A multi-mode steering adaptive pre-adjustment method based on front scene recognition

Abstract

The application is suitable for the field of automobile control technology, and provides a multi-mode steering adaptive pre-adjustment method based on front scene recognition, comprising: constructing a two-dimensional mapping function library of each steering mode; collecting images, correcting distortion, inverse perspective transformation and semantic segmentation, identifying target scenes and calculating comprehensive confidence; after identifying the standard, calculating the minimum trigger distance according to the current vehicle speed, and triggering the switching when the remaining distance is less than or equal to the value; taking the ratio of the distance traveled by the vehicle to the total transition distance as the normalized progress, and weighting and mixing the mapping outputs of the current mode and the target mode through the mixed weight, to realize the smooth transition of the steering characteristics; after entering the scene, the authenticity is verified through the asymmetric frame number and the hysteresis threshold, and when the environment no longer meets the requirement, the original mode is reversely restored through the Sigmoid smoothing mechanism. The application realizes the early prediction and seamless switching of the steering mode, and improves the control stability and driving comfort in the low-speed complex scene.

Description

Technical Field

[0001] This invention belongs to the field of automotive control technology, and in particular relates to a multi-mode adaptive pre-adjustment method for steering based on forward scene recognition. Background Technology

[0002] With the rapid development of automotive intelligent technology, adaptive control of steering systems has become an important technological direction for improving driving experience and driving safety. Drivers have significantly different needs for steering response in different driving scenarios. For example, a stable steering feel is needed at high speeds, while a more precise and less prone to oversteering steering response is required in low-speed, complex scenarios such as narrow roads and sharp curves. Therefore, methods that can adaptively adjust steering modes according to driving scenarios have become a research hotspot.

[0003] In existing technologies, such as the Chinese invention patent application number 202510617128.0 entitled "An Adaptive Adjustment Method for Steering Mode Based on Accident Prevention," the method uses ultrasonic radar to detect the distance between the vehicle and obstacles, and switches from standard mode to precision mode after entering a narrow scene. This method achieves adaptive adjustment of the steering mode to the driving scene, but it has the following shortcomings: its mode switching relies on radar detection after the vehicle has entered the scene, which is an "in-scene response" mechanism and has a time delay; at the same time, it does not address the issue of smoothness in steering feel during the switching process, affecting the continuity and comfort of driving.

[0004] Another type of technology, such as the Chinese invention patent application number 201710010868.3, "Driving Mode Switching Control Method for Electric Power Steering System," achieves a smooth change in feel by adding a transition process when the user manually switches driving modes. However, this method also has limitations: its switching decision relies entirely on the user's manual operation and cannot automatically predict and complete mode adaptation in advance based on the road conditions ahead; the switching trigger timing is determined by the user and cannot be automatically completed before the vehicle enters complex road conditions; in addition, this method lacks a verification and error correction mechanism for the effectiveness of the switching decision.

[0005] Therefore, how to overcome the problems of delayed response, abrupt switching, and lack of automatic prediction and fault tolerance in the existing steering mode switching is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-mode adaptive pre-adjustment method for steering based on the recognition of the foreground scene, which aims to solve the problems mentioned in the background art.

[0007] The present invention is implemented as follows: a multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition includes the following steps:

[0008] Step S1: Collect the corresponding data of steering wheel torque and lateral acceleration of each preset steering mode of the vehicle under different vehicle speed conditions through real vehicle test. With steering wheel torque and vehicle speed as input variables and vehicle lateral acceleration as output variable, construct the two-dimensional mapping function corresponding to each steering mode and store it in the mode library.

[0009] Step S2: Acquire the original image through the vehicle-mounted front-view camera, and sequentially perform distortion correction, image enhancement and noise reduction, and inverse perspective transformation to obtain a bird's-eye view; perform semantic segmentation on the bird's-eye view to identify the target scene type;

[0010] Step S3: When the target scene is identified and the overall confidence level is not lower than the preset threshold, the minimum trigger distance is calculated based on the current vehicle speed. When the current remaining distance is not greater than the minimum trigger distance, a trigger signal is issued, the total transition distance is recorded, and the target mode is passed to the next step.

[0011] Step S4: After receiving the trigger signal, the ratio of the distance the vehicle has traveled since the trigger time to the total transition distance is used as the normalized progress. The normalized progress is used as the independent variable to generate a hybrid weight. In each control cycle, the two-dimensional mapping function corresponding to the current mode and the two-dimensional mapping function corresponding to the target mode are queried according to the current steering wheel torque and vehicle speed, respectively. The hybrid weight is used to perform weighted hybrid output to generate the desired vehicle lateral acceleration, so as to achieve a smooth transition from the current mode to the target mode.

[0012] Step S5: After the vehicle enters the target scene, it continuously performs multi-frame judgment. Through the design of asymmetric frame count and hysteresis threshold, it confirms in real time whether to maintain the current mode. When the environmental features no longer meet the target scene conditions, it triggers a reverse transition and gradually restores the steering mode to the original mode.

[0013] A further technical solution is that, in step S1, a typical low-speed working condition is selected as the evaluation condition; the typical low-speed working condition corresponds to a vehicle speed of 15 km / h, the vehicle is under standard load, and the road surface adhesion coefficient is greater than 0.8.

[0014] Steering wheel torque was collected for each steering mode under this typical low-speed condition. With lateral acceleration Corresponding data; in to Equal intervals are selected within the linear interval discrete data points ,in, For the first Steering wheel torque at each sampling point For the first Lateral acceleration at each sampling point; linear fitting using the least squares method is performed, and the slope of the resulting straight line is the steering sensitivity under this mode. :

[0015] ;

[0016] in, The number of sampling points. The average value of the steering wheel torque across all sampling points. This is the average value of the lateral acceleration at all sampling points.

[0017] Calculate all preset steering modes After the value, The smallest judgment is for Precision Mode, and the middle judgment is for Standard Mode; if the vehicle only provides two modes, the other mode is automatically designated as Standard Mode; the two-dimensional mapping function corresponding to each steering mode. Stored in the pattern library for use in step S4 for pattern switching and smooth weighting.

[0018] In a further technical solution, in step S2, semantic segmentation uses a lightweight network to perform pixel-level classification of the bird's-eye view, and outputs road surface, lane lines and preset obstacle categories. The preset obstacle categories include walls, guardrails, shoulders, barriers, water-filled barriers, cones and parking signs, thereby identifying narrow roads, sharp bends, parking lot entrances and construction areas.

[0019] A further technical solution is that, in step S2, the scene determination is based on a geometric threshold and multi-frame consistency, as follows:

[0020] Narrow path determination: Extract the passable width boundary based on the obstacle information output by semantic segmentation, calculate the passable width, and determine the narrow path if it is less than the width threshold for 10 consecutive frames;

[0021] Sharp curve detection: The curvature radius is fitted using the output lane line information, and a curve is detected when it is less than the curvature threshold for 8 consecutive frames.

[0022] Parking lot entrance determination: Determines the entrance based on the output features of the barrier gate, octagonal parking sign, or the "P" character on the road surface. When the above features appear within a preset distance in front of the vehicle and the same target is detected for 5 consecutive frames.

[0023] Construction area determination: Detect water barriers and / or cones, and determine the construction area when the detection density exceeds a preset threshold for 5 consecutive frames;

[0024] Overall confidence level The calculation formula is:

[0025] ;

[0026] in, Image quality confidence level based on image contrast and sharpness assessment; The semantic segmentation confidence score corresponds to the average pixel classification probability within the target scene region. The temporal consistency confidence score is based on the stability of multi-frame recognition. , and For the weighting coefficients, satisfying ;when If the value is below a preset threshold, it is marked as uncertain, and the system will subsequently revert to the real-time radar switching logic.

[0027] The remaining distance measurement employs a differentiated strategy: visual ranging is primarily used in narrow roads and sharp bends; for parking lot entrances and construction zones where barriers and cones are present, a combination of visual and radar ranging is used, with the minimum value of the two being output as the current remaining distance. .

[0028] A further technical solution, the specific steps of step S3 are as follows:

[0029] The scene type output in step S2 is received at 30 Hz. and When identified as the target scene and At that time, the vehicle enters the pre-adjustment decision-making stage and calculates the minimum trigger distance based on the current vehicle speed. The formula is as follows:

[0030] ;

[0031] in, Vehicle speed The minimum physical distance required to complete a smooth transition; during online operation, based on real-time vehicle speed. Linear interpolation of adjacent calibration points yields the current corresponding... Value, and obtain ;

[0032] when Upon initial establishment, the system issues a trigger signal, at which point the total transition distance is recorded. and pass the target pattern to step S4; when When the vehicle enters the scene, the system reverts to the real-time switching logic based on the vehicle radar, triggering a switch based on the actual distance measured by the radar.

[0033] A further technical solution, the specific steps of step S4 are as follows:

[0034] Upon receiving the trigger signal, the distance traveled by the vehicle since the trigger is obtained by integrating the pulses from the wheel speed sensors. Define the normalized progress as a function of travel distance. as follows:

[0035] ;

[0036] in, Corresponding trigger time, The corresponding vehicle arrives at the starting point of the scene;

[0037] by Using the sigmoid function as the independent variable, the mixed weights are generated. :

[0038] ;

[0039] in, The steepness coefficient of the Sigmoid function for a positive transition;

[0040] Each control cycle reads the current steering wheel torque. and vehicle speed According to the current calculate Query the two-dimensional mapping function corresponding to the current mode respectively. Two-dimensional mapping function corresponding to the target pattern The weighted mixed output generates the desired lateral acceleration of the vehicle. :

[0041] .

[0042] A further technical solution, the specific steps of step S5 are as follows:

[0043] Once the vehicle confirms that it has entered the target scenario, it initiates differentiated closed-loop verification. During the driving process, it continuously performs multi-frame judgments and uses asymmetric frame count and hysteresis threshold design to confirm in real time whether to maintain the current mode or reverse transition back to the original mode.

[0044] In narrow road scenarios, the sum of the minimum detection distances of the ultrasonic radars on the left and right sides is taken as the current passable width; when the width measured in 8 consecutive frames is less than the lower threshold, the target scene verification is passed and the precision mode is maintained; when the width is greater than the upper threshold in 5 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scene and a reverse transition is triggered.

[0045] In sharp curve scenarios, the curvature radius of the lane lines is monitored continuously over multiple frames using visual technology. When the curvature radius is less than the lower threshold for 10 consecutive frames, the target scenario is verified. When the curvature radius is greater than the upper threshold for 5 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scenario, triggering a reverse transition.

[0046] In the construction area scenario, millimeter-wave radar and visual recognition results are combined for joint verification. When both detect the same obstacle in the same spatial location, the target scene verification is successful. If only visual recognition is performed but radar does not obtain a valid echo, the visual result is used for verification and is not counted as a sensor conflict. If radar detects an obstacle but visual recognition fails, or if the spatial locations of the two detection results are significantly inconsistent, the sensor conflict frame statistics are included. When the detection results conflict for more than 10 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scene, triggering a reverse transition.

[0047] In the parking lot entrance scene, the system continuously detects the barrier gate, parking sign, or the "P" character on the road surface. If the same target is detected for 5 consecutive frames, the verification is successful. If the target is lost for 10 consecutive frames, it is determined that the current environment no longer meets the scene characteristics of the target, and a reverse transition is triggered.

[0048] In a further technical solution, in step S5, the reverse transition uses the same Sigmoid smoothing mechanism as in step S4, employing a fixed transition duration, thus the reverse mixing weights are... for:

[0049] ;

[0050] in, To trigger the time after the reverse transition, The central moment of the transition duration, This is the Sigmoid steepness coefficient for the reverse transition.

[0051] The present invention provides a multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition, the advantages of which are as follows:

[0052] (1) Forward-looking prediction and early switching: The target scene is identified in advance by the forward-looking camera, and the minimum trigger distance is calculated based on the vehicle speed. The mode switching is completed before the vehicle enters the scene, eliminating the defect of the traditional solution's switching lag.

[0053] (2) Spatial synchronous smooth transition: The progress is normalized by driving distance and combined with the Sigmoid weight function to achieve smooth and gradual change of steering characteristics. The transition progress is strictly synchronized with the scene position and is not affected by vehicle speed fluctuations, resulting in a natural feel.

[0054] (3) High recognition reliability: It integrates image quality, semantic segmentation and temporal consistency to comprehensively assess confidence and effectively filter false detections; it adopts visual / radar differentiated ranging for different scenarios to balance accuracy and robustness.

[0055] (4) Prevent repeated switching: Asymmetric frame count and hysteresis threshold are used for scene verification. By using different thresholds than exit thresholds, frequent mode oscillations at the boundary are avoided, thus improving system stability.

[0056] (5) Multi-sensor fault tolerance: The radar and vision are combined for joint verification in construction areas and other scenarios. The design of conflict frame statistics ensures that the system can still work reliably when a single sensor fails, thus enhancing environmental adaptability.

[0057] (6) Safety first: The reverse transition adopts Sigmoid smooth recovery, monitors the driver's resistance operation throughout the process and responds first, fully reflecting the human-machine co-driving safety concept. Attached Figure Description

[0058] Figure 1 A flowchart of a multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition provided in an embodiment of the present invention;

[0059] Figure 2 The graph shows the Sigmoid transition weight function.

[0060] Figure 3 This is a schematic diagram illustrating the changes in mapping relationships during the smooth transition of steering modes.

[0061] Figure 4 This is the radar verification and reverse transition closed-loop control logic diagram. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0063] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0064] like Figure 1 As shown, an embodiment of the present invention provides a multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition, comprising the following steps:

[0065] Step S1: Pattern library construction;

[0066] Steering wheel torque was collected under different vehicle speed conditions for each preset steering mode through real-vehicle testing. With lateral acceleration The corresponding data, in terms of steering wheel torque and vehicle speed As input variables, vehicle lateral acceleration Construct a two-dimensional mapping function for the output variable. And store them in the pattern library for later use.

[0067] To determine an accurate model applicable to low-speed, complex scenarios such as narrow roads, sharp bends, parking lot entrances, and construction areas, this embodiment selects a typical low-speed condition as the evaluation condition. The typical low-speed condition corresponds to a vehicle speed of 15 km / h, a standard load condition, and a road surface adhesion coefficient greater than 0.8.

[0068] Under this typical low-speed condition, data on steering wheel torque and lateral acceleration for each steering mode were collected. In everyday driving, lateral acceleration is concentrated... to Within the linear interval, this embodiment selects at equal intervals within this interval. discrete data points ,in, For the first Steering wheel torque at each sampling point For the first Lateral acceleration at each sampling point; linear fitting using the least squares method is performed, and the slope of the resulting straight line is the steering sensitivity under this mode. :

[0069] ;

[0070] in, The number of sampling points. The average value of the steering wheel torque across all sampling points. This is the average value of the lateral acceleration at all sampling points. It reflects the lateral acceleration output generated by a unit steering wheel torque input. The smaller the value, the weaker the vehicle's steering response and the more stable the steering feel under the same torque. This characteristic matches the handling requirements of scenarios such as narrow roads and sharp bends. The vehicle is not overly sensitive to steering wheel input, so as to avoid over-response caused by driver tension or misoperation. The larger the diameter, the more sensitive the steering response, making it suitable for efficient driving on regular roads.

[0071] Calculate all preset steering modes After the value, The smallest judgment is for Precision Mode, and the middle judgment is for Standard Mode; if the vehicle only provides two modes, the other mode is automatically designated as Standard Mode. Two-dimensional mapping functions corresponding to each steering mode. Stored in the pattern library for use in step S4 for pattern switching and smooth weighting.

[0072] Step S2: Forward perception and scene type recognition;

[0073] The vehicle-mounted forward-facing camera acquires raw images at a preset frequency (30 Hz). First, distortion correction is performed to eliminate edge distortion of the wide-angle lens, ensuring the accuracy of scene geometry. Then, image enhancement and noise reduction are performed to improve the clarity of key features such as lane lines and obstacle edges, and to suppress sensor noise and road texture interference, thereby reducing the probability of false detection in subsequent semantic segmentation.

[0074] Inverse perspective transformation constructs a homography matrix from the image plane to the road plane using parameters obtained from camera calibration, converting the forward perspective view into a vertical, bird's-eye view. This transformation restores the true geometric relationship of lane lines that appear larger in the foreground and smaller in the background due to perspective effects.

[0075] Semantic segmentation employs a lightweight network to perform pixel-level classification of the bird's-eye view, outputting road surface, lane lines, and preset obstacle categories. These preset obstacle categories include walls, guardrails, shoulders, barriers, water-filled barriers, traffic cones, and parking signs. Based on this, scene types such as narrow roads, sharp bends, parking lot entrances, and construction zones can be identified, and other scenes can be expanded as needed. Simultaneously, the precise mode is set as the target mode to improve low-speed steering stability.

[0076] Scene determination is based on geometric thresholds and multi-frame consistency, as detailed below:

[0077] Narrow path determination: Based on the obstacle information output by semantic segmentation, the passable width boundary is extracted, and the passable width is calculated. In this embodiment, the narrow path is determined when there are 10 consecutive frames with a width less than the threshold of 3.2 m.

[0078] Sharp curve detection: The curvature radius is fitted using the output lane line information, and is determined when it is less than the curvature threshold of 50 m for 8 consecutive frames.

[0079] Parking lot entrance determination: Based on the output obstacle information such as the barrier gate, octagonal parking sign, or road information such as the "P" character on the road, the determination is made when the above features appear within a preset distance in front of the vehicle and the same target is detected for 5 consecutive frames.

[0080] Construction area determination: When the detection density of temporary isolation facilities such as water-filled barriers and traffic cones exceeds the preset threshold for 5 consecutive frames, a determination is made.

[0081] All of the above thresholds and frame counts can be adjusted according to the actual vehicle sensor configuration.

[0082] Overall confidence level The calculation formula is:

[0083] ;

[0084] in, Image quality confidence level based on image contrast and sharpness assessment; The semantic segmentation confidence score is the average pixel classification probability within the target scene region, used to characterize the overall credibility of the semantic segmentation result in the target region. The temporal consistency confidence score is based on the stability of multi-frame recognition. , and For the weighting coefficients, satisfying In this embodiment, , , .when When the value is below a preset threshold, it is marked as uncertain. In this embodiment, it is taken as... The time stamp is set to uncertain, and the system will subsequently revert to the radar real-time switching logic.

[0085] The remaining distance measurement employs a differentiated strategy: in scenarios without physical obstacles, visual ranging is the primary method; in scenarios with physical obstacles, visual ranging is the primary method, supplemented by radar ranging for verification, and the minimum of the two values ​​is output. Specifically, in scenarios without physical obstacles such as narrow roads and sharp bends, visual ranging is the primary method; in parking lot entrances and construction areas with physical obstacles such as barriers and cones, visual ranging and radar ranging are combined, and the minimum of the two values ​​is output as the current remaining distance. .

[0086] Step S3: Decision-making target pattern and triggering timing;

[0087] The scene type output in step S2 is received at 30 Hz. and When identified as the target scene and At that time, the vehicle enters the pre-adjustment decision-making stage and calculates the minimum trigger distance based on the current vehicle speed. The formula is as follows:

[0088] ;

[0089] in, Vehicle speed The minimum physical distance required to complete a smooth transition. Pre-calibration was achieved through subjective evaluation experiments: at several constant vehicle speeds, multiple drivers experienced the complete handover process under different pre-adjusted distances. The minimum distance at which 90% of drivers rated the handover as "unabrupt" was taken as the calibration point for that speed, and each calibration point was stored as a speed-distance lookup table function. During online operation, the speed was adjusted based on real-time vehicle speed. Linear interpolation of adjacent calibration points yields the current corresponding... Value, and obtain .

[0090] when Upon initial establishment, the system issues a trigger signal, at which point the total transition distance is recorded. And pass the target pattern to step S4. When When the vehicle enters the scene, the system reverts to the real-time switching logic based on the vehicle radar, triggering a switch based on the actual distance measured by the radar.

[0091] Step S4: Smooth transition;

[0092] Upon receiving the trigger signal, the distance traveled by the vehicle since the trigger is obtained by integrating the pulses from the wheel speed sensors. This method uses the travel distance as the transition progress benchmark, enabling synchronization of the vehicle's spatial position relative to the scene's starting point. A normalized progress is defined that varies with the travel distance. as follows:

[0093] ;

[0094] in, Corresponding trigger time, The corresponding vehicle arrives at the starting point of the scene.

[0095] by Using the sigmoid function as the independent variable, the mixed weights are generated. :

[0096] ;

[0097] in, The steepness coefficient of the sigmoid function for a positive transition can be obtained through subjective evaluation and experimental calibration. In this embodiment, we take... ,but:

[0098] ;

[0099] like Figure 2 As shown, the Sigmoid function normalizes the progress. Mixed weights mapped to patterns This enables a smooth transition between the current mode and the precise mode.

[0100] In each control cycle, which takes 10 ms in this embodiment, the current steering wheel torque is read. and vehicle speed According to the current calculate Query the two-dimensional mapping function corresponding to the current mode respectively. Two-dimensional mapping function corresponding to the target pattern The weighted mixed output generates the desired lateral acceleration of the vehicle. :

[0101] ;

[0102] like Figure 3 As shown, The output is entirely determined by the standard pattern mapping function. The output is entirely determined by the precise mode mapping function. During the transition from 0 to 1, the vehicle's steering response characteristics smoothly transition to the precision mode. Only follow change, The transition process is determined by the distance already traveled and is not affected by instantaneous speed fluctuations.

[0103] Step S5: Scenario Authenticity Verification and Fault Tolerance;

[0104] like Figure 4 As shown, once the vehicle confirms it has entered the target scene, it initiates differentiated closed-loop verification. During the journey, continuous multi-frame judgments are performed. Through the design of asymmetric frame counts and hysteresis thresholds, it is determined in real time whether to maintain the current mode or reverse transition back to the original mode. Specifically, potential misjudgments in forward perception are corrected in real time, and the scene is determined to end when environmental features no longer meet the conditions of the target scene, triggering a reverse transition to avoid repeated mode switching at the boundary.

[0105] In narrow road scenarios, the sum of the minimum detection distances of the ultrasonic radars on both the left and right sides is taken as the current passable width. If the width measured in 8 consecutive frames is less than the lower threshold of 3.2 m, the target scene verification is passed and the precision mode is maintained; if the width is greater than the upper threshold of 3.8 m in 5 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scene, and a reverse transition is triggered.

[0106] In sharp curve scenarios, the curvature radius of the lane line is monitored continuously over multiple frames using visual technology. When the curvature radius is less than the lower threshold of 50 m for 10 consecutive frames, the target scene is verified. When the curvature radius is greater than the upper threshold of 60 m for 5 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scene, triggering a reverse transition.

[0107] In the construction area scenario, millimeter-wave radar and visual recognition results are used for joint verification. When both detect the same obstacle in the same spatial location, the target scene verification is successful. If only visual recognition is performed but radar does not obtain a valid echo, the visual result is used for verification and is not counted as a sensor conflict. If radar detects an obstacle but visual recognition fails, or if the spatial locations of the two detection results are significantly inconsistent, the frame count for sensor conflict is recorded. When the detection results conflict for more than 10 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scene, triggering a reverse transition.

[0108] In the parking lot entrance scenario, visual detection is used to continuously detect the barrier gate, parking sign, or the "P" character on the road surface. In this embodiment, the verification is successful if the same target is detected for 5 consecutive frames; if the target is lost for 10 consecutive frames, it is determined that the current environment no longer meets the scene characteristics of the target, triggering a reverse transition. The number of consecutive detection frames and the hysteresis threshold can be obtained through real vehicle testing and can be adjusted according to sensor performance.

[0109] The reverse transition uses the same Sigmoid smoothing mechanism as step S4, employing a fixed transition duration, and then the reverse mixing weights... for:

[0110] ;

[0111] in, To trigger the time after the reverse transition, The central moment of the transition duration, This is the Sigmoid steepness coefficient for the reverse transition. In this embodiment, we take... Preset total reverse transition time Then the central time Substituting, we get:

[0112] ;

[0113] Gradually restore the steering mode from the current mode to the original mode to ensure a smooth driving experience.

[0114] In addition, the method also includes:

[0115] During the transition, the driver's steering wheel input is continuously monitored. If the driver is detected to have obvious resistance, such as applying torque that deviates significantly from the expected range of the current mode, or rapidly turning the steering wheel (in this embodiment, the angular velocity is greater than 80 ° / s), the transition is paused or the weight change rate is reduced to prioritize responding to the driver's intentions. The driver is alerted to the mode change information through the instrument panel or steering wheel vibration, but without interfering with driving.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition, characterized in that, Includes the following steps: Step S1: Collect the corresponding data of steering wheel torque and lateral acceleration of each preset steering mode of the vehicle under different vehicle speed conditions through real vehicle test. With steering wheel torque and vehicle speed as input variables and vehicle lateral acceleration as output variable, construct the two-dimensional mapping function corresponding to each steering mode and store it in the mode library. Step S2: Acquire the original image through the vehicle-mounted front-view camera, and sequentially perform distortion correction, image enhancement and noise reduction, and inverse perspective transformation to obtain a bird's-eye view; perform semantic segmentation on the bird's-eye view to identify the target scene type; Step S3: When the target scene is identified and the overall confidence level is not lower than the preset threshold, the minimum trigger distance is calculated based on the current vehicle speed. When the current remaining distance is not greater than the minimum trigger distance, a trigger signal is issued, the total transition distance is recorded, and the target mode is passed to the next step. Step S4: After receiving the trigger signal, the ratio of the distance the vehicle has traveled since the trigger time to the total transition distance is used as the normalized progress. The normalized progress is used as the independent variable to generate a hybrid weight. In each control cycle, the two-dimensional mapping function corresponding to the current mode and the two-dimensional mapping function corresponding to the target mode are queried according to the current steering wheel torque and vehicle speed, respectively. The hybrid weight is used to perform weighted hybrid output to generate the desired vehicle lateral acceleration, so as to achieve a smooth transition from the current mode to the target mode. Step S5: After the vehicle enters the target scene, it continuously performs multi-frame judgment. Through the design of asymmetric frame count and hysteresis threshold, it confirms in real time whether to maintain the current mode. When the environmental features no longer meet the conditions of the target scene, a reverse transition is triggered to gradually restore the steering mode to the original mode. In step S2, the scene determination is based on a geometric threshold and multi-frame consistency, as follows: Narrow path determination: Extract the passable width boundary based on the obstacle information output by semantic segmentation, calculate the passable width, and determine the narrow path if it is less than the width threshold for 10 consecutive frames; Sharp curve detection: The curvature radius is fitted using the output lane line information, and a curve is detected when it is less than the curvature threshold for 8 consecutive frames. Parking lot entrance determination: Determines the entrance based on the output features of the barrier gate, octagonal parking sign, or "P" character on the road surface when the above features appear within a preset distance in front of the vehicle and the same target is detected for 5 consecutive frames. Construction area determination: Detect water barriers and / or cones, and determine the construction area when the detection density exceeds a preset threshold for 5 consecutive frames; Overall confidence level The calculation formula is: ； in, Image quality confidence level based on image contrast and sharpness assessment; The semantic segmentation confidence score corresponds to the average pixel classification probability within the target scene region. The temporal consistency confidence score is based on the stability of multi-frame recognition. , and For the weighting coefficients, satisfying ;when If the value is below a preset threshold, it is marked as uncertain, and the system will subsequently revert to the real-time radar switching logic. The remaining distance measurement employs a differentiated strategy: visual ranging is primarily used in narrow roads and sharp bends; for parking lot entrances and construction zones where barriers and cones are present, a combination of visual and radar ranging is used, with the minimum value of the two being output as the current remaining distance. ; The specific steps of step S3 are as follows: The scene type output in step S2 is received at 30 Hz. and When identified as the target scene and At that time, the vehicle enters the pre-adjustment decision-making stage and calculates the minimum trigger distance based on the current vehicle speed. The formula is as follows: ； in, Vehicle speed The minimum physical distance required to complete a smooth transition; during online operation, based on real-time vehicle speed. Linear interpolation of adjacent calibration points yields the current corresponding... Value, and obtain ; when Upon initial establishment, the system issues a trigger signal, at which point the total transition distance is recorded. and pass the target pattern to step S4; when When the vehicle enters the scene, the switching logic will revert to real-time switching based on the vehicle radar, triggering a switch based on the actual distance measured by the radar. The specific steps of step S4 are as follows: Upon receiving the trigger signal, the distance traveled by the vehicle since the trigger is obtained by integrating the pulses from the wheel speed sensors. Define the normalized progress as a function of travel distance. as follows: ； in, Corresponding trigger time, The corresponding vehicle arrives at the starting point of the scene; by Using the sigmoid function as the independent variable, the mixed weights are generated. : ； in, The steepness coefficient of the Sigmoid function for a positive transition; Each control cycle reads the current steering wheel torque. and vehicle speed According to the current calculate Query the two-dimensional mapping function corresponding to the current mode respectively. Two-dimensional mapping function corresponding to the target pattern The weighted mixed output generates the desired lateral acceleration of the vehicle. : 。 2. The multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition according to claim 1, characterized in that, In step S1, a typical low-speed operating condition is selected as the evaluation condition. Typical low-speed operating conditions correspond to a vehicle speed of 15 km / h, the vehicle being under standard load, and a road surface adhesion coefficient greater than 0.8; Steering wheel torque was collected for each steering mode under this typical low-speed condition. With lateral acceleration Corresponding data; in to Equal intervals are selected within the linear interval discrete data points ,in, For the first Steering wheel torque at each sampling point For the first Lateral acceleration at each sampling point; linear fitting using the least squares method is performed, and the slope of the resulting straight line is the steering sensitivity under this mode. : ； in, The number of sampling points. The average value of the steering wheel torque across all sampling points. This is the average lateral acceleration across all sampling points; Calculate all preset steering modes After the value, The smallest judgment is for Precision Mode, and the middle judgment is for Standard Mode; if the vehicle only provides two modes, the other mode is automatically designated as Standard Mode; the two-dimensional mapping function corresponding to each steering mode. Stored in the pattern library for use in step S4 for pattern switching and smooth weighting.

3. The multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition according to claim 2, characterized in that, In step S2, semantic segmentation uses a lightweight network to perform pixel-level classification of the bird's-eye view, outputting road surface, lane lines, and preset obstacle categories. The preset obstacle categories include walls, guardrails, shoulders, barriers, water-filled barriers, cones, and parking signs, thereby identifying narrow roads, sharp bends, parking lot entrances, and construction zones.

4. The multi-mode adaptive pre-adjustment method for steering based on foreground scene recognition according to claim 3, characterized in that, The specific steps of step S5 are as follows: Once the vehicle confirms that it has entered the target scenario, it initiates differentiated closed-loop verification. During the driving process, it continuously performs multi-frame judgments and uses asymmetric frame count and hysteresis threshold design to confirm in real time whether to maintain the current mode or reverse transition back to the original mode. In narrow road scenarios, the sum of the minimum detection distances of the ultrasonic radars on the left and right sides is taken as the current passable width; when the width measured in 8 consecutive frames is less than the lower threshold, the target scene verification is passed and the precision mode is maintained; when the width is greater than the upper threshold in 5 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scene and a reverse transition is triggered. In sharp curve scenarios, the curvature radius of the lane lines is monitored continuously over multiple frames using visual technology. When the curvature radius is less than the lower threshold for 10 consecutive frames, the target scenario is verified. When the curvature radius is greater than the upper threshold for 5 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scenario, triggering a reverse transition. In the construction area scenario, the results of millimeter-wave radar and visual recognition are combined for joint verification; when both detect the same obstacle in the same spatial location, the target scene verification is successful. If only visual recognition is achieved but radar does not obtain a valid echo, the visual result is used for verification and is not counted as a sensor conflict; if radar detects an obstacle but visual recognition fails, or if the spatial positions of the two detection results are significantly inconsistent, it is included in the sensor conflict frame statistics; when the detection results conflict for more than 10 consecutive frames, it is determined that the current environment no longer meets the characteristics of the target scene, and a reverse transition is triggered. In the parking lot entrance scenario, the system continuously detects the barrier gate, parking sign, or the "P" character on the road surface using vision. If the same target is detected for 5 consecutive frames, the verification is successful. If the target is lost for 10 consecutive frames, it is determined that the current environment no longer meets the scene characteristics of the target, and a reverse transition is triggered.

5. The multi-mode adaptive pre-adjustment method for steering based on forward scene recognition according to claim 4, characterized in that, In step S5, the reverse transition uses the same Sigmoid smoothing mechanism as in step S4, employing a fixed transition duration, and the reverse mixing weights are... for: ； in, To trigger the time after the reverse transition, The central moment of the transition duration, This is the Sigmoid steepness coefficient for the reverse transition.

Images