 meters of safe space on both sides for U-turns. Existing U-turn assistance systems lack special modes such as crab assist, failing to achieve lateral vehicle movement. This necessitates multiple directional adjustments when making U-turns in narrow lanes, resulting in complex and time-consuming operations. According to actual test data, traditional methods require an average of 3-5 forward and reverse maneuvers in a 5.0-meter-wide lane, taking approximately 45-60 seconds, significantly increasing the driver's workload and psychological stress. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for adaptive assisted U-turn of large new energy vehicles based on multiple sensors. It acquires information about the vehicle's surrounding environment and vehicle status through multi-source sensor fusion technology, performs U-turn operations through an optimized four-wheel steering coordination strategy and adaptive torque distribution algorithm, and dynamically adjusts the torque output of each drive wheel by real-time monitoring of parameters such as road surface adhesion coefficient, vehicle center of gravity height, and axle load distribution, effectively solving the problem of tire slippage under low-adhesion road surface conditions.

[0007] The technical solution to achieve the purpose of this invention is as follows: An adaptive assisted U-turn method for large new energy vehicles based on multiple sensors includes the following steps: S01: Acquire millimeter-wave radar data, binocular camera data, speedometer data, and inertial navigation system data; fuse the data from each sensor to obtain information about the vehicle's surrounding environment and vehicle status. S02: Based on the acquired environmental and vehicle status information, and combined with the preset large SUV parameter database and scene feature database, select the U-turn mode and generate the U-turn trajectory; S03: Establish a torque optimization model that includes center of gravity height compensation. By monitoring the road surface adhesion coefficient, vehicle center of gravity height, and axle load distribution in real time, dynamically adjust the torque output of each drive wheel. Control the output of each wheel through the four-wheel motor drive according to the calculated torque distribution ratio, and control the four-wheel steering to perform a U-turn operation.

[0008] In the preferred technical solution, the formula for fusing the data from each sensor in step S01 is as follows: D_fusion = ω1D_radar + ω2D_vision + ω3D_LDV + ω4D_SINS Where D_fusion represents the fusion distance, D_radar, D_vision, D_LDV, and D_SINS represent the distance data measured by millimeter-wave radar, machine vision, LDV lidar, and SINS inertial navigation system, respectively; ω1, ω2, ω3, and ω4 are the corresponding fusion weights, and ω1+ω2+ω3+ω4=1. The fusion weights are dynamically adjusted based on the real-time confidence levels of each sensor and environmental conditions. The fusion weights consist of two parts: a base weight and a dynamically adjusted weight, i.e., ωi = ωi_base + Δωi, where ωi_base is the base weight of the sensor and Δωi is the dynamic adjustment amount calculated based on the real-time status.

[0009] In the preferred technical solution, the adjustment strategy for the fusion weights includes: The dynamic adjustment amount Δω1 of the weight ω1 of the millimeter-wave radar is: Δω1 = 0.15×(Q_radar×M_radar×D_radar - 1) Where Q_radar is the signal quality factor, M_radar is the multipath interference factor, and D_radar is the detection range factor; The dynamic adjustment amount Δω2 of the weight ω2 of the binocular camera is: Δω2 = 0.2×(L_vision×C_vision×R_vision - 1) Where L_vision is the illumination condition factor, C_vision is the image sharpness factor, and R_vision is the target recognition confidence factor; The dynamic adjustment amount Δω3 of the weight ω3 of the laser velocimeter is: Δω3 = 0.1×(A_LDV×S_LDV×U_LDV - 1) Where A_LDV is the measurement accuracy factor, S_LDV is the speed range factor, and U_LDV is the data consistency factor; The dynamic adjustment amount Δω4 of the weight ω4 of the inertial navigation system is: Δω4 = 0.1×(E_SINS×K_SINS×G_SINS - 1) Where E_SINS is the cumulative error factor, K_SINS is the motion state factor, and G_SINS is the external reference factor.

[0010] In the preferred technical solution, step S01 further includes: By tightly coupling and integrating velocimeter data and inertial navigation system data, and employing a local anomaly factor algorithm to perform real-time anomaly detection on the laser velocimeter data, accurate positioning in satellite-free environments is achieved. The local anomaly factor algorithm includes: The laser speedometer data is preprocessed, including noise reduction, outlier removal, and conversion of the sensor coordinate system to the vehicle coordinate system. For each data point p_i in the dataset, calculate its Euclidean distance to all other data points, sort the distances in ascending order, and the k-th smallest distance is called the k-distance of point p_i, denoted as k-distance(p_i). Construct the k-distance neighborhood N_k(p_i) of p_i. For any two data points p_i and p_j, the reachability distance of point p_j with respect to point p_i is defined as: reachability_distance(p_i, p_j) = max{k-distance(p_i), d(p_i, p_j)}; The local reachability density of a data point p_i is defined as the reciprocal of the average reachability distance of all data points in its k-neighborhood: LRD(p_i)=1 / (Σ_{p_j∈N_k(p_i)} reachability_distance(p_i, p_j) / k); The Local Outlier Factor (LOF) of a data point p_i is defined as the ratio of the average local reachability of all data points in its k-neighborhood to its own local reachability: LOF(p_i) = (Σ_{p_j∈N_k(p_i)} LRD(p_j) / k) / LRD(p_i). When the LOF is within the threshold range, the point is considered a normal data point; when the LOF is greater than the threshold, it is considered an outlier data point. When abnormal data points are detected, the system adopts a hierarchical processing strategy: for mild anomalies with LOF values ​​between 1.5 and 2.0, linear interpolation is used for correction; for moderate anomalies with LOF values ​​between 2.0 and 3.0, quadratic polynomial fitting is used for correction; for severe anomalies with LOF values ​​> 3.0, the data point is directly removed; and the weight of the speedometer data at that moment is reduced in subsequent fusion calculations.

[0011] In a preferred technical solution, the fusion of data from various sensors further includes spatiotemporal synchronization of the data, wherein the spatiotemporal synchronization includes: Time synchronization: Set hardware timestamps for data from each sensor to ensure hardware timestamp synchronization; During fusion computing, the time reference of the inertial navigation system data is used to extract the data of each sensor within the time window from the buffer; for sensors that do not have matching data within the time window, a linear interpolation method is used for time alignment. Using attitude and velocity information provided by the inertial navigation system, a vehicle motion model is established. This motion model is then used to predict the position of obstacles detected by the camera from the detection time to the current time, achieving prediction compensation synchronization; and Spatial synchronization: By pre-calibrating or online-calibrating the sensor extrinsic parameters, the data in each sensor coordinate system are uniformly transformed to the vehicle coordinate system; Dynamic compensation and online calibration: The vehicle pitch and roll angles are monitored in real time through the inertial navigation system. When the pitch angle changes beyond a certain angle, the height component in the sensor extrinsic parameters is adjusted to compensate for the height change caused by suspension compression. The natural feature matching method is adopted. When the vehicle passes through an environment with known features, the feature positions detected by each sensor are compared with the known positions to calculate the extrinsic error. When the error exceeds the threshold, the extrinsic update is triggered.

[0012] In the preferred technical solution, the torque optimization model that includes center of gravity height compensation is as follows: V_wheel = V_base + k1δ + k2(1-μ) + k3H Where V_wheel represents the actual rotational speed of the wheel, V_base represents the base speed calculated by the system, k1 is the steering compensation coefficient, k2 is the adhesion compensation coefficient, k3 is the center of gravity compensation coefficient, δ is the steering angle, μ is the road adhesion coefficient, and H is the vehicle center of gravity height.

[0013] In a preferred embodiment, the dynamic adjustment of the torque output of each drive wheel further includes a hierarchical optimization process, comprising: The first layer is torque limitation based on the adhesion coefficient. The maximum available torque T_max = μmg is calculated based on the road adhesion coefficient μ, where m is the vehicle mass and g is the gravitational acceleration. The second layer is axle load distribution based on center of gravity height. According to the vehicle's center of gravity height H and steering angle δ, the torque distribution ratio between the front and rear axles is dynamically adjusted, where the center of gravity height H is positively correlated with the proportion of torque distributed to the rear axle. The third layer is differential compensation based on steering angle. According to the steering angle δ, a closed-loop feedback control strategy is used to increase the torque compensation of the outer wheel so that its speed matches the theoretical speed difference.

[0014] In a preferred technical solution, a U-turn mode is selected from a preset mode determination matrix based on the ratio of lane width to vehicle width and the road surface adhesion coefficient. The U-turn mode includes at least a four-wheel steering mode and a crab assist mode. When the crab assist mode is selected, a crab turn trajectory is generated, and two-stage crab steering control is performed. In the first stage, the front and rear wheels are controlled to turn in the same direction to a preset angle range, so that the vehicle can move laterally. In the second stage, after the lateral translation is completed, the system switches to four-wheel steering mode and controls the front and rear wheels to turn in opposite directions to complete the remaining turn.

[0015] This invention also discloses a multi-sensor-based adaptive assisted U-turn system for large new energy vehicles, used to implement the aforementioned multi-sensor-based adaptive assisted U-turn method for large new energy vehicles, comprising: The environmental perception module acquires data from millimeter-wave radar, binocular camera, speedometer, and inertial navigation system, and fuses the data from each sensor to obtain information about the vehicle's surrounding environment and vehicle status. The mode selection and path planning module selects the U-turn mode and generates the U-turn trajectory based on the acquired environmental information and vehicle status information, combined with the preset large SUV parameter database and scene feature library. The torque distribution and steering control module establishes a torque optimization model that includes center of gravity height compensation. By monitoring the road surface adhesion coefficient, vehicle center of gravity height, and axle load distribution in real time, it dynamically adjusts the torque output of each drive wheel. The four-wheel motor drive controls the output of each wheel according to the calculated torque distribution ratio, and controls the four-wheel steering to perform a U-turn operation.

[0016] The present invention also discloses a computer storage medium storing a computer program, which, when executed, implements the above-mentioned adaptive assisted U-turn method for large new energy vehicles based on multiple sensors.

[0017] Compared with the prior art, the significant advantages of this invention are: 1. This invention employs a four-source fusion scheme of radar, vision, laser, and inertial navigation. Through the coordinated operation of millimeter-wave radar, binocular surround-view cameras, laser velocimeters, and an inertial navigation system, it achieves high-precision positioning and obstacle detection in satellite-free environments such as underground parking garages. The adaptive torque distribution strategy of this invention significantly improves safety and stability under low-adhesion road surface conditions. By establishing a torque optimization model that includes center-of-gravity height compensation, the system can dynamically adjust torque distribution based on parameters such as road surface adhesion coefficient, vehicle center-of-gravity height, and axle load distribution, reducing the risk of tire slippage by 95%. On low-adhesion road surfaces with μ=0.4, after adopting the technical solution of this invention, the vehicle roll control accuracy is improved by 45%, and the side slip angle during steering is always controlled within 3 degrees, significantly improving driving stability.

[0018] 2. This invention also innovatively proposes a tightly coupled integrated positioning scheme of SINS and dual 2D-LDV. By detecting abnormal data through the LOF algorithm and dynamically adjusting the weights of each sensor using an adaptive filter, the positioning error in satellite-free environments can be controlled within 10 centimeters.

[0019] 3. This invention significantly improves space adaptability through its optimized four-wheel steering and crab-walking mode for large SUVs. By coordinating front and rear wheel steering in opposite directions, the minimum turning radius of large SUVs can be reduced from the traditional 5.5-6.5 meters to 4.95 meters, a reduction of 10-24%. The crab-walking assist mode achieves lateral movement by steering the front and rear wheels in the same direction, reducing the space required for U-turns by more than 30%. Actual test data shows that in a 5.0-meter-wide underground parking garage lane, the U-turn efficiency is increased by 70% after adopting the technical solution of this invention, eliminating the need for multiple turns and achieving a success rate of over 95% on the first attempt. Attached Figure Description

[0020] Figure 1 This is a flowchart of the adaptive assisted U-turn method for large new energy vehicles based on multiple sensors in this embodiment; Figure 2 This is a schematic diagram of the overall architecture of this embodiment; Figure 3 This is a schematic diagram illustrating the principle of multi-sensor fusion distance calculation. Figure 4 Flowchart for determining the scene mode in an underground parking garage; Figure 5 This is a schematic diagram of the trajectory in the crab-like movement assist mode. Detailed Implementation

[0021] The principle of this invention is as follows: This invention constructs a closed-loop adaptive control system for large SUVs in underground parking garage scenarios. The system first acquires information about the vehicle's surrounding environment and vehicle status through multi-source sensor fusion technology. Then, based on a large SUV parameter database and an underground parking garage scenario feature database, it performs intelligent mode selection and path planning. Finally, it executes a U-turn operation through an optimized four-wheel steering coordination strategy and an adaptive torque distribution algorithm. From a control theory perspective, this system adopts a closed-loop control architecture of "perception-decision-execution." First, a multi-sensor system collects real-time information about the vehicle's surrounding environment, including obstacle distance, relative speed, and spatial geometric features. Then, an adaptive fusion algorithm is used to weight and process the data from each sensor to obtain accurate and reliable environmental perception results. Finally, based on the perception results, the optimal U-turn trajectory is planned, and trajectory tracking is achieved through precise wheel speed control.

[0022] Example 1: like Figure 1 As shown, a multi-sensor-based adaptive assisted U-turn method for large new energy vehicles includes the following steps: S01: Acquire millimeter-wave radar data, binocular camera data, speedometer data, and inertial navigation system data; fuse the data from each sensor to obtain information about the vehicle's surrounding environment and vehicle status. S02: Based on the acquired environmental and vehicle status information, and combined with the preset large SUV parameter database and scene feature database, select the U-turn mode and generate the U-turn trajectory; S03: Establish a torque optimization model that includes center of gravity height compensation. By monitoring the road surface adhesion coefficient, vehicle center of gravity height, and axle load distribution in real time, dynamically adjust the torque output of each drive wheel. Control the output of each wheel through the four-wheel motor drive according to the calculated torque distribution ratio, and control the four-wheel steering to perform a U-turn operation.

[0023] In a preferred embodiment, the formula for fusing the data from each sensor in step S01 is as follows: D_fusion = ω1D_radar + ω2D_vision + ω3D_LDV + ω4D_SINS Where D_fusion represents the fusion distance, D_radar, D_vision, D_LDV, and D_SINS represent the distance data measured by millimeter-wave radar, machine vision, LDV lidar, and SINS inertial navigation system, respectively; ω1, ω2, ω3, and ω4 are the corresponding fusion weights, and ω1+ω2+ω3+ω4=1. The fusion weights are dynamically adjusted based on the real-time confidence levels of each sensor and environmental conditions. The fusion weights consist of two parts: a base weight and a dynamically adjusted weight, i.e., ωi = ωi_base + Δωi, where ωi_base is the base weight of the sensor and Δωi is the dynamic adjustment amount calculated based on the real-time status.

[0024] In a preferred embodiment, the adjustment strategy for the fusion weights includes: The dynamic adjustment amount Δω1 of the weight ω1 of the millimeter-wave radar is: Δω1 = 0.15×(Q_radar×M_radar×D_radar - 1) Where Q_radar is the signal quality factor, M_radar is the multipath interference factor, and D_radar is the detection range factor; The dynamic adjustment amount Δω2 of the weight ω2 of the binocular camera is: Δω2 = 0.2×(L_vision×C_vision×R_vision - 1) Where L_vision is the illumination condition factor, C_vision is the image sharpness factor, and R_vision is the target recognition confidence factor; The dynamic adjustment amount Δω3 of the weight ω3 of the laser velocimeter is: Δω3 = 0.1×(A_LDV×S_LDV×U_LDV - 1) Where A_LDV is the measurement accuracy factor, S_LDV is the speed range factor, and U_LDV is the data consistency factor; The dynamic adjustment amount Δω4 of the weight ω4 of the inertial navigation system is: Δω4 = 0.1×(E_SINS×K_SINS×G_SINS - 1) Where E_SINS is the cumulative error factor, K_SINS is the motion state factor, and G_SINS is the external reference factor.

[0025] In a preferred embodiment, step S01 further includes: By tightly coupling and integrating velocimeter data and inertial navigation system data, and employing a local anomaly factor algorithm to perform real-time anomaly detection on the laser velocimeter data, accurate positioning in satellite-free environments is achieved. The local anomaly factor algorithm includes: The laser speedometer data is preprocessed, including noise reduction, outlier removal, and conversion of the sensor coordinate system to the vehicle coordinate system. For each data point p_i in the dataset, calculate its Euclidean distance to all other data points, sort the distances in ascending order, and the k-th smallest distance is called the k-distance of point p_i, denoted as k-distance(p_i). Construct the k-distance neighborhood N_k(p_i) of p_i. For any two data points p_i and p_j, the reachability distance of point p_j with respect to point p_i is defined as: reachability_distance(p_i, p_j) = max{k-distance(p_i), d(p_i, p_j)}; The local reachability density of a data point p_i is defined as the reciprocal of the average reachability distance of all data points in its k-neighborhood: LRD(p_i)=1 / (Σ_{p_j∈N_k(p_i)} reachability_distance(p_i, p_j) / k); The Local Outlier Factor (LOF) of a data point p_i is defined as the ratio of the average local reachability density of all data points in its k-neighborhood to its own local reachability density: LOF(p_i) = (Σ_{p_j∈N_k(p_i)} LRD(p_j) / k) / LRD(p_i). When LOF≈1 (e.g., 1...), the LOD(p_i) value is determined. + When the LOF is 0.1, the point is considered a normal data point; when the LOF is greater than the threshold of 1, it is considered an abnormal data point. When abnormal data points are detected, the system adopts a hierarchical processing strategy: for mild anomalies with LOF values ​​between 1.5 and 2.0, linear interpolation is used for correction; for moderate anomalies with LOF values ​​between 2.0 and 3.0, quadratic polynomial fitting is used for correction; for severe anomalies with LOF values ​​> 3.0, the data point is directly removed; and the weight of the speedometer data at that moment is reduced in subsequent fusion calculations.

[0026] In a preferred embodiment, fusing the data from each sensor further includes spatiotemporal synchronization of the data from each sensor, wherein the spatiotemporal synchronization includes: Time synchronization: Set hardware timestamps for data from each sensor to ensure hardware timestamp synchronization; During fusion computing, the time reference of the inertial navigation system data is used to extract the data of each sensor within the time window from the buffer; for sensors that do not have matching data within the time window, a linear interpolation method is used for time alignment. Using attitude and velocity information provided by the inertial navigation system, a vehicle motion model is established. This motion model is then used to predict the position of obstacles detected by the camera from the detection time to the current time, achieving prediction compensation synchronization; and Spatial synchronization: By pre-calibrating or online-calibrating the sensor extrinsic parameters, the data in each sensor coordinate system are uniformly transformed to the vehicle coordinate system; Dynamic compensation and online calibration: The vehicle pitch and roll angles are monitored in real time through the inertial navigation system. When the pitch angle changes beyond a certain angle, the height component in the sensor extrinsic parameters is adjusted to compensate for the height change caused by suspension compression. The natural feature matching method is adopted. When the vehicle passes through an environment with known features, the feature positions detected by each sensor are compared with the known positions to calculate the extrinsic error. When the error exceeds the threshold, the extrinsic update is triggered.

[0027] In a preferred embodiment, the torque optimization model including center of gravity height compensation is as follows: V_wheel = V_base + k1δ + k2(1-μ) + k3H Where V_wheel represents the actual rotational speed of the wheel, V_base represents the base speed calculated by the system, k1 is the steering compensation coefficient, k2 is the adhesion compensation coefficient, k3 is the center of gravity compensation coefficient, δ is the steering angle, μ is the road adhesion coefficient, and H is the vehicle center of gravity height.

[0028] In a preferred embodiment, dynamically adjusting the torque output of each drive wheel further includes a hierarchical optimization process, comprising: The first layer is torque limitation based on the adhesion coefficient. The maximum available torque T_max = μmg is calculated based on the road adhesion coefficient μ, where m is the vehicle mass and g is the gravitational acceleration. The second layer is axle load distribution based on center of gravity height. According to the vehicle's center of gravity height H and steering angle δ, the torque distribution ratio between the front and rear axles is dynamically adjusted, where the center of gravity height H is positively correlated with the proportion of torque distributed to the rear axle. The third layer is differential compensation based on steering angle. According to the steering angle δ, a closed-loop feedback control strategy is used to increase the torque compensation of the outer wheel so that its speed matches the theoretical speed difference.

[0029] In a preferred embodiment, a U-turn mode is selected from a preset mode determination matrix based on the ratio of lane width to vehicle width and the road surface adhesion coefficient. The U-turn mode includes at least a four-wheel steering mode and a crab assist mode. When the crab assist mode is selected, a crab turn trajectory is generated, and two-stage crab steering control is performed. In the first stage, the front and rear wheels are controlled to turn in the same direction to a preset angle range, so that the vehicle can move laterally. In the second stage, after the lateral translation is completed, the system switches to four-wheel steering mode and controls the front and rear wheels to turn in opposite directions to complete the remaining turn.

[0030] In another embodiment, a computer storage medium stores a computer program that, when executed, implements the aforementioned multi-sensor-based adaptive assisted U-turn method for large new energy vehicles. The reconstruction method described above will not be elaborated further here.

[0031] In another embodiment, a multi-sensor-based adaptive assisted U-turn system for large new energy vehicles is provided to implement the aforementioned multi-sensor-based adaptive assisted U-turn method for large new energy vehicles, comprising: The environmental perception module acquires data from millimeter-wave radar, binocular camera, speedometer, and inertial navigation system, and fuses the data from each sensor to obtain information about the vehicle's surrounding environment and vehicle status. The mode selection and path planning module selects the U-turn mode and generates the U-turn trajectory based on the acquired environmental information and vehicle status information, combined with the preset large SUV parameter database and scene feature library. The torque distribution and steering control module establishes a torque optimization model that includes center of gravity height compensation. By monitoring the road surface adhesion coefficient, vehicle center of gravity height, and axle load distribution in real time, it dynamically adjusts the torque output of each drive wheel. The four-wheel motor drive controls the output of each wheel according to the calculated torque distribution ratio, and controls the four-wheel steering to perform a U-turn operation.

[0032] Specifically, the workflow of a multi-sensor-based adaptive assisted U-turn system for large new energy vehicles is described below using a preferred embodiment as an example: This embodiment addresses a scenario where a large new energy SUV makes a U-turn in the main passageway of an underground parking garage. The specific conditions for this scenario are: lane width 5.0 meters, epoxy flooring, road surface adhesion coefficient μ=0.4 in a wet environment, and vehicle parameters: length 5.3 meters, width 1.95 meters, wheelbase 3.1 meters, center of gravity height 0.9 meters, and curb weight 2.6 tons. This scenario is the most representative U-turn scenario in underground parking garages, as the lane width is only 2.56 times the vehicle width, the space is relatively narrow, and the low-adhesion road surface increases the difficulty of control.

[0033] In this embodiment, the perception layer configuration includes: a 77GHz 4D millimeter-wave radar (6-meter detection radius, obstacle distance accuracy ±3cm) arranged at four corners, a binocular surround-view camera group (180-degree field of view), dual 2D-LDV laser speedometers (speed range 0-20km / h, accuracy ±0.1km / h), a SINS inertial navigation system (attitude angle accuracy ±0.1 degrees), and a road adhesion coefficient sensor for epoxy flooring (detection range μ=0.2-0.8). The decision control layer configuration includes: a large SUV parameter database (storing parameters such as wheelbase, track width, and center of gravity height for different models), an underground parking garage scene feature database (containing parameters such as typical lane width and column spacing), and an adaptive torque distribution algorithm module. The execution layer configuration includes: a four-wheel independent steering drive (front and rear wheel steering angle ±12 degrees), an electronic differential (supporting front and rear axle torque distribution), a four-wheel independent motor drive (torque control accuracy ±1N·m), and a distributed electronic braking system.

[0034] like Figure 2 As shown, the implementation steps include: Step 1: Function Activation and System Self-Check. When the driver activates the assisted U-turn function, the system first performs a self-check to confirm that the ABS, EPS, and four-wheel steering systems are fault-free. If the system detects that the seat belts are properly worn and the current vehicle speed is 12 km / h (meeting the activation condition of ≤15 km / h), the system enters the ready state.

[0035] Step Two: Environmental Perception and Parameter Acquisition. All sensors in the perception layer begin working synchronously: the 4D millimeter-wave radar detects a distance of 1.0 meter to the left pillar, 1.2 meters to the right pillar, and 8.5 meters to the wall in front; the binocular camera identifies the lane markings on both sides and calculates the lane width to be 5.0 meters; the dual 2D-LDV measures the current vehicle speed as 12 km / h, with the vertical speed close to zero; the road adhesion coefficient sensor acquires the current road surface μ=0.4; and the SINS provides the vehicle's current attitude information (pitch angle 0.5 degrees, roll angle 0.2 degrees, and initial yaw angle). The central control unit constructs a three-dimensional environmental model using a four-source fusion algorithm, with fusion weights set to ω1=0.4 (radar), ω2=0.2 (vision), ω3=0.25 (LDV), and ω4=0.15 (SINS).

[0036] Step 3: Mode Selection and Path Planning. For example... Figure 4As shown, the system queries the mode determination matrix based on the parameters of a large SUV (5.3 meters in length and 1.95 meters in width) and the current environmental parameters (5.0 meters in lane width and μ=0.4). Since the lane width is 2.56 times the vehicle width and the adhesion coefficient μ=0.4≥0.4, the system determines to adopt the "four-wheel steering mode". The path planning module generates the U-turn trajectory based on the improved Bézier curve algorithm, sets the turning radius to 4.95 meters, controls the trajectory offset to within 10 centimeters, and sets the safe distance between the turning center and the wall to 30 centimeters.

[0037] Step 4: Adaptive Torque Calculation and Distribution. This is the core step of this embodiment. The system uses a torque optimization formula that includes center of gravity height compensation to calculate the target torque for each drive wheel. First, obtain the basic parameters: vehicle center of gravity height H = 0.9 meters, road adhesion coefficient μ = 0.4, current steering angle δ = 12 degrees (front wheel), total vehicle mass m = 2600 kg, and gravitational acceleration g = 9.8 m / s². Then, calculate the vertical load on each drive wheel: the front axle load is approximately 45% of the total vehicle mass (1170 kg), and the rear axle load is approximately 55% (1430 kg). This is because the power battery of large SUVs is usually located at the rear of the chassis. Next, calculate the target wheel speed: the target speed of the front wheel corresponds to 8 km / h, and the target speed of the rear wheel corresponds to 6 km / h. This is to utilize the larger load on the rear axle to provide better driving force. Finally, apply the torque optimization formula: V_wheel = V_base + k1δ + k2(1-μ) + k3H, where V_wheel represents the actual wheel rotation speed, and V_base represents the base speed calculated by the system. Where k1 is the steering compensation coefficient (value 0.05), k2 is the adhesion compensation coefficient (value 0.3), and k3 is the center of gravity compensation coefficient (value 0.15). The calculation shows that the total torque is reduced by 28%, and the front-to-rear axle torque distribution ratio is adjusted to 55:45, that is, the front axle torque is 45% of the total torque, and the rear axle torque is 55% of the total torque.

[0038] Step 5: Steering and Drive Coordinated Control. The EPS system controls the front wheel steering angle to 12 degrees and the rear wheel steering in the opposite direction to 10 degrees, creating a four-wheel steering synergy effect and effectively shortening the turning radius. The steering angular velocity adopts a dynamic adjustment strategy, with an initial angular velocity of 0.3 rad / s, gradually increasing to 0.7 rad / s as the steering angle increases to avoid sudden changes in body roll. The four independent motor drives precisely control the output of each wheel according to the calculated torque distribution ratio, and the electronic differential monitors the speed difference between the left and right wheels in real time, ensuring smooth steering through torque fine-tuning.

[0039] Step Six: Attitude Monitoring and Dynamic Adjustment. During the U-turn, the gyroscope monitors the vehicle's sideslip angle in real time. When the detected sideslip angle is 2.5 degrees (below the 3-degree threshold), the system maintains the current control parameters. When the vehicle turns to a 90-degree position, the distance to the pillar on the right side decreases to 0.8 meters, and the system automatically adjusts the steering angle by 0.5 degrees to maintain a safe distance. The entire U-turn process lasts approximately 18 seconds, with a heading angle change of 180 degrees.

[0040] Step 7: Function Exit and Low-Speed ​​Cruise. After the heading angle reaches 180 degrees, the instrument cluster will display "Please take over steering," and the system will enter crawl mode (vehicle speed ≤ 3km / h). The driver lightly presses the accelerator pedal, and the system will detect the acceleration operation, exit the assisted mode, and switch to the underground parking garage low-speed cruise mode, maintaining the vehicle speed below 10km / h.

[0041] Regarding the implementation details of the adaptive torque distribution algorithm, this invention employs a layered optimization strategy. The first layer is a torque limiting layer based on the adhesion coefficient. The maximum usable torque T_max = μmg is calculated based on the road adhesion coefficient μ, where m is the vehicle mass and g is the acceleration due to gravity. When μ = 0.4, T_max is approximately 10192 N·m (2600 kg × 9.8 m / s² × 0.4). The second layer is an axle load distribution layer based on the center of gravity height. Considering that large SUVs have a high center of gravity (0.8-1.0 meters), significant load transfer occurs during steering, increasing the vertical load on the outer wheels and decreasing the vertical load on the inner wheels. Therefore, the front and rear axle torque distribution ratio needs to be dynamically adjusted. The third layer is a differential compensation layer based on the steering angle. During steering, the outer wheels travel a longer path, requiring an appropriate increase in the torque of the outer wheels to maintain speed balance.

[0042] Mechanism for dynamically adjusting the front and rear axle torque distribution ratio The system calculates the torque distribution ratio between the front and rear axles in real time based on the road surface adhesion coefficient μ and the vehicle's center of gravity height H. Specific implementation details are as follows: Step one involves a coarse distribution based on the coefficient of adhesion. When μ ≥ 0.6 (high-adhesion road surface), a balanced 50:50 front-to-rear axle distribution is used; when 0.4 ≤ μ < 0.6 (medium-adhesion road surface), the distribution is adjusted to 45% front axle and 55% rear axle; when μ < 0.4 (low-adhesion road surface), the distribution is further adjusted to 40% front axle and 60% rear axle. This layered strategy ensures that on low-adhesion road surfaces, the rear axle bears more torque (greater rear axle load), preventing front axle slippage.

[0043] Step two involves fine-tuning based on the center of gravity height. Considering the rear-mounted nature of the power battery in large SUVs, the static load on the rear axle typically accounts for 55-60%. During steering, centrifugal force causes load transfer, increasing the vertical load on the outer wheels and decreasing the vertical load on the inner wheels. The system fine-tunes the torque distribution using the k3H term; for every 0.1-meter increase in H, the front axle torque ratio decreases by 1.5%, and the rear axle torque ratio increases accordingly.

[0044] Step three involves dynamic compensation based on the steering angle. During steering, the outer wheel travels a longer path than the inner wheel. The system monitors the speed difference between the inner and outer wheels in real time via an electronic differential. When the speed difference deviates from the theoretical value by more than 5%, the system automatically increases the torque output to the outer wheel to ensure coordinated speeds of all four wheels.

[0045] Precise control of torque on the outer wheels to maintain speed balance The torque compensation for the outer wheels adopts a closed-loop feedback control strategy, which includes five steps: Step 1: Calculate the theoretical speed difference. Based on the current steering angle δ, calculate the steering radius R = L / tan(δ), where L is the wheelbase. Then calculate the theoretical speed ratio of the inner and outer wheels: ratio = (R+W / 2) / (RW / 2), where W is the track width.

[0046] Step 2: Measure the actual speed difference. The actual speeds n_FL, n_FR, n_RL, and n_RR of the four wheels are measured in real time using wheel speed sensors, and the ratio of the actual speeds of the inner and outer wheels is calculated.

[0047] Step 3: Calculate the torque compensation amount. The formula for calculating the torque compensation amount ΔT of the outer wheel is: ΔT = k1×δ×T_base×(1 - n_actual / n_theoretical), where k1 is the steering compensation coefficient, δ is the current steering angle, T_base is the base torque, n_actual is the actual speed of the outer wheel, and n_theoretical is the theoretical speed of the outer wheel.

[0048] Step 4: Apply torque compensation. The calculated torque compensation amount ΔT is superimposed on the target torque of the outer wheel and precisely output through the motor controller. For example, when turning left, the right wheel is the outer wheel, and the system automatically increases the torque output of the right front wheel and the right rear wheel.

[0049] Step 5: Closed-loop feedback control. The electronic differential continuously monitors the speed difference between the inner and outer wheels. When the speed difference returns to within ±2% of the theoretical value, the compensation amount is gradually reduced to avoid overcompensation that could lead to vehicle instability.

[0050] The system architecture of this invention adopts a layered modular design, mainly including the following four layers: the first layer is the perception layer, responsible for collecting environmental information and vehicle status information; the second layer is the decision control layer, responsible for mode selection, path planning, and control command generation; the third layer is the execution layer, responsible for the specific execution of steering, driving, and braking; and the fourth layer is the human-machine interaction module, responsible for function activation and status display. From a data flow perspective, the raw data collected by the perception layer is preprocessed and then transmitted to the decision control layer. The control commands generated by the decision control layer are transmitted to the execution layer via the CAN bus or FlexRay bus. The execution layer feeds back the execution results to the decision control layer, forming a closed-loop control.

[0051] The key technical aspects of this invention mainly include four aspects: First, multi-source sensor fusion technology, which achieves high-precision environmental perception through steps such as spatiotemporal synchronization, data association, and state estimation; second, SINS and dual 2D-LDV tightly coupled positioning technology, which achieves accurate positioning in satellite-free environments through LOF anomaly detection and adaptive filtering; third, mode selection technology based on large SUV parameters, which achieves intelligent mode switching by establishing a scene determination matrix; and finally, adaptive torque distribution technology, which is the core innovation of this invention, which achieves stable control under low-adhesion road surface conditions by establishing a torque optimization model that includes center of gravity height compensation.

[0052] like Figure 3 As shown, the four-source fusion algorithm of this invention adopts an adaptive weight adjustment mechanism. Its core idea is to dynamically adjust the fusion weights ω1-ω4 according to the real-time confidence level of each sensor and environmental conditions to ensure optimal perception performance in different scenarios. The fusion formula is D_fusion = ω1D_radar + ω2D_vision + ω3D_LDV + ω4D_SINS. D_fusion represents the fused distance. D_radar, D_vision, D_LDV, and D_SINS represent the distance data measured by millimeter-wave radar, machine vision, LDV lidar, and SINS inertial navigation system, respectively. ω1, ω2, ω3, and ω4 are the corresponding adaptive weight coefficients, ω1+ω2+ω3+ω4=1, and each weight coefficient varies between 0 and 1.

[0053] Confidence-based weight calculation model The weight of each sensor consists of two parts: the base weight and the dynamically adjusted weight, i.e., ω_i = ω_i_base + Δω_i, where ω_i_base is the reference weight of the sensor and Δω_i is the dynamically adjusted amount calculated based on the real-time state. The reference weight is determined according to the reliability statistics of the sensor in a typical underground garage scenario and is initially set as ω1_base = 0.4 (radar), ω2_base = 0.2 (vision), ω3_base = 0.25 (LDV), ω4_base = 0.15 (SINS). The dynamically adjusted amount Δω_i is calculated based on a multi-factor evaluation model that comprehensively considers factors such as the sensor's own state, environmental conditions, and historical performance.

[0054] Dynamic adjustment strategy for the radar weight ω1 The weight ω1 of the 77GHz 4D millimeter-wave radar is mainly affected by three factors: the signal quality factor Q_radar, the multipath interference factor M_radar, and the detection distance factor D_radar. The signal quality factor is calculated by analyzing the signal-to-noise ratio (SNR) of the radar echo. When SNR > 15dB, Q_radar = 1.0; when 10dB < SNR ≤ 15dB, Q_radar = 0.8; when SNR ≤ 10dB, Q_radar = 0.5. The multipath interference factor is determined by detecting the number of false targets in the radar echo. When the number of false targets ≤ 2, M_radar = 1.0; when 3 ≤ the number of false targets ≤ 5, M_radar = 0.7; when the number of false targets > 5, M_radar = 0.4. The detection distance factor is adjusted according to the distance of the target from the radar. For close-range targets (<3 meters), D_radar = 1.0; for medium-range targets (3 - 6 meters), D_radar = 0.9; for long-range targets (>6 meters), D_radar = 0.7. The comprehensive weight adjustment amount is Δω1 = 0.15×(Q_radar×M_radar×D_radar - 1), which means that when the radar is in good working condition, its weight can be increased from the reference value of 0.4 to 0.55; when the radar is severely interfered, its weight can be reduced to 0.25.

[0055] Dynamic adjustment strategy for the vision weight ω2 The weight ω2 of the binocular surround-view camera is mainly affected by the illumination condition factor L_vision, the image sharpness factor C_vision, and the target recognition confidence factor R_vision. The illumination conditions in an underground parking garage are complex and variable. The illumination condition factor is evaluated using the average image brightness value. When the average brightness is 80-120 candela / m², L_vision = 1.0 (ideal lighting); when the average brightness is 50-80 or 120-150 candela / m², L_vision = 0.7 (acceptable lighting); and when the average brightness is <50 or >150 candela / m², L_vision = 0.3 (insufficient lighting or overexposure). The image sharpness factor is evaluated using the image edge gradient intensity, calculated using the Laplacian operator. When the gradient magnitude > 500, C_vision = 1.0 (sharp); when 200 < gradient magnitude ≤ 500, C_vision = 0.6 (moderate); and when the gradient magnitude ≤ 200, C_vision = 0.3 (blurred). The target recognition confidence factor is based on the softmax probability output by the deep learning network. When confidence > 0.9, R_vision = 1.0; when 0.7 < confidence ≤ 0.9, R_vision = 0.7; and when confidence ≤ 0.7, R_vision = 0.4. The dynamic adjustment range of the visual weights is relatively large, Δω2 = 0.2 × (L_vision × C_vision × R_vision - 1). Under ideal conditions, ω2 can be increased to 0.4, while under adverse conditions it can drop to 0.0 (i.e., completely disabling visual perception).

[0056] Dynamic adjustment strategy for LDV weight ω3 The weight ω3 of the dual 2D-LDV laser speedometer is mainly affected by the measurement accuracy factor A_LDV, the speed range factor S_LDV, and the data consistency factor U_LDV. The measurement accuracy factor is evaluated using the standard deviation of multiple LDV measurements. When the standard deviation is <0.05 km / h, A_LDV = 1.0; when 0.05 ≤ standard deviation < 0.1 km / h, A_LDV = 0.8; and when the standard deviation is ≥0.1 km / h, A_LDV = 0.5. The speed range factor is determined based on the current vehicle speed. LDV has the highest accuracy at low speeds: when the vehicle speed is <5 km / h, S_LDV = 1.0; when 5 ≤ vehicle speed < 10 km / h, S_LDV = 0.9; and when the vehicle speed is ≥10 km / h, S_LDV = 0.7. The data consistency factor is assessed by evaluating the difference between the two LDV measurements. When the difference between the two LDV measurements is <5%, U_LDV = 1.0; when the difference is between 5% and 10%, U_LDV = 0.7; and when the difference is >10%, U_LDV = 0.4. The dynamic adjustment of the LDV weight is Δω3 = 0.1 × (A_LDV × S_LDV × U_LDV - 1), with a weight adjustment range of 0.15–0.35.

[0057] Dynamic adjustment strategy of SINS weight ω4 The weight ω4 of the SINS inertial navigation system is mainly affected by the cumulative error factor E_SINS, the motion state factor K_SINS, and the external reference factor G_SINS. SINS uses high-precision gyroscopes and accelerometers, but accumulated errors occur after prolonged operation. The cumulative error factor is determined based on the time since the last zero-speed calibration: E_SINS = 1.0 when calibration time < 30 seconds, E_SINS = 0.8 when 30 seconds ≤ calibration time < 60 seconds, and E_SINS = 0.5 when calibration time ≥ 60 seconds. The motion state factor is determined based on the vehicle's motion state: K_SINS = 1.0 when the vehicle is moving at a constant or low speed, K_SINS = 0.8 when the vehicle is accelerating or decelerating, and K_SINS = 0.6 when the vehicle is making a sharp turn. The external reference factor is determined based on whether other sensors provide position reference: G_SINS = 0.7 when radar or vision provides a reliable position reference (reducing the SINS weight to avoid conflicts), and G_SINS = 1.3 when there is no external reference (increasing the SINS weight). The dynamic adjustment of the SINS weight is Δω4 = 0.1×(E_SINS×K_SINS×G_SINS - 1), and the weight adjustment range is 0.05-0.25.

[0058] Weight normalization and smoothing Since the dynamic adjustment of the weights of each sensor is performed independently, the total weight after adjustment may not be equal to 1. Therefore, normalization is required: ω_i_normalized = ω_i / (ω1+ω2+ω3+ω4). Furthermore, to avoid abrupt changes in the fusion result caused by sudden weight shifts, this invention employs a first-order hysteresis filter to smooth the weights: ω_i_filtered(k) = 0.7×ω_i_filtered(k-1) + 0.3×ω_i_normalized(k), where k is the current time and k-1 is the previous time. This smoothing process ensures the gradual change characteristics of the weights and avoids jitter in the fusion result caused by instantaneous changes in sensor states.

[0059] Example of weight configuration in typical scenarios In a typical underground parking garage U-turn scenario, the weighting configuration dynamically changes with environmental conditions. For example, when a vehicle enters a well-lit main passageway of an underground parking garage, with good lighting (L_vision=1.0), clear images (C_vision=1.0), and no radar interference (M_radar=1.0), the weighting configuration is ω1=0.45, ω2=0.30, ω3=0.15, and ω4=0.10. The visual perception weighting is significantly improved, allowing full utilization of lane line recognition information. When the vehicle enters a dimly lit corner area, with insufficient lighting (L_vision=0.3) and blurry images (C_vision=0.6), the weighting automatically adjusts to ω1=0.55, ω2=0.05, ω3=0.25, and ω4=0.15. The visual perception weighting is significantly reduced, while the radar and LDV weightings are correspondingly increased. When the vehicle approaches the metal pillar, radar multipath interference becomes severe (M_radar=0.4). The weights are further adjusted to ω1=0.30, ω2=0.10, ω3=0.40, and ω4=0.20, with LDV weights becoming dominant, ensuring reliable perception results even when the radar is interfered with. This adaptive weight adjustment mechanism enables the system to maintain optimal perception performance in various complex environments.

[0060] Specific implementation steps of the LOF anomaly detection algorithm The LOF (Local Outlier Factor) algorithm is a density-based anomaly detection method. This invention applies it to anomaly detection in the output data of a dual 2D-LDV laser velocimeter. Compared with traditional statistical methods (such as the 3σ criterion), the LOF algorithm can better handle datasets with local density variations, making it particularly suitable for local anomaly patterns that may appear in LDV data in underground parking garage environments.

[0061] Step 1: Data Preprocessing and Feature Extraction The dual 2D-LDV laser speedometer outputs the longitudinal velocity v_x and lateral velocity v_y of a vehicle at a frequency of 50Hz. The raw data is first preprocessed. Preprocessing includes three sub-steps: first, noise reduction, using a median filter to eliminate impulse noise, with a window size of 5 sampling points; second, outlier removal, marking points with speed variations exceeding 2 km / h as suspicious outliers; and finally, coordinate transformation, converting the LDV-measured velocity from the sensor coordinate system to the vehicle coordinate system using the formula v_vehicle = R_sensor_to_vehicle × v_LDV, where R_sensor_to_vehicle is the rotation matrix determined by the installation angle. The preprocessed data sequence is denoted as {(v_x(i), v_y(i)) | i=1,2,…,N}, where N is the number of data points within the sliding window. In this invention, N=100, corresponding to a data length of 2 seconds.

[0062] Step 2: k-distance neighborhood construction For each data point p_i in the dataset, calculate its Euclidean distance to all other data points d(p_i, p_j) = sqrt((v_x(i)-v_x(j))² + (v_y(i)-v_y(j))²). Sort the distances in ascending order to obtain the distance sequence d_1 ≤ d_2 ≤ … ≤ d_{N-1}. The k-th smallest distance is called the k-distance of point p_i, denoted as k-distance(p_i). In this invention, the neighborhood parameter k is set to 20. This value is based on the following considerations: if k is too small (e.g., k<10), the neighborhood range will be too narrow, making it easy to misjudge normal fluctuations as anomalies; if k is too large (e.g., k>30), the neighborhood range will be too wide, reducing the sensitivity to local anomalies. Through extensive real-vehicle testing, it was found that k=20 achieves the best balance between anomaly detection rate and false alarm rate.

[0063] Construction process of k-distance neighborhood N_k(p_i) After determining k-distance(p_i), it is necessary to construct the k-distance neighborhood N_k(p_i) of point p_i. In other words, the construction of N_k(p_i) is a neighborhood search process based on the radius threshold of k-distance(p_i). Specifically, N_k(p_i) is defined as the set of all data points in the dataset whose distance to point p_i does not exceed k-distance(p_i), i.e., N_k(p_i). p_j∈ D | d(p_i,p_j)≤ k-distance(p_i) , where D is the entire dataset.

[0064] From a mathematical perspective, the construction of N_k(p_i) requires strict distance criteria. First, a multidimensional hypersphere neighborhood is constructed with point p_i as the center and k-distance(p_i) as the radius. Second, all other data points p_j (j≠i) in the dataset are traversed, and the Euclidean distance d(p_i,p_j) from each point p_j to the center point p_i is calculated. Then, all data points p_j satisfying d(p_i,p_j)≤k-distance(p_i) are included in the neighborhood N_k(p_i). It's important to note that according to the definition of k-distance, N_k(p_i) should contain at least k data points (i.e., the k nearest neighbors). However, in practical applications, since multiple data points may have a distance to p_i that is exactly equal to k-distance(p_i), the number of data points in N_k(p_i) may be greater than k.

[0065] From an engineering implementation perspective, the construction process of N_k(p_i) employs an efficient neighborhood search algorithm. Due to the limited computing resources of in-vehicle embedded systems, this invention uses a fast nearest neighbor search algorithm based on a KD-tree, reducing the time complexity of the neighborhood search from O(N) to O(logN). The specific implementation steps are as follows: First, a KD-tree index structure is constructed based on the N data points within the sliding window; then, using point p_i as the query point, a k-nearest neighbor search is performed in the KD-tree to obtain the k nearest data points and their corresponding distance values; finally, these k data points, along with all other data points whose distance is equal to k-distance(p_i), are combined to form N_k(p_i).

[0066] From a data structure perspective, N_k(p_i) is stored using a dynamic linked list structure to accommodate changes in the neighborhood size. Each node in the linked list stores three key pieces of information: the index of its neighbor p_j, the distance value d(p_i, p_j), and the reachability_distance(p_i, p_j). This data structure design allows subsequent reachability distance calculations and local reachability density calculations to be performed directly by traversing the linked list, avoiding redundant distance calculations and significantly improving the algorithm's execution efficiency.

[0067] From a physical perspective, N_k(p_i) represents the local data distribution characteristics around point p_i. Under normal driving conditions, the speed data output by LDV exhibits a clustered distribution in the two-dimensional speed space. At this time, the data points in N_k(p_i) are relatively evenly distributed in all directions, and the value of k-distance(p_i) is small. When abnormal data occurs, the abnormal points deviate from the normal data clustering area, resulting in an uneven distribution of data points in N_k(p_i), and the value of k-distance(p_i) increases significantly. For example, when the vehicle is traveling at a constant speed, the speed data points are concentrated in... Nearby, the k-distance(p_i) of a normal point might be only 0.3 km / h, and its N_k(p_i) contains 20 neighboring points within 0.3 km / h; however, when vibration disturbances cause the speed to jump to... At that time, the k-distance(p_i) of the outlier may reach more than 2.5 km / h, and its neighboring points in N_k(p_i) are sparsely distributed and may contain multiple boundary points with a distance of exactly 2.5 km / h.

[0068] Step 3: Calculate reachable distance For any two data points p_i and p_j, the reachability distance of point p_j with respect to point p_i is defined as: reachability_distance(p_i, p_j) = max{k-distance(p_i), d(p_i, p_j)}. This definition ensures that when p_j is within the k-distance neighborhood of p_i, the reachability distance is equal to k-distance(p_i) (i.e., the neighborhood radius), and when p_j is outside the k-distance neighborhood of p_i, the reachability distance is equal to the actual distance d(p_i, p_j). The concept of reachability distance enables the LOF algorithm to distinguish between normal and abnormal data points: normal data points are surrounded by a higher density of data points, resulting in a smaller k-distance; abnormal data points are surrounded by a lower density of data points, resulting in a larger k-distance.

[0069] Step 4: Calculation of Locally Reachable Density The Local Reachability Density (LRD) of a data point p_i is defined as the reciprocal of the average reachability distance of all data points within its k-neighborhood: LRD(p_i) = 1 / (Σ_{p_j∈N_k(p_i)}reachability_distance(p_i, p_j) / k), where N_k(p_i) is the k-distance neighborhood of p_i. Local reachability density reflects the density of data around a data point; regions with high density have large LRD values, and regions with low density have small LRD values. Under normal driving conditions, the speed data output by LDV changes smoothly, and the data points are densely distributed, with LRD values ​​typically ranging from 2.5 to 4.0 km⁻¹. When abnormal data occurs, the abnormal points deviate from the normal data distribution, and their LRD values ​​decrease significantly, typically below 1.5 km⁻¹.

[0070] Step 5: Calculation of Local Anomaly Factors The Local Outlier Factor (LOF) of a data point p_i is defined as the ratio of the average local reachability density of all data points in its k-neighborhood to its own local reachability density: LOF(p_i) = (Σ_{p_j∈N_k(p_i)} LRD(p_j) / k) / LRD(p_i). The LOF value reflects the degree of anomalousness of a data point relative to its neighborhood: when LOF≈1, it indicates that the density of the point is similar to that of other points in its neighborhood, and it is a normal data point; when LOF>1, it indicates that the density of the point is lower than the average density of its neighborhood, and it may be an anomalous data point; the larger the LOF value, the higher the degree of anomalousness. This invention sets the anomaly threshold to LOF_threshold=1.5. When LOF(p_i)>1.5, the data point p_i is determined to be anomalous. This threshold selection is based on ROC curve analysis. On a real-vehicle test dataset, when LOF_threshold=1.5, the true positive rate reaches 92%, and the false positive rate is controlled below 5%.

[0071] Step Six: Strategies for Handling Abnormal Data Upon detecting outlier data points, the system employs a tiered processing strategy. For mild outliers with LOF values ​​between 1.5 and 2.0, linear interpolation is used for correction, employing two normal data points before and after the outlier: p_corrected(i) = 0.5 × p(i-1) + 0.5 × p(i+1). For moderate outliers with LOF values ​​between 2.0 and 3.0, a quadratic polynomial fitting method is used for correction, fitting a polynomial using five normal data points before and after the outlier, and then calculating the correction value for the outlier's location. For severe outliers with LOF values ​​> 3.0, the data point is directly removed and not included in subsequent fusion calculations. Corrected data points are marked as "corrected," and their confidence level is reduced in subsequent weight calculations. Specifically, the LDV measurement accuracy factor A_LDV at that moment is reduced to 0.6, thereby correspondingly reducing the LDV weight.

[0072] Step 7: Real-time performance optimization To ensure the real-time performance of the LOF algorithm in automotive embedded systems, this invention employs three optimization measures. First, a sliding window strategy is used, performing LOF calculations only on the most recent N=100 data points to avoid the cumulative computational load caused by calculating the entire dataset. Second, an incremental update strategy is adopted; when a new data point arrives, only the affected neighborhood relationships are updated, rather than recalculating the distance matrix for all data points, reducing the computational time complexity from O(N²) to O(kN). Third, fixed-point arithmetic is used instead of floating-point arithmetic; the speed data is amplified 100 times and converted to integers for distance calculation, improving computational speed while maintaining accuracy. The optimized LOF algorithm achieves a single execution time of less than 5ms on an automotive processor (1.5GHz), meeting the real-time requirement of a 50Hz data update frequency.

[0073] Engineering Validation of the LOF Algorithm To verify the effectiveness of the LOF algorithm, this invention designed a specific comparative test. Test scenarios included: normal driving (stable LDV data), vibration interference (LDV data jitter caused by a vehicle passing over a speed bump), and multipath effects (LDV laser beam reflection from a wall causing data jumps). Test results showed that in the normal driving scenario, the false alarm rate of the LOF algorithm was only 3.2%, lower than the 8.7% of the traditional 3σ criterion; in the vibration interference scenario, the anomaly detection rate of the LOF algorithm reached 94.5%, higher than the 78.3% of the 3σ criterion; in the multipath effect scenario, the LOF algorithm could detect anomalies and trigger a correction mechanism within 20ms, while traditional methods required 50-80ms. The test data fully demonstrates the superiority of the LOF algorithm in LDV data anomaly detection.

[0074] Regarding spatial adaptability, this invention optimizes the four-wheel steering coordination strategy for the steering characteristics of large SUVs and adds a crab-walking assist mode. This mode achieves lateral movement of the vehicle by steering the front and rear wheels in the same direction by 6-8 degrees, reducing the space required for U-turns by more than 30%. Furthermore, this invention establishes a dedicated parameter database based on the size parameters of large SUVs (length 5.0-5.3 meters, width 1.9-2.0 meters) and designs scenario-based mode selection logic according to the characteristics of underground parking garage scenarios (lane width, column spacing, ramp angle, etc.).

[0075] Specific methods for sensor spatiotemporal synchronization This invention innovatively solves the problem of traditional methods requiring periodic manual calibration. The method utilizes stable natural features such as columns and walls in an underground parking garage environment, and recursively updates extrinsic parameters through feature matching and extended Kalman filtering (EKF).

[0076] The specific implementation steps are as follows: First, extract stable features such as columns and walls from radar, vision, and LDV data; second, match the extracted features with a pre-built environmental map; then, calculate the deviation between the current external parameters and the calibration external parameters based on the matching results. The advantage of this online calibration mechanism is that it eliminates the need for a dedicated calibration field, utilizing the environmental characteristics during normal vehicle operation to complete the calibration, greatly improving the system's practicality and long-term stability. Real-world testing shows that after 100 hours of continuous operation, the external parameter error can still be controlled within ±2 cm.

[0077] Time synchronization methods Time synchronization employs a layered synchronization strategy combining hardware timestamps and software interpolation. The first layer is hardware timestamp synchronization, with the system equipped with a high-precision GPS timing module (receiving satellite signals on the ground) and a crystal clock (maintaining a time reference in the underground parking garage), achieving a clock accuracy of ±1μs. Each sensor's data is timestamped at the moment of acquisition, with the timestamp format being a Unix timestamp (millisecond level) plus a microsecond-level counter value. The data update cycle for the 77GHz 4D millimeter-wave radar is 20ms (50Hz), for the binocular surround-view camera it is 33ms (30Hz), for the dual 2D-LDV it is 20ms (50Hz), and for the SINS it is 10ms (100Hz).

[0078] The second layer is timestamp-based data alignment. Because the arrival times of data from different sensors to the central control unit vary, radar data has a delay of approximately 5ms, camera data approximately 15ms (including image processing time), LDV data approximately 3ms, and SINS data approximately 2ms. To eliminate these delay differences, the central control unit maintains a circular buffer to store the sensor data by timestamp. When fusion calculations are needed, the time reference t_SINS of the SINS data is used as a reference to extract the data from each sensor within the time window [t_SINS-Δt, t_SINS+Δt], where Δt = 10ms. For sensors without precisely matching data within the time window, linear interpolation is used for time alignment. For example, when t_SINS=1000ms, the nearest radar data points are t_radar1=995ms and t_radar2=1015ms. Then, the radar data at time t_SINS is calculated by interpolation: D_radar(1000) = D_radar(995) + (D_radar(1015)-D_radar(995))×(1000-995) / (1015-995).

[0079] The third layer is prediction compensation synchronization. For sensors with large delays (such as cameras, with a delay of about 15ms), simple interpolation may still produce large errors in high-speed motion scenarios. Therefore, a motion model-based prediction compensation method is adopted: utilizing the high-frequency (100Hz) attitude and velocity information provided by SINS, a vehicle motion model is established: x(t+Δt) = x(t) + v(t)·Δt + 0.5·a(t)·Δt², where x is the position, v is the velocity, and a is the acceleration. When the camera data delay is Δt=15ms, this motion model is used to predict the position of the obstacle detected by the camera from the detection time t_capture to the current time t_current, thereby compensating for the error caused by the time delay. Experimental results show that this prediction compensation method achieves a position prediction error of less than 2 cm at a vehicle speed of 10km / h and a delay of 15ms.

[0080] Spatial synchronization methods Spatial synchronization employs a strategy combining unified coordinate transformation and extrinsic parameter calibration. First, a unified vehicle coordinate system is established, with the origin at the rear axle center, the X-axis pointing forward, the Y-axis pointing to the left side of the vehicle, and the Z-axis pointing vertically upward. Data from each sensor needs to be transformed from its own coordinate system to the vehicle coordinate system using the formula: P_vehicle = R_sensor × P_sensor + T_sensor, where R_sensor is a 3×3 rotation matrix and T_sensor is a 3×1 translation vector. These two parameters are collectively referred to as sensor extrinsic parameters.

[0081] Sensor extrinsic parameters are obtained through high-precision calibration. The calibration process utilizes a dedicated calibration field, which includes multiple calibration boards with known precise coordinates (positional accuracy ±1mm). Taking a binocular surround-view camera as an example, during calibration, the vehicle is positioned at the center of the calibration field, and the camera captures images of the calibration boards. The camera extrinsic parameters are then solved using the Perspective-n-Point (PnP) algorithm. To improve calibration accuracy, each sensor captures images of the calibration boards at at least 15-20 different locations. A nonlinear least squares method is used to optimize the extrinsic parameters, ensuring a reprojection error of less than 0.5 pixels. The calibration of the 77GHz 4D millimeter-wave radar uses a corner reflector (radar cross-section RCS > 10㎡). By measuring the distance and angle from the radar to the corner reflector and combining this with the precise position of the corner reflector, the radar extrinsic parameters are solved. The calibration of the dual 2D-LDV uses a laser tracker to measure the offset of the LDV's installation position relative to the vehicle coordinate system, achieving a calibration accuracy of ±0.5mm. The calibration of SINS is achieved through a turntable. The SINS is mounted on the turntable, and the turntable is rotated at different angles to record the attitude changes output by the SINS, thereby calibrating the relative relationship between the SINS coordinate system and the vehicle coordinate system.

[0082] During real-time operation, the data from each sensor is transformed into the vehicle coordinate system through extrinsic parameter transformation. For example, if a 77GHz 4D millimeter-wave radar detects polar coordinate data of an obstacle ahead (distance r = 5.2m, azimuth θ = 15°, elevation φ = 3°), it is first converted to the radar Cartesian coordinate system: x_radar = r·cos(φ)·cos(θ) = 5.2×cos(3°)×cos(15°) ≈ 5.02m, y_radar = r·cos(φ)·sin(θ) = 5.2×cos(3°)×sin(15°) ≈ 1.34m, z_radar = r·sin(φ) = 5.2×sin(3°) ≈ 0.27m. Then, it is transformed into the vehicle coordinate system through extrinsic parameter transformation: P_vehicle = R_radar × [x_radar, y_radar, z_radar]ᵀ + T_radar. For a radar installed in the center of the front bumper of a vehicle, its extrinsic parameters are approximately R_radar=I (identity matrix) and T_radar=[2.5, 0,0.5]ᵀ (unit: meters). Therefore, the position of the obstacle in the vehicle coordinate system is x_vehicle = 5.02+2.5 = 7.52m, y_vehicle = 1.34+0 = 1.34m, z_vehicle = 0.27+0.5 = 0.77m.

[0083] Dynamic compensation and online calibration Considering that vehicles may experience body deformation during operation (such as suspension compression due to load changes) and slight displacement of sensor mounting positions (loosening due to long-term vibration), this invention also incorporates dynamic compensation and online calibration mechanisms. Dynamic compensation primarily addresses body deformation by using SINS to monitor the vehicle's pitch and roll angles in real time. When the pitch angle change exceeds 1°, the height component in the sensor's extrinsic parameters is automatically adjusted to compensate for the height change caused by suspension compression. Online calibration employs a natural feature matching method. When the vehicle passes over an environment with known features (such as pillars or walls), the feature positions detected by each sensor are compared with the known positions to calculate the extrinsic parameter error. When the error exceeds a threshold, an extrinsic parameter update is triggered. This online calibration mechanism ensures the long-term accuracy of the sensor's extrinsic parameters. Real-world testing shows that after 100 hours of continuous operation, the extrinsic parameter error remains within ±2 cm.

[0084] Performance metrics of spatiotemporal synchronization After the aforementioned spatiotemporal synchronization processing, the time synchronization accuracy of the data from each sensor reached ±5ms, and the spatial synchronization accuracy reached ±3cm (for obstacles less than 10 meters away). This high-precision spatiotemporal synchronization lays a solid foundation for subsequent data fusion, effectively avoiding fusion errors caused by spatiotemporal inconsistencies. For example, in testing, it was found that without spatiotemporal synchronization, the positions of the same obstacle detected by radar and vision differed by up to 0.5 meters, causing the fusion algorithm to misidentify it as two obstacles. After spatiotemporal synchronization, the positional deviation was reduced to within 0.03 meters, and the fusion algorithm was able to correctly identify it as the same obstacle, increasing the fusion accuracy from 85% to 99%.

[0085] Another embodiment, application of crab-walking assist mode in narrow passages. This embodiment addresses a scenario where a large new energy SUV makes a U-turn in a narrow passageway of an underground parking garage. The specific conditions for this scenario are: a lane width of 4.5 meters (only 2.3 times the vehicle width), an epoxy floor surface, a road adhesion coefficient μ=0.6 in a dry environment, and vehicle parameters identical to those in Embodiment 1. This scenario presents a much narrower space, making a single U-turn difficult with traditional four-wheel steering; therefore, a crab-assisted mode is required.

[0086] The technical configuration of this embodiment is basically the same as that of Embodiment 1. The difference is that the mode selection module adds the judgment condition for the crab assist mode, and the steering execution unit supports the function of steering in the same direction for the front and rear wheels. In terms of human-machine interaction, when the system determines that the crab assist mode needs to be used, the instrument cluster will display a special crab mode icon and prompt the driver that "crab mode has been activated, please pay attention to lateral movement".

[0087] Implementation steps, such as Figure 5 As shown, it includes: Steps one through three are similar to those in Example 1, except for the mode selection result in step three. Since the lane width is 4.5 meters (2.3 times the vehicle width) and the adhesion coefficient μ=0.6≥0.4, the system determines to use the "crab assist mode".

[0088] Step 4: Crab-like Path Planning. The path planning module generates a special crab-like U-turn trajectory. This trajectory first achieves lateral movement of the vehicle by steering the front and rear wheels in the same direction, moving the vehicle from one side of the lane to the other. Then, the remaining U-turn is completed using traditional four-wheel steering. Specifically, the plan is as follows: first, control the front and rear wheels to turn 7 degrees to the left in the same direction, the vehicle moves laterally to the left by about 1.2 meters, then resumes straight driving, and then performs a standard four-wheel steering U-turn.

[0089] Step 5: Crab-like Steering Control. The EPS system first controls the front and rear wheels to simultaneously turn 7 degrees to the left (same direction steering). The four independent motor drives the vehicle forward at a low speed (3km / h), causing the vehicle to move laterally in a manner similar to a crab walking. During the lateral movement, the 4D millimeter-wave radar monitors the distance to the left pillar in real time. When the distance decreases to 0.4 meters, the system controls the front and rear wheels to return to center, ending the lateral movement phase.

[0090] Step Six: Four-wheel steering U-turn. After the vehicle completes lateral movement, the system switches to four-wheel steering mode, turning the front wheels 12 degrees to the left and the rear wheels 10 degrees to the right to execute a standard U-turn. Since the additional lateral space gained through the crab mode allows for a more effortless four-wheel steering U-turn.

[0091] Step 7: Exit Function. After the U-turn is completed, the system prompts the driver to take over and exits the assisted mode after detecting the driver's operation.

[0092] Regarding the implementation details of the crab-like steering assist mode, this invention employs a special steering angle control strategy. The steering angle of the front and rear wheels in the same direction is set at 6-8 degrees. This is based on the following considerations: if the angle is too small (less than 6 degrees), the lateral translation effect will be insignificant, and the advantages of the crab-like steering mode cannot be effectively utilized; if the angle is too large (greater than 8 degrees), it will lead to accelerated tire wear and make it easy to scrape against side obstacles during lateral translation. During lateral translation, the system strictly controls the vehicle speed below 5 km / h to ensure the safety and controllability of lateral movement.

[0093] In terms of safety, the Crab Assist mode adds extra safety monitoring logic. When the system detects a lateral distance of less than 0.5 meters, it automatically reduces the lateral movement speed to below 2 km / h; when the distance is less than 0.3 meters, it immediately stops the lateral movement and alerts the driver. Furthermore, the system monitors the road surface adhesion coefficient; when μ < 0.4, Crab Assist mode is disabled because lateral movement on low-adhesion surfaces can easily lead to vehicle instability.

[0094] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for adaptive assisted U-turn of large new energy vehicles based on multiple sensors, characterized in that, Includes the following steps: S01: Acquire millimeter-wave radar data, binocular camera data, speedometer data, and inertial navigation system data; fuse the data from each sensor to obtain information about the vehicle's surrounding environment and vehicle status. S02: Based on the acquired environmental and vehicle status information, and combined with the preset large SUV parameter database and scene feature database, select the U-turn mode and generate the U-turn trajectory; S03: Establish a torque optimization model that includes center of gravity height compensation. By monitoring the road surface adhesion coefficient, vehicle center of gravity height, and axle load distribution in real time, dynamically adjust the torque output of each drive wheel. Control the output of each wheel through the four-wheel motor drive according to the calculated torque distribution ratio, and control the four-wheel steering to perform a U-turn operation.

2. The adaptive assisted U-turn method for large new energy vehicles based on multiple sensors according to claim 1, characterized in that, The formula for fusing the data from each sensor in step S01 is as follows: D_fusion = ω1D_radar + ω2D_vision + ω3D_LDV + ω4D_SINS Where D_fusion represents the fusion distance, D_radar, D_vision, D_LDV, and D_SINS represent the distance data measured by millimeter-wave radar, machine vision, LDV lidar, and SINS inertial navigation system, respectively; ω1, ω2, ω3, and ω4 are the corresponding fusion weights, and ω1+ω2+ω3+ω4=1. The fusion weights are dynamically adjusted based on the real-time confidence levels of each sensor and environmental conditions. The fusion weights consist of two parts: a base weight and a dynamically adjusted weight, i.e., ωi = ωi_base + Δωi, where ωi_base is the base weight of the sensor and Δωi is the dynamic adjustment amount calculated based on the real-time status.

3. The adaptive assisted U-turn method for large new energy vehicles based on multiple sensors according to claim 2, characterized in that, The strategies for adjusting the fusion weights include: The dynamic adjustment amount Δω1 of the weight ω1 of the millimeter-wave radar is: Δω1 = 0.15×(Q_radar×M_radar×D_radar - 1) Where Q_radar is the signal quality factor, M_radar is the multipath interference factor, and D_radar is the detection range factor; The dynamic adjustment amount Δω2 of the weight ω2 of the binocular camera is: Δω2 = 0.2×(L_vision×C_vision×R_vision - 1) Where L_vision is the illumination condition factor, C_vision is the image sharpness factor, and R_vision is the target recognition confidence factor; The dynamic adjustment amount Δω3 of the weight ω3 of the laser velocimeter is: Δω3 = 0.1×(A_LDV×S_LDV×U_LDV - 1) Where A_LDV is the measurement accuracy factor, S_LDV is the speed range factor, and U_LDV is the data consistency factor; The dynamic adjustment amount Δω4 of the weight ω4 of the inertial navigation system is: Δω4 = 0.1×(E_SINS×K_SINS×G_SINS - 1) Where E_SINS is the cumulative error factor, K_SINS is the motion state factor, and G_SINS is the external reference factor.

4. The adaptive assisted U-turn method for large new energy vehicles based on multiple sensors according to claim 1, characterized in that, Step S01 also includes: By tightly coupling and integrating velocimeter data and inertial navigation system data, and employing a local anomaly factor algorithm to perform real-time anomaly detection on the laser velocimeter data, accurate positioning in satellite-free environments is achieved. The local anomaly factor algorithm includes: The laser speedometer data is preprocessed, including noise reduction, outlier removal, and conversion of the sensor coordinate system to the vehicle coordinate system. For each data point p_i in the dataset, calculate its Euclidean distance to all other data points, sort the distances in ascending order, and the k-th smallest distance is called the k-distance of point p_i, denoted as k-distance(p_i). Construct the k-distance neighborhood N_k(p_i) of p_i. For any two data points p_i and p_j, the reachability distance of point p_j with respect to point p_i is defined as: reachability_distance(p_i, p_j) = max{k-distance(p_i), d(p_i, p_j)}; The local reachability density of a data point p_i is defined as the reciprocal of the average reachability distance of all data points in its k-neighborhood: LRD(p_i)=1 / (Σ_{p_j∈N_k(p_i)} reachability_distance(p_i, p_j) / k); The Local Outlier Factor (LOF) of a data point p_i is defined as the ratio of the average local reachability of all data points in its k-neighborhood to its own local reachability: LOF(p_i) = (Σ_{p_j∈N_k(p_i)} LRD(p_j) / k) / LRD(p_i). When the LOF is within the threshold range, the point is considered a normal data point; when the LOF is greater than the threshold, it is considered an outlier data point. When abnormal data points are detected, the system adopts a hierarchical processing strategy: for mild anomalies with LOF values ​​between 1.5 and 2.0, linear interpolation is used for correction; for moderate anomalies with LOF values ​​between 2.0 and 3.0, quadratic polynomial fitting is used for correction; for severe anomalies with LOF values ​​> 3.0, the data point is directly removed; and the weight of the speedometer data at that moment is reduced in subsequent fusion calculations.

5. The adaptive assisted U-turn method for large new energy vehicles based on multiple sensors according to claim 1, characterized in that, The fusion of data from various sensors also includes spatiotemporal synchronization of the data, wherein the spatiotemporal synchronization includes: Time synchronization: Set hardware timestamps for data from each sensor to ensure hardware timestamp synchronization; During fusion computing, the time reference of the inertial navigation system data is used to extract the data of each sensor within the time window from the buffer; for sensors that do not have matching data within the time window, a linear interpolation method is used for time alignment. Using attitude and velocity information provided by the inertial navigation system, a vehicle motion model is established. This motion model is then used to predict the position of obstacles detected by the camera from the detection time to the current time, achieving prediction compensation synchronization; and Spatial synchronization: By pre-calibrating or online-calibrating the sensor extrinsic parameters, the data in each sensor coordinate system are uniformly transformed to the vehicle coordinate system; Dynamic compensation and online calibration: The vehicle pitch and roll angles are monitored in real time through the inertial navigation system. When the pitch angle changes beyond a certain angle, the height component in the sensor extrinsic parameters is adjusted to compensate for the height change caused by suspension compression. The natural feature matching method is adopted. When the vehicle passes through an environment with known features, the feature positions detected by each sensor are compared with the known positions to calculate the extrinsic error. When the error exceeds the threshold, the extrinsic update is triggered.

6. The adaptive assisted U-turn method for large new energy vehicles based on multiple sensors according to claim 1, characterized in that, The torque optimization model that includes center of gravity height compensation is as follows: V_wheel = V_base + k1δ + k2(1-μ) + k3H Where V_wheel represents the actual rotational speed of the wheel, V_base represents the base speed calculated by the system, k1 is the steering compensation coefficient, k2 is the adhesion compensation coefficient, k3 is the center of gravity compensation coefficient, δ is the steering angle, μ is the road adhesion coefficient, and H is the vehicle center of gravity height.

7. The adaptive assisted U-turn method for large new energy vehicles based on multiple sensors according to claim 1, characterized in that, The dynamic adjustment of torque output for each drive wheel also includes a hierarchical optimization process, including: The first layer is torque limitation based on the adhesion coefficient. The maximum available torque T_max = μmg is calculated based on the road adhesion coefficient μ, where m is the vehicle mass and g is the gravitational acceleration. The second layer is axle load distribution based on center of gravity height. According to the vehicle's center of gravity height H and steering angle δ, the torque distribution ratio between the front and rear axles is dynamically adjusted, where the center of gravity height H is positively correlated with the proportion of torque distributed to the rear axle. The third layer is differential compensation based on steering angle. According to the steering angle δ, a closed-loop feedback control strategy is used to increase the torque compensation of the outer wheel so that its speed matches the theoretical speed difference.

8. The adaptive assisted U-turn method for large new energy vehicles based on multiple sensors according to claim 1, characterized in that, Based on the ratio of lane width to vehicle width and the road surface adhesion coefficient, a U-turn mode is selected from a preset mode determination matrix. The U-turn mode includes at least a four-wheel steering mode and a crab assist mode. When the crab assist mode is selected, a crab turn trajectory is generated, and two-stage crab steering control is performed. In the first stage, the front and rear wheels are controlled to turn in the same direction to a preset angle range, so that the vehicle can move laterally. In the second stage, after the lateral translation is completed, the system switches to four-wheel steering mode and controls the front and rear wheels to turn in opposite directions to complete the remaining turn.

9. A multi-sensor-based adaptive assisted U-turn system for large new energy vehicles, used to implement the multi-sensor-based adaptive assisted U-turn method for large new energy vehicles as described in any one of claims 1-8, characterized in that, include: The environmental perception module acquires data from millimeter-wave radar, binocular camera, speedometer, and inertial navigation system, and fuses the data from each sensor to obtain information about the vehicle's surrounding environment and vehicle status. The mode selection and path planning module selects the U-turn mode and generates the U-turn trajectory based on the acquired environmental information and vehicle status information, combined with the preset large SUV parameter database and scene feature library. The torque distribution and steering control module establishes a torque optimization model that includes center of gravity height compensation. By monitoring the road surface adhesion coefficient, vehicle center of gravity height, and axle load distribution in real time, it dynamically adjusts the torque output of each drive wheel. The four-wheel motor drive controls the output of each wheel according to the calculated torque distribution ratio, and controls the four-wheel steering to perform a U-turn operation.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, it implements the adaptive assisted U-turn method for large new energy vehicles based on multiple sensors as described in any one of claims 1-8.

Images