Intelligent vehicle lane-changing trajectory automatic generation method and system based on deep learning

Abstract

The application relates to a deep learning-based intelligent vehicle lane-changing trajectory automatic generation method and system, and belongs to the technical field of automatic driving intelligent vehicle trajectory planning and control. A preselected lane-changing trajectory point set is acquired, a trajectory curvature sequence and a curvature space variation rate sequence are extracted. Adjacent points satisfying a polarity reversal condition are detected by traversing the curvature sequence, a curvature reversal point is obtained, and a curvature space variation rate peak value in the neighborhood of the curvature reversal point is extracted. A dynamic threshold value is calculated based on a longitudinal vehicle speed and a chassis load rigidity distribution parameter, the local trajectory is subjected to curvature reconstruction when the peak value exceeds the threshold value, a target trajectory point set and a control instruction are generated, and a preset space range is updated based on feedback optimization parameters. The application realizes cross-modal coupling of chassis mechanics constraints and trajectory planning, and can obviously inhibit yaw oscillation problems in continuous lane changing.

Description

Technical Field

[0001] This invention belongs to the field of trajectory planning and control technology for autonomous driving intelligent vehicles, specifically a method and system for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning. Background Technology

[0002] The mainstream approach to autonomous driving trajectory planning employs end-to-end deep learning or parametric curve methods, aiming to output smooth lane-change trajectories at the geometric level. However, existing methods generally treat the vehicle as an ideal rigid body, assuming that tire grip is constant or changes linearly. This assumption is acceptable for steady-state cornering, but it leads to serious physical deviations under continuous S-shaped lane-change conditions: at the curvature polarity reversal point, the abrupt change in lateral acceleration causes an instantaneous reversal of the vehicle's roll momentum, and the weight is rapidly transferred from the rear wheel diagonal to the front wheel side, causing the rear axle grip model to collapse.

[0003] Current technology fails to detect the limits of chassis mechanical transfer, still outputting a geometric trajectory requiring a high steering response rate. This causes the rate of change of curvature space to exceed the dynamic load transfer limits of the suspension and tires, resulting in instantaneous micro-slippage of the rear wheels. While the underlying PID controller detects the yaw angle error and attempts to forcefully correct it, because the tires have already entered the nonlinear slip zone, this triggers 2-3 cycles of severe yaw oscillations, seriously threatening safety and tracking robustness.

[0004] Existing approaches fall into two categories: one is to improve the mathematical smoothness of the trajectory, attempting to replace physical perception with geometric smoothness; the other is to design complex adaptive laws at the lower control layer to passively resist disturbances. Both have significant limitations: the former ignores the physical boundaries of the chassis's mechanical counterweight under dynamic conditions, and global smoothing may still violate the suspension response limits; the latter passively responds, correcting after sideslip occurs, resulting in slow convergence and instability. Therefore, it is urgent to address the root cause of the underlying control oscillations and tracking loss-of-control problems in continuous lane changes caused by the trajectory planning's failure to perceive dynamic axle load transfer limits. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for automatically generating lane change trajectories of intelligent vehicles based on deep learning. The aim is to transform the static counterweight characteristics of the chassis mechanics into a physical forbidden zone in the dynamic curvature derivative space, and after the deep learning trajectory is output, to implement local rigid limiting and reconstruction based on the current vehicle speed only at the trajectory curvature polarity reversal point, thereby preventing the occurrence of rear wheel sideslip and PID oscillation from the physical root.

[0006] The objective of this application can be achieved through the following technical solution: Firstly, a method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning, comprising the following steps:

[0007] Obtain a set of pre-selected lane change trajectory points output by a preset trajectory generation network, and extract the trajectory curvature sequence and curvature space rate of change sequence based on the set of pre-selected lane change trajectory points;

[0008] Traverse the trajectory curvature sequence, and when adjacent trajectory points satisfy the preset curvature polarity reversal condition, obtain the trajectory curvature reversal point based on the adjacent trajectory points, and extract the peak value of the curvature space change rate centered on the trajectory curvature reversal point within the preset spatial range in the curvature space change rate sequence.

[0009] The longitudinal driving speed of the target vehicle is obtained, and based on the longitudinal driving speed and the preset chassis load rigidity distribution parameters, the curvature space change rate threshold for the trajectory curvature reversal point is obtained.

[0010] When the peak value of the curvature space change rate is greater than the curvature space change rate threshold, the pre-selected lane change trajectory point set is subjected to local curvature reconstruction within a preset space range to obtain the corresponding target trajectory point set;

[0011] Based on the target trajectory point set, a target control command for the target vehicle is generated, feedback optimization parameters characterizing the confidence level of the target control command are obtained, and the preset spatial range is updated based on these parameters.

[0012] Secondly, the intelligent vehicle lane-changing trajectory automatic generation system based on deep learning includes the following modules:

[0013] The extraction module is used to obtain a set of pre-selected lane change trajectory points output by a preset trajectory generation network, and extract the trajectory curvature sequence and the curvature space change rate sequence based on the set of pre-selected lane change trajectory points.

[0014] The peak extraction module is used to traverse the trajectory curvature sequence. When adjacent trajectory points satisfy the preset curvature polarity reversal condition, the trajectory curvature reversal point is obtained based on the adjacent trajectory points. The peak value of the curvature space change rate is extracted from the curvature space change rate sequence with the trajectory curvature reversal point as the center and within a preset spatial range.

[0015] The threshold acquisition module is used to obtain the longitudinal driving speed of the target vehicle, and based on the longitudinal driving speed and the preset chassis load rigidity distribution parameters, obtain the curvature space change rate threshold for the trajectory curvature reversal point.

[0016] The reconstruction module is used to perform local curvature reconstruction on the preselected lane change trajectory point set within a preset space range to obtain the corresponding target trajectory point set when the peak value of the curvature space change rate is greater than the curvature space change rate threshold.

[0017] The feedback optimization module is used to generate target control commands for the target vehicle based on the target trajectory point set, obtain feedback optimization parameters characterizing the confidence level of the target control commands, and update the preset spatial range based on them.

[0018] Thirdly, a computer storage medium stores computer-executable instructions, which, when executed, implement the deep learning-based intelligent vehicle lane-changing trajectory automatic generation method described in the first aspect.

[0019] Compared with the prior art, the beneficial effects of this application are:

[0020] This application constructs a cross-modal reconstruction architecture with strong coupling between geometric programming and mechanical constraints. It implements rigid amplitude limiting correction based on vehicle speed only in the local neighborhood of the curvature polarity reversal point, fundamentally eliminating the conditions for rear wheel sideslip. Gaussian decay feedback optimization and a preset spatial range dynamic expansion mechanism are introduced to achieve adaptive closed-loop: when tracking deteriorates, the reconstruction range is automatically expanded to make curvature reversal smoother, supplemented by a maximum upper bound constraint to ensure convergence. This significantly shortens the PID yaw oscillation convergence time during S-shaped lane changes, completely eliminates high-speed tail-wagging and drifting phenomena, and significantly improves tracking robustness and safety under extreme weight conditions. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the steps of the deep learning-based intelligent vehicle lane-changing trajectory automatic generation method of this application;

[0022] Figure 2 This is a schematic diagram of the modules of the intelligent vehicle lane-changing trajectory automatic generation system based on deep learning in this application. Detailed Implementation

[0023] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only to illustrate selected embodiments of this application.

[0024] Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item has been defined in one figure, it does not need to be further defined and explained in subsequent figures. The terms first, second, etc. are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] With the continuous development of autonomous driving technology, deep learning methods have been widely used in the field of intelligent vehicle trajectory planning. The existing mainstream technical approach usually uses end-to-end deep neural networks or parametric methods such as polynomials and Bézier curves for trajectory planning. Its core objective is to output a continuous and smooth lane-changing trajectory at the geometric level, and rely on the underlying PID control algorithm to track and execute the trajectory.

[0026] However, existing deep learning trajectory planning methods suffer from a long-overlooked systemic flaw in engineering applications: when building trajectory planning models, vehicles are generally treated as ideal rigid bodies, and it is assumed that tire grip (i.e., ground adhesion) remains constant or exhibits a simple linear change during lane changes. This rigidity assumption is acceptable under steady-state turning conditions, but it will lead to serious deviations from physical reality under continuous lane change (i.e., S-shaped obstacle avoidance) conditions.

[0027] Specifically, when an intelligent vehicle executes an S-shaped lane-changing trajectory, at the inflection point where the curvature polarity reverses (i.e., switching from a left turn to a right turn, or vice versa), the vehicle's lateral acceleration undergoes a sudden polarity change. For vehicles with a specific front-to-rear weight distribution, this sudden change causes an instantaneous reversal of the vehicle's roll momentum, resulting in a rapid and drastic transfer of vehicle weight from the rear wheel diagonal to the front wheel side. The original high rear axle grip model collapses instantly, and existing deep learning networks, failing to perceive this chassis mechanical transfer limit during the planning phase, will still output a geometric trajectory requiring a high steering response rate.

[0028] The engineering consequences of the aforementioned root cause defect can be described as follows: The rate of change of curvature space (i.e., the change in curvature per unit distance traveled, physically equivalent to the steering wheel rotation rate) at the curvature inversion point of the trajectory output by deep learning exceeds the dynamic load transfer limit that the suspension and tires can withstand. The rear wheels will instantly lose traction and experience micro-slippage. After detecting the yaw angle error, the underlying PID controller will output a large corrective control value. However, since the tires have already entered the nonlinear slip region, the strong intervention of the PID controller not only fails to mitigate the error but also directly triggers severe yaw oscillations lasting for 2 to 3 control cycles (commonly known as the "dragon-drawing phenomenon"), seriously affecting vehicle driving safety and trajectory tracking robustness.

[0029] To address the aforementioned issues, existing technologies generally follow two paths: one is to continuously improve the mathematical smoothness of the trajectory at the global trajectory planning layer, for example, by using higher-order spline curves or Bézier curves, attempting to replace the active perception of physical limits with geometric smoothness; the other is to design more complex adaptive control laws at the lower control layer, attempting to compensate for the disconnect between the planning layer and the execution layer with passive anti-disturbance processing.

[0030] Both approaches have inherent limitations: the former ignores the nonlinear physical boundaries of specific chassis mechanical weights under dynamic conditions, and the globally smooth trajectory may still violate the suspension response limits at curvature polarity reversal points; the latter is a passive response, attempting to correct after vehicle sideslip has already occurred, resulting in long convergence times and unstable performance. Therefore, how to fundamentally solve the systemic technical defects of underlying control oscillations and trajectory tracking loss of control caused by deep learning trajectory planning's failure to perceive dynamic axle load transfer limits under conventional continuous lane-changing conditions is a technical problem that urgently needs to be solved in this field.

[0031] Therefore, such as Figure 1 As shown, this application provides a method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning, including the following steps:

[0032] Obtain a set of pre-selected lane change trajectory points output by a preset trajectory generation network, and extract the trajectory curvature sequence and curvature space rate of change sequence based on the set of pre-selected lane change trajectory points;

[0033] Traverse the trajectory curvature sequence, and when adjacent trajectory points satisfy the preset curvature polarity reversal condition, obtain the trajectory curvature reversal point based on the adjacent trajectory points, and extract the peak value of the curvature space change rate centered on the trajectory curvature reversal point within the preset spatial range in the curvature space change rate sequence.

[0034] The longitudinal driving speed of the target vehicle is obtained, and based on the longitudinal driving speed and the preset chassis load rigidity distribution parameters, the curvature space change rate threshold for the trajectory curvature reversal point is obtained.

[0035] When the peak value of the curvature space change rate is greater than the curvature space change rate threshold, the pre-selected lane change trajectory point set is subjected to local curvature reconstruction within a preset space range to obtain the corresponding target trajectory point set;

[0036] Based on the target trajectory point set, a target control command for the target vehicle is generated, feedback optimization parameters characterizing the confidence level of the target control command are obtained, and the preset spatial range is updated based on these parameters.

[0037] In another implementation, such as Figure 2 As shown, this application also provides a deep learning-based intelligent vehicle lane-changing trajectory automatic generation system, including the following modules:

[0038] The extraction module is used to acquire a pre-selected track change trajectory point set output by a preset trajectory generation network, and extract the trajectory curvature sequence and the curvature space change rate sequence based on the pre-selected track change trajectory point set. The extraction module is the data preprocessing entry point of the system. It calculates the geometric curvature point by point by calling the three-point difference processing algorithm, and calculates the curvature space change rate segment by segment by discrete differential processing, providing operable feature sequence data for subsequent modules.

[0039] The peak extraction module is used to traverse the trajectory curvature sequence. When adjacent trajectory points satisfy a preset curvature polarity reversal condition, the module obtains the trajectory curvature reversal point based on the adjacent trajectory points. It then extracts the peak value of the curvature space change rate within a preset spatial range centered on the trajectory curvature reversal point from the curvature space change rate sequence. This peak extraction module is the core of the system's hazard feature identification. By determining the polarity reversal when the product of the curvatures of two adjacent points is less than zero, it accurately anchors high-risk physical areas in the S-shaped trajectory. Furthermore, by maximizing the absolute value, it extracts the peak change rate of the worst operating conditions, ensuring that no over-limit actions in any direction are missed.

[0040] The threshold acquisition module is used to obtain the longitudinal driving speed of the target vehicle. Based on the longitudinal driving speed and preset chassis load rigidity distribution parameters, it obtains the curvature space change rate threshold for the trajectory curvature reversal point. This threshold acquisition module is the core of the system's physical constraint boundary calculation. By simultaneously solving the chassis dynamics rigid body model and the tire friction mechanics formula, it transforms the chassis hardware characteristics into a safety limit boundary that dynamically changes with real-time vehicle speed, and obtains it in real-time on the embedded platform using a lookup table interpolation method.

[0041] The reconstruction module is used to perform local curvature reconstruction on the pre-selected lane change trajectory point set within a preset spatial range to obtain the corresponding target trajectory point set when the peak value of the curvature space change rate exceeds the curvature space change rate threshold. The reconstruction module is the core execution unit of the system, including a clamping processing submodule, a safety curvature integration submodule, and a spatial position iterative reconstruction submodule. These three submodules work sequentially, replacing the original curvature curve exceeding physical limits with a local planar point set that meets chassis safety constraints, and smoothly merging it with the original trajectory to ensure a smooth transition between the reconstructed segment and the preceding and following original trajectories in terms of position and tangential direction.

[0042] The feedback optimization module is used to generate target control commands for the target vehicle based on the target trajectory point set, obtain feedback optimization parameters characterizing the confidence level of the target control commands, and update the preset spatial range based on these parameters. The feedback optimization module is the core of the system's adaptive closed-loop control. It uses a Gaussian decay function to comprehensively map the lateral displacement deviation and yaw rate deviation executed at the lower level into a confidence level evaluation value, and uses a dynamic expansion formula with a maximum spatial range constraint to achieve adaptive updates of the preset spatial range, forming a complete closed-loop control architecture with cross-modal coupling of the planning layer, execution layer, and feedback layer.

[0043] Step 1: Obtaining the pre-selected lane change trajectory point set and extracting its feature sequence;

[0044] The preset trajectory generation network is the data input source for the entire method. It is built into the processing terminal of the target vehicle and is responsible for generating a preliminary lane-changing trajectory based on real-time perception data. The input of the network consists of two parts of physical feature data of different modalities fused in the spatiotemporal dimension: one is the environmental spatial feature vector, which is extracted from the sequence of track or road environment images collected by the vehicle's front-end perception sensors, and after image preprocessing (including binarization, difference ratio and operation, etc.), the discrete boundary coordinate point set representing the left and right edges of the road ahead is extracted;

[0045] The second component is the vehicle state time series vector, which consists of historical motion state data collected by onboard sensors at a fixed sampling period, including longitudinal speed, lateral acceleration, and yaw rate over the current control cycle and multiple consecutive historical control cycles. The system aligns and concatenates these two data components in the time domain to generate a multimodal input feature matrix. This matrix is ​​then input into the network and processed through multiple nonlinear mapping and trajectory decoding output layers. Finally, a pre-selected lane-change trajectory point set is output in the target vehicle's local reference coordinate system, in the form of an ordered sequence of horizontal and vertical coordinates.

[0046] ;

[0047] Where (xᵢ, yᵢ) represents the x and y coordinates of the i-th look-ahead trajectory point in the vehicle's local coordinate system, and N is the total number of look-ahead points. This point set is immediately passed to the subsequent feature sequence extraction process after generation, serving as the initial input baseline for the entire physical constraint clamping method.

[0048] After obtaining the pre-selected lane change trajectory point set, the system uses three-point difference processing to obtain the geometric curvature of each trajectory point and combines them into a trajectory curvature sequence. Since the trajectory output by the deep learning network is a discrete set of coordinate points rather than a continuous mathematical equation, the three-point difference method must be introduced for numerical differentiation. The three-point difference method utilizes the current point... and one adjacent point before and after it. and (A total of three points) jointly estimate the geometric curvature of the current point. Compared with two-point difference, it can utilize first-order and second-order derivative information at the same time, resulting in higher calculation accuracy and eliminating phase lag error, which is suitable for the real-time operation requirements of low-computing-power embedded controller platforms.

[0049] Using the coordinate step size Δt as the parameter interval between adjacent points, the formula for calculating the three-point central difference of the first derivative (tangent direction vector) is as follows:

[0050] ;

[0051] The formula for calculating the second derivative (curvature vector) is:

[0052] ;

[0053] Substituting the first and second derivatives into the classical plane curve curvature formula, we obtain the geometric curvature κᵢ of the i-th trajectory point:

[0054] ;

[0055] In this context, the sign of κᵢ represents the turning direction (positive for left turn, negative for right turn), and the amplitude represents the steepness of the curve. By iterating through all trajectory points in the pre-selected lane change trajectory point set and performing the above three-point difference processing point by point, a complete trajectory curvature sequence can be obtained:

[0056] ;

[0057] Based on the extracted trajectory curvature sequence, the system further extracts the geometric curvature difference between adjacent trajectory points and their corresponding physical arc length difference, obtains the curvature space change rate of each trajectory point through discrete differential processing, and combines them into a curvature space change rate sequence.

[0058] The rate of change of curvature space (i.e., the first derivative of curvature with respect to arc length) is physically equivalent to the angular velocity of the steering wheel. This is the core characteristic quantity for determining whether the trajectory violates the physical limits of the chassis in this invention. (Extracting adjacent trajectory points...) and Geometric curvature difference between And the corresponding difference in physical arc length between the two points. (Approximated by Euclidean distance), then the first... The rate of change of curvature space at each trajectory point is:

[0059] ;

[0060] By traversing all adjacent point pairs and performing the above discrete differentiation process point by point, the complete sequence of curvature space change rates can be obtained:

[0061] ;

[0062] The absolute value of each element in this sequence represents the degree of steering wheel rotation at that location, and is the direct object of subsequent high-risk area identification and physical constraint. The trajectory curvature sequence and the curvature space change rate sequence together serve as the basic input data for downstream identification and processing modules.

[0063] Step 2: Anchoring of the trajectory curvature reversal point and extraction of the peak value of the rate of change of curvature space;

[0064] During continuous lane changes (S-shaped obstacle avoidance), the lateral acceleration polarity reversal of the vehicle is the core cause of the instantaneous transfer of dynamic load on the rear wheels, leading to sideslip. The system traverses the trajectory curvature sequence along the look-forward direction, sequentially extracting the first geometric curvature of adjacent first trajectory points. The second geometric curvature of the second trajectory point And perform the curvature polarity reversal condition judgment: when the product of the first geometric curvature and the second geometric curvature is less than zero;

[0065] The determination that the first and second trajectory points satisfy the curvature polarity reversal condition has the physical meaning that when the curvature signs of two adjacent trajectory points are opposite, the vehicle's steering direction reverses between these two points from left to right (or from right to left), meaning the trajectory has passed through the zero-value point of curvature (curvature zero-crossing point). To adapt to low-computing-power embedded controller platforms, this invention adopts a dimensionality reduction approximation strategy, directly using the position corresponding to the first trajectory point (i.e., the point with index i) as the trajectory curvature reversal point. All subsequent local operations are performed with this position index as the reference center. This processing method fully meets the engineering accuracy requirements because subsequent steps extract the peak value within the neighborhood of the reversal point, and a small deviation of the reference center does not affect the effectiveness of peak value identification.

[0066] After anchoring the position index i of the trajectory curvature reversal point, the system extracts a subset of candidate trajectory points within the preset spatial range based on this position index and the preset spatial range. The preset spatial range is a spatial length parameter (in meters) representing physical distance, used to define the size of a local neighborhood window centered on the trajectory curvature reversal point. At the code implementation level, this is converted into the boundary span of the array index by the ratio of the arc length interval between adjacent points. Subsequently, the system obtains a set of candidate curvature space change rates corresponding to the candidate trajectory point subset from the curvature space change rate sequence, and extracts the curvature space change rate with the largest absolute value as the peak curvature space change rate. :

[0067] The use of absolute values ​​to extract peak values ​​here is determined by the physical nature of the vehicle chassis: the fundamental cause of rear wheel sideslip is the intensity of dynamic axle load transfer, which is unrelated to the direction of steering wheel rotation (left or right). Regardless of whether the vehicle pulls the steering wheel to the left or right at an extremely high speed, the amplitude of the transient roll impact force borne by the suspension is equivalent. If only the mathematical maximum value is taken instead of the absolute value, when there is a very large negative rate of change (corresponding to a sudden right turn at extremely high speed, which is physically extremely dangerous) and a very small positive rate of change (physically safe) within the window, the trajectory will be incorrectly judged as safe, allowing a fatal trajectory to proceed.

[0068] Step 3: Calculate the threshold of curvature space change rate based on longitudinal driving speed and chassis load rigidity distribution parameters;

[0069] The chassis load stiffness distribution parameters refer to a set of static calibration constants that describe the mechanical characteristics of the target vehicle chassis, including the rear axle track width, vehicle mass, center of gravity height, and rear axle roll stiffness distribution coefficient. These parameters are determined by the mechanical structure at the time of vehicle manufacture and can be obtained through offline calibration, without the need for online real-time measurement.

[0070] Among them, the rear axle roll stiffness distribution coefficient This comprehensively reflects the ratio of roll stiffness between the front and rear axles of a vehicle—for a specific weighted vehicle with a front-to-rear axle load ratio of approximately 4.5:5.5. Approximately equal to the rear axle load ratio of 0.55, this represents the load transfer share borne by the rear axle under roll conditions. At the point of trajectory curvature reversal, the vehicle's lateral acceleration... A sudden change in polarity occurs, resulting in a change in the dynamic load transfer of the rear axle. This determines whether the vehicle loses control due to the collapse of rear wheel traction. Based on rigid body dynamics and tire friction mechanics, the causal model is as follows:

[0071] ;

[0072] Where m is the total vehicle mass. For the height of the center of mass, Given the rear axle track, with the chassis hardware parameters fixed, the only dynamic variable affecting rear wheel grip is the product of the square of the vehicle speed and the trajectory curvature.

[0073] To prevent overshoot of roll momentum due to response lag during reverse rotation of the chassis suspension in an S-curve, the derivative of axle load transfer with respect to time must be less than the suspension's ultimate tolerance damping constant. By transforming the time-domain shaft load transfer rate constraint into a spatial-domain curvature derivative constraint, a curvature spatial rate of change threshold for the curvature reversal point of the trajectory can be obtained. The analytical formula, combined with longitudinal driving speed and the preset load transfer critical damping Obtain the threshold of the rate of change of curvature space. :

[0074] ;

[0075] Curvature space rate of change threshold With vehicle speed The equation has a strict inverse relationship with the cube, meaning that the higher the vehicle speed, the more drastically the permissible steering wheel speed must decrease exponentially. This mathematical relationship fundamentally reveals the physical cause of the high-speed S-shaped lane-changing phenomenon—in traditional trajectory planning, as vehicle speed increases, the constraint on the rate of curvature change is not tightened simultaneously, inevitably leading to axle load collapse under high-speed conditions. To adapt to low-computing-power controller platforms, the system pre-converts the above analytical formula offline into a one-dimensional safety mapping lookup table indexed by vehicle speed, and uses a linear interpolation scheme to calculate the safety mapping lookup table based on the current longitudinal speed during real-time vehicle operation. The calculation can be solved directly by looking up the table, without having to repeatedly perform floating-point operations online.

[0076] Step 4: Local curvature reconstruction of the pre-selected lane change trajectory point set;

[0077] First, boundary checks are performed on the peak value and threshold of the rate of change of curvature space: when At that time, the steering wheel speed near the curvature reversal point of the original trajectory is within the physical tolerance range of the chassis, and no correction is required. The original pre-selected lane change trajectory point set is directly issued as the target trajectory point set.

[0078] when At this point, the original trajectory will inevitably cause the dynamic load on the rear axle to exceed the suspension damping limit. The system then activates a local nonlinear clamping correction mechanism: it acquires all curvature space change rates in the candidate curvature space change rate set whose absolute values ​​are greater than the curvature space change rate threshold, and forcibly truncates and replaces them with the curvature space change rate threshold to obtain an updated safe curvature space change rate set. All elements in the safe curvature space change rate set satisfy the condition that their absolute values ​​do not exceed... This ensures, from a physical perspective, that the steering wheel turning speed at any point in the subsequent trajectory does not exceed the physical limit of the chassis dynamic load transfer.

[0079] After obtaining the set of safe curvature space change rates, the system performs numerical integration along the driving arc length direction based on the set of safe curvature space change rates to obtain a local safe curvature sequence within a preset spatial range. The curvature values ​​at the boundaries corresponding to the preset spatial range are then used as the basis for this process. As the initial condition for integration, we integrate step by step along the arc length direction to recover the local safe curvature at each arc length position:

[0080] ;

[0081] in, The initial curvature value at the beginning of the local correction interval boundary (taken from the original trajectory boundary point that has not been modified). The set of rates of change of curvature in the safe curvature space at the arc length position The rate of change of curvature space at that point, The distance from the current target point to the starting point of the boundary is the arc length. Let be the integral variable during the integration process. With the maximum points They play completely different roles in mathematical grammar and in physical terms;

[0082] This represents the total arc length (a defined upper limit) of the distance between the target point and the starting point that we want to calculate. From 0 to The process variables are dynamically scanned during the integration process and disappear after integration is completed. The two cannot be used interchangeably to maintain the complete clarity of the integration formula logic.

[0083] Because the local safety curvature sequence differs from the original curvature sequence after clamping correction, the corresponding trajectory geometric coordinates also change. Therefore, iterative spatial location reconstruction is necessary to eliminate data gaps. Iterative spatial location reconstruction includes the following three sub-steps:

[0084] 1) Based on the local safety curvature sequence, perform a heading angle integral along the driving arc length direction to obtain the target heading angle corresponding to each arc length position in the reconstructed trajectory. The formula is as follows:

[0085] ;

[0086] in, Indicates the position of arc length The target heading angle at that location, This represents the initial heading angle (taken from the vehicle's heading angle at the beginning of the local correction interval boundary, determined by the tangent direction of the unmodified original trajectory boundary point). This indicates the local safety curvature sequence at the arc length position. The curvature value at that point, For integral variables.

[0087] 2) Based on the sine and cosine values ​​corresponding to each target heading angle, spatial coordinates are accumulated along the travel arc length direction to obtain the lateral and longitudinal correction amounts corresponding to each arc length position. The specific formula is as follows:

[0088] ;

[0089] ;

[0090] in, and These represent the arc length positions of the reconstructed local plane point set. The x and y coordinates of the location and The x and y coordinates of the starting position of the boundary are respectively represented. Indicates the position of arc length The target heading angle at that location. The three sets of integral formulas mentioned above together restore the modified safety curvature curve to a high-dimensional local coordinate point set that satisfies physical constraints, completing the accurate mapping from safety curvature space to physical plane coordinates.

[0091] 3) Overlap and add the lateral and longitudinal correction values ​​with the x and y coordinates corresponding to the starting positions of the boundary, respectively, to obtain the reconstructed local planar point set. Finally, based on the reconstructed local planar point set, replace the original trajectory points within the preset spatial range in the pre-selected lane change trajectory point set, while retaining the original trajectory points outside the preset spatial range unchanged, thereby obtaining the target trajectory point set.

[0092] Through the aforementioned local curvature reconstruction process, this invention achieves precise physical compliance correction of the deep learning trajectory output: it preserves the original geometry of the trajectory in non-dangerous areas, performs only the minimum necessary local clamping and reconstruction in the neighborhood of the curvature polarity reversal point, maintains the original intent of the top-level planning to the greatest extent, and ensures that the steering wheel turning speed of the trajectory at the curvature reversal point is strictly constrained within the chassis suspension dynamic load transfer limit.

[0093] Step 5: Generate target control commands based on the target trajectory point set;

[0094] The target vehicle's control system operates with a fixed control cycle (e.g., once every 10 milliseconds). The target trajectory point set provides a complete path plan covering multiple forward points. The underlying actuators can only perform one fixed action per control cycle. Therefore, a look-ahead distance strategy and spatial interpolation techniques must be combined to extract the target control curvature corresponding to the current control cycle from the target trajectory point set. Specifically, the system dynamically calculates the look-ahead distance based on the current longitudinal speed, measures the target trajectory point set sequence to the pre-aiming point at that look-ahead distance, and accurately obtains the curvature value at the virtual pre-aiming point through linear or spline interpolation between adjacent points, which serves as the target control curvature.

[0095] After obtaining the target control curvature, the system inputs the target control curvature into a preset steering mapping model to output the target front wheel steering angle. The steering mapping model uses a simplified two-degree-of-freedom bicycle model, and calculates the corresponding target front wheel steering angle using Ackermann steering geometry. :

[0096] ;

[0097] in, The target steering angle for the front wheels. A fixed wheelbase (i.e., the distance between the center points of the front and rear wheels) is used for vehicles. The target control curvature is determined by the intercepted curve. The target front wheel steering angle is obtained. Subsequently, the system generates a pulse width modulation (PWM) signal with a corresponding control duty cycle based on the target front wheel steering angle, and sends it to the target vehicle as a target control command. The generation of the PWM signal is achieved through a pre-calibrated linear mapping equation from angle to duty cycle:

[0098] ;

[0099] in, This is the reference duty cycle when the servo is in the absolute homing state (corresponding to a typical analog servo with a duty cycle of approximately 7.5%). This is the proportional coefficient of the servo motor hardware (determined through calibration experiments). The system writes the calculated duty cycle value into the compare register of the microcontroller's internal hardware timer. The timer hardware automatically and continuously outputs a PWM square wave corresponding to the duty cycle on the pin, driving the front wheel servo motor to precisely deflect to the target angle, thus achieving target trajectory tracking.

[0100] Since the target curvature sequence entering the bottom-level servo has been physically constrained and clamped, the steering angular velocity of the vehicle when making a reverse angle in an S-curve is strictly limited within the suspension response capability. The chassis load gets sufficient mechanical transfer buffer time when alternating between left and right, and the rear axle tires always stay within the linear zone of adhesion, which fundamentally ensures the rapid convergence of the PID and achieves smooth cornering without yaw oscillation.

[0101] Step Six: Obtaining feedback optimization parameters and adaptively updating the preset space range;

[0102] First, the system obtains the lateral displacement deviation, which characterizes the difference between the actual driving state and the desired driving state of the target vehicle after executing the target control command. and yaw rate deviation Lateral displacement deviation represents the lateral distance (in meters) by which the vehicle's actual driving position deviates from the target trajectory. In engineering implementation, it is obtained through image processing from the front-facing camera combined with a pre-calibrated inverse perspective transformation matrix (IPM). Yaw angular velocity deviation represents the deviation (in radians / second) between the vehicle's actual rotational angular velocity and the theoretical rotational angular velocity expected based on the target curvature and the current vehicle speed. It is obtained by subtracting the measured angular velocity from the onboard inertial measurement unit (IMU) from the kinematically expected angular velocity.

[0103] Lateral displacement deviation is an integral error, which reacts relatively slowly and represents the macroscopic accumulation of historical errors; yaw rate deviation is a differential error, which reacts extremely quickly and can detect abnormal changes in steering angular velocity before the vehicle begins to veer off course. Combining these two physical quantities—one slow and one fast, one macroscopic and one microscopic—constitutes the most comprehensive state assessment basis for the effectiveness of the underlying control execution.

[0104] After obtaining the lateral displacement deviation and yaw rate deviation, the system uses a preset displacement penalty weight. And yaw angle penalty weight Gaussian decay calculation is performed on the lateral displacement deviation and yaw rate deviation to obtain normalized feedback optimization parameters. The formula for calculating Gaussian decay is:

[0105] ;

[0106] in, The feedback optimization parameter (i.e., confidence level) has a value range strictly between 0 and 1. Characterizes the weight of the lateral displacement penalty; Characterizes the actual lateral displacement deviation; Characterizes the weight of the yaw angle penalty; It characterizes the actual yaw rate deviation.

[0107] To ensure the dimensionality consistency of the values ​​within the exponential function, the lateral displacement penalty weight... The yaw angle penalty weight includes a normalization factor (with dimensions m⁻²) that is inversely proportional to the square of the actual lateral displacement deviation. It includes a normalization factor (dimensions s² / rad²) that is inversely proportional to the square of the actual yaw rate deviation. Through the aforementioned penalty weights that have a physical dimensionality reduction effect, the length deviation and rotational velocity deviation, which cannot be directly added in physics, are transformed into dimensionless pure proportional values, ensuring the rigor of the confidence assessment model at both the mathematical and physical levels.

[0108] When the vehicle perfectly follows the target trajectory (with deviations approaching zero), A value close to 1.0 indicates that the target control command was executed with high confidence; when the vehicle body skids or veers (deviation increases sharply), The exponentially rapid decline to near zero accurately reflects the collapse of confidence in the underlying control execution, thus forming an effective alarm signal.

[0109] When feedback optimization parameters Less than the preset confidence threshold This indicates that the underlying control execution is showing a trend of tracking degradation, requiring an expansion of the spatial range of local curvature reconstruction to smooth the trajectory changes at curvature reversal points. The system is based on a preset spatial expansion coefficient. and maximum preset space range For the current preset spatial range Perform an update to obtain the updated preset space range. The corresponding iterative update formula is:

[0110] ;

[0111] Among them, the introduction Operation and Upper bound constraints are a key safety design feature to prevent the dynamic expansion mechanism from causing infinite numerical divergence: if the system remains in an unstable state for a long period of time ( consistently below The compensation term in the formula will continue to accumulate. If there is no upper bound constraint, the spatial range will expand infinitely, causing all curves to be forcibly straightened and lose their turning ability, thus triggering a new risk of loss of control.

[0112] The introduction of this feature rigidly limits the expansion of the spatial range to the physical upper bound corresponding to the camera's line-of-sight distance or chassis dynamics constraints, ensuring the absolute convergence of the system under any continuous perturbation. Updated preset spatial range This will be applied to the local curvature reconstruction task in all subsequent control cycles until the next feedback evaluation triggers a new update. When the feedback optimization parameters... Greater than or equal to the confidence level safety threshold At this time, the current preset space range remains unchanged to prevent frequent adjustments from interfering with system stability when control is good.

[0113] Through the aforementioned adaptive closed-loop mechanism, this invention constructs an anti-skid system with dynamic adjustment characteristics: when the vehicle is driving smoothly (high confidence), the system maintains a small preset space range to preserve the flexible steering response of the trajectory; when the vehicle shows signs of skidding (low confidence), the system automatically expands the preset space range and forcibly extends the curvature of the trajectory to reverse its course, sacrificing micro-geometric response for macro-stability, thus forming an active defense against the physical limits of the chassis.

[0114] Through the above technical solution, the present invention can achieve the following technical effects: (1) The present invention injects the chassis physical limit into the trajectory planning layer in reverse for the first time, and constructs a cross-modal reconstruction architecture with strong coupling between geometric planning and mechanical constraints. Unlike the traditional global mathematical smoothing or passive adaptive control approach, the present invention only implements rigid amplitude correction based on the current vehicle speed in the local neighborhood of the curvature polarity reversal point of the deep learning trajectory output, and actively injects the dynamic axle load transfer limit of the chassis suspension as the physical constraint of the trajectory planning in advance, thereby eliminating the conditions for rear wheel sideslip from the root.

[0115] (2) This invention achieves dynamic calculation of the curvature space change rate threshold. By modeling the chassis load rigidity distribution parameters such as vehicle mass, center of gravity height, rear axle track, and rear axle roll stiffness distribution coefficient in conjunction with real-time longitudinal driving speed, an analytical formula for the safety threshold, which is inversely proportional to the cube of the vehicle speed, is given, ensuring the physical legality of the limited-amplitude maneuver and the effectiveness of anti-skid measures at any driving speed. This formula reveals from first-principles principles the physical law that the allowable steering wheel turning rate must decrease sharply by a cubic exponential rate as the vehicle speed increases, providing rigorous mechanical support for the entire technical solution.

[0116] (3) This invention introduces a feedback optimization parameter based on Gaussian decay and a dynamic expansion mechanism for a preset spatial range to construct a complete adaptive closed loop. By evaluating the execution confidence of the target control command in real time, the system automatically expands the spatial range of local curvature reconstruction when the underlying control shows a trend of tracking degradation, so that the trajectory completes the curvature polarity reversal earlier and more smoothly. By introducing the maximum preset spatial range as an upper bound constraint, the infinite divergence of the dynamic expansion mechanism is effectively prevented, ensuring the absolute convergence of the system under any continuous disturbance.

[0117] (4) Without adding any sensors or computing hardware, this invention can significantly shorten the convergence time of the bottom-level PID yaw oscillation during continuous S-shaped lane changes by simply reconstructing the cross-layer coupling of software algorithms, completely eliminating the tail-swing and dragon-drawing phenomena during high-speed S-shaped lane changes, and significantly improving the trajectory tracking robustness and driving safety of autonomous vehicles under extreme chassis counterweight.

[0118] In another embodiment, this application also provides a computer storage medium storing computer-executable instructions, which, when executed, implement the aforementioned method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning.

[0119] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning, characterized in that, Includes the following steps: Obtain a set of pre-selected lane change trajectory points output by a preset trajectory generation network, and extract the trajectory curvature sequence and curvature space change rate sequence based on the set of pre-selected lane change trajectory points; Traverse the trajectory curvature sequence, and when adjacent trajectory points satisfy the preset curvature polarity reversal condition, obtain the trajectory curvature reversal point based on the adjacent trajectory points, and extract the peak value of the curvature space change rate centered on the trajectory curvature reversal point within the preset spatial range in the curvature space change rate sequence. The longitudinal driving speed of the target vehicle is obtained, and based on the longitudinal driving speed and the preset chassis load rigidity distribution parameters, the curvature space change rate threshold for the trajectory curvature reversal point is obtained. When the peak value of the curvature space change rate is greater than the curvature space change rate threshold, the pre-selected lane change trajectory point set is subjected to local curvature reconstruction within a preset space range to obtain the corresponding target trajectory point set; Based on the target trajectory point set, a target control command for the target vehicle is generated, a feedback optimization parameter characterizing the confidence level of the target control command is obtained, and the preset spatial range is updated based on it. The process of obtaining the threshold of the rate of change of curvature space includes: The chassis load rigidity distribution parameters include the rear axle track width. Overall vehicle quality Center of mass height and rear axle roll stiffness distribution coefficient ; Based on the parameters in the chassis load rigidity distribution parameters, combined with the longitudinal travel speed and the preset load transfer critical damping Obtain the curvature space rate of change threshold for the curvature reversal point of the trajectory. ; 。 2. The method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning according to claim 1, characterized in that, The process of extracting the trajectory curvature sequence and the curvature space rate of change sequence includes: Based on the position of each trajectory point in the pre-selected lane change trajectory point set, the geometric curvature of each trajectory point is obtained through three-point difference processing, and then combined into a trajectory curvature sequence. Extract the geometric curvature difference and its corresponding physical arc length difference between adjacent trajectory points in the trajectory curvature sequence, obtain the curvature space change rate of each trajectory point through discrete differential processing, and combine them into a curvature space change rate sequence.

3. The method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning according to claim 2, characterized in that, The process of extracting the peak value of the rate of change of curvature space includes: Extract the first geometric curvature of the first trajectory point and the second geometric curvature of the second trajectory point from the adjacent first trajectory point in the trajectory curvature sequence; When the product of the first geometric curvature and the second geometric curvature is less than zero, it is determined that the first trajectory point and the second trajectory point satisfy the curvature polarity reversal condition, and the position corresponding to the first trajectory point is taken as the trajectory curvature reversal point. Based on the position index of the trajectory curvature reversal point in the pre-selected lane change trajectory point set and the preset spatial range, a subset of candidate trajectory points within the preset spatial range is extracted; In the curvature space change rate sequence, a set of candidate curvature space change rates corresponding to the candidate trajectory point subset is obtained, and the curvature space change rate with the largest absolute value is extracted as the peak curvature space change rate.

4. The method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning according to claim 3, characterized in that, The process of obtaining the target trajectory point set includes: Obtain each curvature space rate of change in the candidate curvature space rate of change set whose absolute value is greater than the curvature space rate of change threshold, and forcibly truncate and replace it with the curvature space rate of change threshold to obtain an updated safe curvature space rate of change set. Numerical integration processing is performed along the driving arc length direction based on the set of safe curvature spatial change rates to obtain a local safe curvature sequence within a preset spatial range. Spatial position iterative reconstruction is then performed by combining the boundary start position and initial heading angle at the corresponding boundary of the preset spatial range to obtain a reconstructed local planar point set. The original trajectory points within a preset spatial range in the pre-selected lane change trajectory point set are replaced by the reconstructed local plane point set to obtain the target trajectory point set.

5. The method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning according to claim 4, characterized in that, The process of iterative reconstruction of spatial location includes: Based on the local safety curvature sequence, the heading angle is integrated along the driving arc length direction to obtain the target heading angle corresponding to each arc length position in the reconstructed trajectory; Based on the sine and cosine values ​​corresponding to each target heading angle, spatial coordinates are accumulated along the driving arc length direction to obtain the lateral and longitudinal correction amounts corresponding to each arc length position. The horizontal correction amount and the vertical correction amount are respectively superimposed and added to the horizontal and vertical coordinates corresponding to the starting position of the boundary to obtain the reconstructed local planar point set; ； ； ； in, Indicates the position of arc length The target heading angle at that location, Indicates the initial heading angle. This indicates the local safety curvature sequence at the arc length position. The curvature value at that point, and These represent the arc length positions of the reconstructed local plane point set. The x and y coordinates of the location and The x-coordinate and y-coordinate represent the starting position of the boundary, respectively.

6. The method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning according to claim 1, characterized in that, The process of generating target control instructions includes: Based on the target trajectory point set, the target control curvature corresponding to the current control cycle is extracted. The target control curvature is input into the preset steering mapping model to output the target front wheel steering angle. Based on the target front wheel steering angle, a pulse width modulation signal with a corresponding control duty cycle is generated and sent to the target vehicle as a target control command.

7. The method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning according to claim 1, characterized in that, The process of updating the preset space range includes: Obtain the lateral displacement deviation, which characterizes the difference between the actual driving state and the desired driving state of the target vehicle after executing the target control command. and yaw rate deviation ; Based on preset displacement penalty weights And yaw angle penalty weight Gaussian decay calculation is performed on the lateral displacement deviation and yaw rate deviation to obtain normalized feedback optimization parameters. ; When the feedback optimization parameters Less than the preset confidence threshold At that time, based on the preset spatial expansion coefficient and maximum preset space range For the current preset spatial range Update to obtain the updated preset space range And apply it to subsequent tasks.

8. A deep learning-based intelligent vehicle lane-changing trajectory automatic generation system, characterized in that, Includes the following modules: The extraction module is used to obtain a set of pre-selected lane change trajectory points output by a preset trajectory generation network, and extract the trajectory curvature sequence and the curvature space change rate sequence based on the set of pre-selected lane change trajectory points. The peak extraction module is used to traverse the trajectory curvature sequence. When adjacent trajectory points satisfy the preset curvature polarity reversal condition, the trajectory curvature reversal point is obtained based on the adjacent trajectory points. The peak value of the curvature space change rate is extracted from the curvature space change rate sequence with the trajectory curvature reversal point as the center and within a preset spatial range. The threshold acquisition module is used to obtain the longitudinal driving speed of the target vehicle, and based on the longitudinal driving speed and the preset chassis load rigidity distribution parameters, obtain the curvature space change rate threshold for the trajectory curvature reversal point. The reconstruction module is used to perform local curvature reconstruction on the preselected lane change trajectory point set within a preset space range to obtain the corresponding target trajectory point set when the peak value of the curvature space change rate is greater than the curvature space change rate threshold. The feedback optimization module is used to generate target control commands for the target vehicle based on the target trajectory point set, obtain feedback optimization parameters characterizing the confidence level of the target control commands, and update the preset spatial range based on them. The process of obtaining the threshold of the rate of change of curvature space includes: The chassis load rigidity distribution parameters include the rear axle track width. Overall vehicle quality Center of mass height and rear axle roll stiffness distribution coefficient ; Based on the parameters in the chassis load rigidity distribution parameters, combined with the longitudinal travel speed and the preset load transfer critical damping Obtain the curvature space rate of change threshold for the curvature reversal point of the trajectory. ; 。 9. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement the method for automatically generating lane-changing trajectories of intelligent vehicles based on deep learning, as described in any one of claims 1-7.

Images