An automatic avoidance method and system for an intelligent vehicle

Abstract

The application relates to the technical field of automatic avoidance, in particular to an automatic avoidance method and system for an intelligent automobile. The method comprises the following steps: obtaining the driving state of the intelligent automobile in a high-speed state, and monitoring the video frames of the scattered goods in the front lane through a vehicle-mounted camera; based on the video frames, multi-target marking is carried out to obtain the marking data of the scattered goods. Then, the marking data is used to predict the trajectory of the scattered goods, and the avoidance behavior is adjusted according to the driving state of the intelligent automobile to generate automatic avoidance behavior adjustment data; finally, the automatic avoidance strategy model is constructed through the adjustment data, and the model is sent to the intelligent automobile control center to implement the automatic avoidance of the intelligent automobile. The automatic avoidance technology is more accurate through the optimization processing of the automatic avoidance technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic avoidance, and in particular to an automatic avoidance method and system for intelligent vehicles. BACKGROUND

[0002] With the rapid development of intelligent vehicle technology, automatic driving has gradually moved from the experimental stage to practical application. On highways and urban roads, intelligent vehicles can obtain real-time vehicle surrounding environment information through sensors, cameras, radars and other devices, and make judgments and decisions based on this information. However, during driving, unexpected situations such as scattered goods on the front lane, traffic accidents or natural obstacles still pose a threat to vehicle safety. These unexpected obstacles not only increase the complexity of driving, but also easily lead to traffic accidents and affect traffic flow. Therefore, how to quickly and effectively detect and avoid these obstacles to ensure the safe driving of intelligent vehicles in complex environments is one of the key problems that need to be solved in intelligent driving systems. However, the traditional automatic avoidance method for intelligent vehicles has the problems of low accuracy of scattered goods trajectory prediction and large automatic avoidance error. SUMMARY

[0003] Therefore, it is necessary to provide an automatic avoidance method and system for intelligent vehicles to solve at least one of the above technical problems.

[0004] To achieve the above-mentioned purpose, an automatic avoidance method for intelligent vehicles, the method comprising the following steps:

[0005] Step S1: Obtain the driving state of the intelligent vehicle in high-speed state, and monitor the video frames of the scattered goods in the front lane through the vehicle-mounted camera; based on the video frames, multi-target labeling of the scattered goods is performed to obtain multi-target labeling data of the scattered goods;

[0006] Step S2: Based on the multi-target labeling data of the scattered goods, the multi-target scattered goods trajectory is predicted, and then the automatic avoidance behavior adjustment is adjusted according to the driving state of the intelligent vehicle to obtain automatic avoidance behavior adjustment data;

[0007] Step S3: An automatic avoidance strategy model is constructed through the automatic avoidance behavior adjustment data, and the automatic avoidance strategy model is sent to the intelligent vehicle control center to execute the automatic avoidance of the intelligent vehicle.

[0008] Preferably, the present application also provides an automatic avoidance system for intelligent vehicles for executing the automatic avoidance method for intelligent vehicles as described above, the automatic avoidance system for intelligent vehicles comprising:

[0009] The scattered goods marking module is configured to acquire a driving state of the intelligent vehicle in a high-speed state, and monitor video frames of scattered goods in a front lane through a vehicle-mounted camera; multi-target marking of the scattered goods is performed based on the video frames, and multi-target marking data of the scattered goods is obtained;

[0010] The avoidance behavior adjustment module is configured to perform multi-target goods scattered trajectory prediction based on the multi-target marking data of the scattered goods, and then perform automatic avoidance behavior adjustment according to the driving state of the intelligent vehicle, so as to obtain automatic avoidance behavior adjustment data.

[0011] The avoidance model construction module is configured to construct an automatic avoidance strategy model through the automatic avoidance behavior adjustment data, and send the automatic avoidance strategy model to a control center of the intelligent vehicle, so as to execute automatic avoidance of the intelligent vehicle.

[0012] The beneficial effects of the present application are that by obtaining the driving state of the intelligent vehicle and the video frames of the scattered goods in the front lane monitored by the vehicle-mounted camera, the obstacle information in the environment where the vehicle is located can be captured in real time. The video frame data provides a basis for subsequent goods marking and path prediction, ensuring the real-time and accuracy of the data. Through the multi-target marking technology based on video frames, the system can accurately identify and label the specific position and form of multiple scattered goods, providing accurate target information for subsequent avoidance decisions. This step provides high-quality environmental perception data for the entire system and is the basis for automatic avoidance decisions. After obtaining and marking the positions of the scattered goods, based on these multi-target marking data, the system further predicts the scattering trajectory of the goods. This process combines the driving state of the intelligent vehicle (such as speed, direction, etc.) for dynamic trajectory prediction and generates the future motion trajectory of the goods. Through the prediction of the future trajectory, the system can judge the relative position change of the obstacle and the vehicle in advance, and then adjust the avoidance strategy according to the real-time driving state of the intelligent vehicle. The combination of prediction and real-time state adjustment effectively improves the accuracy and response speed of the avoidance strategy, avoiding collisions due to delayed judgment. According to the generated automatic avoidance behavior adjustment data, the system constructs an automatic avoidance strategy model. This model combines the avoidance needs in different scenarios and can automatically adjust the avoidance strategy according to the current vehicle speed, obstacle position, and dynamic changes, etc. to ensure that the avoidance behavior remains efficient and safe in various complex situations. After transmitting this model to the intelligent vehicle control center, the system can execute these avoidance instructions in real time to achieve automatic avoidance operations. Through this process, the intelligent vehicle can autonomously judge and safely avoid obstacles during high-speed driving, significantly improving driving safety and reducing the risk of traffic accidents. Therefore, the present application optimizes the traditional automatic avoidance method for an intelligent vehicle, solves the problem of low accuracy in predicting the scattering trajectory of goods and large automatic avoidance errors in the traditional automatic avoidance method for an intelligent vehicle, improves the accuracy of predicting the scattering trajectory of goods, and reduces the error of automatic avoidance. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 Figure 1 is a schematic flow chart of the steps of the automatic avoidance method for an intelligent vehicle.

[0014] Figure 2 Figure 2 is a detailed implementation step flow chart of step S2 in the automatic avoidance method for an intelligent vehicle. Figure 1 DETAILED DESCRIPTION

[0015] Referring to Figure 1 An automatic avoidance method for an intelligent vehicle, the method comprising the following steps:

[0016] ​Step S1: Obtain the intelligent vehicle driving state in the high-speed state, and monitor the video frames of the scattered goods in the front lane through the vehicle-mounted camera; based on the video frames, multi-target labeling of the scattered goods is performed to obtain multi-target labeling data of the scattered goods;

[0017] In the embodiment of the application, the driving state data transmitted by the vehicle CAN bus is read in real time through the vehicle-mounted OBD interface, including vehicle speed, acceleration, steering angle, GPS positioning coordinates and yaw rate information output by the inertial measurement unit; when the vehicle is in a high-speed running state, the video frame sequence of the front lane is obtained in real time through the binocular high-definition vehicle-mounted camera installed in the middle of the vehicle head. The vehicle control system continuously inputs 60 frames of video data per second to the image processing unit through the embedded multi-thread image acquisition bus interface. The image processing unit first uses the fixed threshold histogram equalization algorithm to perform global brightness equalization on the original video frame, and then uses the edge sharpening method based on the Sobel convolution kernel to enhance the target edge features. The sharpened image data is input to the two-dimensional region grid division module, which divides the fixed grid into 16 equal areas in the horizontal direction and 9 equal areas in the vertical direction, and performs background modeling based on the HSV color space in each region. The dynamic abnormal area is identified by the frame difference method and region motion detection. In each abnormal area, a multi-target detection network based on the improved YOLO structure is used to identify scattered goods and label target bounding boxes using multi-scale features extracted by residual blocks and feature pyramids. The target class and boundary coordinates output by the detection network are smoothed in time sequence by Kalman filtering to generate continuous frame-to-frame goods tracking identification data, and finally form multi-target labeling data of scattered goods, providing input basis for subsequent trajectory prediction.

[0018] Step S2: Based on the multi-target labeling data of the scattered goods, the multi-target goods scattering trajectory is predicted, and then the automatic avoidance behavior adjustment data is obtained according to the automatic avoidance behavior adjustment of the intelligent vehicle driving state.

[0019] In this embodiment of the invention, using multi-target labeled data of scattered goods as input, morphological analysis is performed on the geometric contours of all labeled goods, transforming the boundary coordinate set into a rectangular circumscribed envelope region. Area and angular curvature data for each target are obtained through connected component analysis. Based on this, initial parabolic trajectory parameters are calculated using inter-frame position difference, extracting the initial velocity direction and the initial parabolic height difference. Then, wind speed and direction data are read in real-time from the environmental sensors built into the intelligent vehicle, and added to the force model of the scattered objects through vector decomposition. During the three-dimensional motion analysis of each target, a spatial velocity vector decomposition algorithm based on Newtonian mechanics is used to apply wind field environmental factors to the force angle and force area of ​​the goods, forming dynamic force posture data. Combining the three-dimensional spatial coordinate change sequence, a particle filtering algorithm is used to probabilistically predict the scattered positions at several future time steps, outputting a set of scattered goods trajectory points. The intelligent vehicle control unit synchronously receives real-time speed, steering angle, and acceleration information of the vehicle, establishing its own dynamic state matrix. The trajectory prediction submodule calculates the potential collision probability distribution based on the dynamic relative radius relationship between the vehicle's driving path and the cargo scattering trajectory. The avoidance calculation unit then generates automatic avoidance behavior adjustment data using Bayesian inference methods. This data includes vehicle steering amplitude adjustment commands, safety distance control parameters after deceleration, and power distribution correction coefficients for each axle, ensuring a one-to-one correspondence between the prediction process and the timing of the behavior adjustment data.

[0020] Step S3: Construct an automatic avoidance strategy model using automatic avoidance behavior adjustment data, and send the automatic avoidance strategy model to the intelligent vehicle control center to execute the automatic avoidance of the intelligent vehicle;

[0021] In this embodiment of the invention, automatic obstacle avoidance behavior adjustment data is used as input. First, feature decomposition is performed on the obstacle avoidance behavior data to extract key quantitative features such as vehicle speed change rate, steering wheel angle response time, lateral acceleration amplitude, and longitudinal deceleration distribution. Then, standardization and normalization are applied to ensure consistent numerical distributions across different feature dimensions. To improve the accuracy of the strategy model, the training and validation sets are divided according to a time series window. Convolutional operations are used to encode the temporal context behavior features into fixed-length vectors, which are then input into a multi-layer decision structure based on the random forest algorithm to construct an initial strategy decision tree group. This decision tree group filters effective variables by ranking the importance of multi-dimensional features and performs cross-prediction on the validation set to generate intermediate automatic obstacle avoidance strategy data. Then, error convergence iteration is performed on the decision output, and gradient-weighted averaging is used to correct the obstacle avoidance response lag, ultimately forming an optimized automatic obstacle avoidance strategy model. The optimized model is transmitted to the vehicle control center via a communication bus. The control center issues corresponding control signals to the vehicle's steering motor, brake actuator, and power distribution module based on the path adjustment commands generated by the model, realizing the automatic obstacle avoidance function of the intelligent vehicle in actual operation.

[0022] Step S1 includes the following steps:

[0023] Step S11: Obtain the driving status of the intelligent vehicle at high speed and monitor video frames of goods scattered in the lane ahead using the onboard camera;

[0024] Step S12: Sharpen the video frames to obtain video sharpening data;

[0025] Step S13: Divide the video sharpening data into a regional grid to obtain video region division data;

[0026] Step S14: Based on the video region segmentation data, perform multi-target marking of scattered goods to obtain multi-target marking data of scattered goods.

[0027] In this embodiment of the invention, after the video frames are initially read, image sharpening processing is performed. The processing unit first calls grayscale channel conversion in each frame to convert the color channels into a single brightness matrix. Then, high-pass filtering based on Laplacian convolution kernels is performed on this brightness matrix to enhance edge gradients. The result after high-pass filtering is then superimposed with the original grayscale image to highlight high-frequency components, forming a preliminary sharpened image. To further maintain the balance between texture details and edge sharpness, the system uses a bilateral filtering mechanism to suppress excessive noise caused by sharpening, separating noise points from real feature edges in the weight space. The video sharpening data output by the above two-stage operations forms clear boundary lines at each pixel, providing a clear contrast between scattered objects and road surface light reflection areas in high-speed driving scenes, laying a clear image foundation for subsequent regional analysis.

[0028] Video sharpening data is input into the region segmentation process. A fixed-size regular grid is used to spatially decompose each frame of the image to form video region segmentation data. The segmentation process divides the image into 16×9 squares horizontally and vertically using a uniform partitioning method. Each cell stores the brightness center value, edge gradient magnitude, and color uniformity of its corresponding pixel. Then, at the global level, the edge feature density distribution is calculated for all grid points, and a threshold is selected to identify areas with high feature density as potential target areas. The entire process is achieved through fast region scanning based on integral images, eliminating the need to repeatedly traverse each pixel. In consecutive frames, the brightness variation and color difference fluctuation values ​​of the same grid area are calculated sequentially to identify motion feature sets and distinguish between static backgrounds and dynamic abnormal objects. Finally, a calibrated grid matrix is ​​formed, with each cell containing a description of the region's content attributes, including motion state marker codes, edge texture statistics, and offset trend vectors, providing a spatial index reference for multi-target labeling.

[0029] Based on the aforementioned video region segmentation data, multi-target labeling of scattered goods is implemented in the detected potential dynamic regions. First, the motion contour region is obtained within each region through inter-frame differencing. Then, a foreground target recognition operation based on convolutional feature extraction is used to cluster the shape features of goods appearing in consecutive frames. This process utilizes morphological closing operations to connect incomplete regions to obtain complete shape boundaries. The coordinates of the four vertices of the enclosing rectangle are sequentially extracted for each shape boundary, and the temporal movement path of the region's center point is smoothed. Subsequent recognition operations use a fixed-category sample comparison method to match and judge different geometric features of the goods, identifying categories such as metal parts, plastic shells, and boxes, each with a unique number. All labeling information is detected frame-by-frame in the temporal dimension. The target number is continuously updated using the migration trajectory of the target's center point in the frame sequence, generating a multi-target labeling result file. This file records the region index, detection timestamp, geometric boundary, and category number of each goods target in the video. Finally, scattered goods multi-target labeling data is output, providing a complete identification basis for the next stage of trajectory analysis and avoidance adjustment.

[0030] refer to Figure 2 The aforementioned step S2 includes the following steps:

[0031] Step S21: Perform geometric morphology analysis on the scattered cargo multi-target marking data to obtain multi-target cargo geometric morphology data;

[0032] In this embodiment of the invention, geometric morphology analysis is performed on the multi-target marker data of scattered goods. Each target has fixed two-dimensional boundary information, which is described by the set of vertex coordinates obtained in the previous identification. The system first calculates the perimeter and contour closure of the target shape using a contour extraction algorithm, and fills in the missing boundary parts with linear interpolation. Then, a complete target area mask is generated within the closed contour using a pixel-filling scanning method, and the number of pixels occupied by the target is calculated to extract its equivalent area value. To analyze the geometric curvature characteristics of the target, the contour point sequence is smoothed using an arc length parameterization method to obtain a continuous edge curvature sequence, and then the morphological structure radian is calculated by endpoint radian integral to reflect the aerodynamic characteristics of the goods shape. While performing geometric morphology calculation for each target, the system performs three-dimensional reconstruction of the target center point coordinates and obtains a depth estimate through the parallax of a binocular camera, thereby forming geometric shape description data in three-dimensional space, and finally outputs multi-target goods geometric morphology data, establishing an accurate morphological basis for subsequent trajectory prediction.

[0033] Step S22: Perform initial parabolic state analysis on the scattered cargo multi-target marking data to obtain the initial parabolic state of the multi-target cargo;

[0034] In this embodiment of the invention, initial parabolic state analysis is performed on multi-target tagged data of scattered goods. First, the center point position sequence of the same target in adjacent video frames is read, and the initial velocity and direction of the target are calculated using time intervals and pixel displacements. For the projectile characteristics under high-speed conditions, a fixed inter-frame difference algorithm is used to calculate the rate of change of displacement in the vertical direction to determine the initial launch angle. Subsequently, acceleration estimation is performed on the continuous center point positions of each target to obtain the initial parabolic motion trend. Simultaneously, the system analyzes crosswind data detected by vehicle sensors, reading the ambient wind speed and direction angle, thereby correcting the initial velocity component of the target from a dynamic perspective. To ensure the stability of the analysis results, a time-series sliding window averaging method is used on multi-frame data to eliminate jitter noise, ultimately generating initial parabolic state data for multi-target goods. This data includes the initial displacement vector, initial velocity direction angle, vertical and horizontal velocity decompositions, and environmental correction parameters, providing deterministic input for the next step of trajectory prediction calculation.

[0035] Step S23: Based on the geometric shape data of the multi-objective cargo and the initial parabolic state of the multi-objective cargo, predict the scattering trajectory of the multi-objective cargo to output the predicted scattering trajectory of the multi-objective cargo.

[0036] In this embodiment of the invention, the trajectory prediction of multiple target cargo is performed based on the geometric shape data and initial parabolic state of the multiple target cargo. By analyzing the position and velocity variables of each target in the three-dimensional coordinate system, a discrete prediction algorithm based on time steps is used to calculate the position changes at several subsequent time points. During the prediction process, the aforementioned shape area and shape structure curvature data are added as correction factors for air resistance, converting aerodynamic effects into velocity attenuation ratios, and superimposing them with the decomposed amount of environmental wind direction to form a comprehensive motion vector. The system performs iterative calculations of spatial displacement for each target in three directions, outputting the three-dimensional position coordinates of the target at each time step. To prevent the accumulation of measurement errors, weighted smoothing is introduced throughout the trajectory prediction process, fusing historical points and predicted points according to time series weights to form a continuous and stable trajectory curve. Finally, a set of predicted trajectories for multiple target cargo is generated for all targets. Each trajectory in this set consists of a discrete point sequence, containing time labels, spatial coordinates, and velocity direction information, which can be used in the avoidance analysis stage.

[0037] Step S24: Adjust the automatic avoidance behavior based on the driving status of the intelligent vehicle and the predicted trajectory of multi-target cargo scattering to obtain automatic avoidance behavior adjustment data;

[0038] In this embodiment of the invention, automatic avoidance behavior adjustment is performed based on the intelligent vehicle's driving state and the predicted trajectory of scattered multi-target cargo. First, vehicle driving state parameters, including real-time vehicle speed, acceleration, vehicle attitude angle, and steering angle information, are read to establish a dynamic vehicle state set. Then, by spatially mapping with the predicted trajectory of the scattered cargo, the spatial distance between the vehicle's current position and the trajectory points of the scattered cargo is calculated at each time step, thereby obtaining dynamic relative radius data. After each distance calculation, the system performs a collision risk assessment. When the relative radius of any predicted point is lower than a safety threshold, the avoidance adjustment process is triggered. The avoidance adjustment process involves three levels of calculation: the first level calculates the lateral displacement that the vehicle should adjust based on the current route offset; the second level calculates the required speed correction value based on the difference between the vehicle's speed and the predicted speed of the cargo ahead; and the third level determines the steering angle correction range based on the vehicle's wheel yaw characteristics. Finally, the three layers of output results are superimposed to generate automatic avoidance behavior adjustment data. This data describes the vehicle's dynamic adjustment trajectory over the next few seconds in a multi-dimensional numerical manner, including acceleration / deceleration curves, steering angle adjustment plans, and vehicle attitude change sequences, and is provided to the automatic control system to execute real-time avoidance actions.

[0039] Step S23 includes the following steps:

[0040] Step S231: Extract the morphological area and morphological curvature from the geometric morphology data of multi-target cargo; obtain the current environmental wind conditions on the highway;

[0041] Step S232: Analyze the initial drop velocity and attitude of the multi-target cargo in its initial parabolic state to obtain the initial parabolic velocity and parabolic attitude of the multi-target cargo;

[0042] Step S233: Based on the environmental wind conditions, fit the scattering trajectory evolution of the shape area, shape structure curvature, initial parabolic velocity and parabolic attitude of the multi-target cargo to output cargo scattering trajectory evolution data;

[0043] Step S234: Perform multi-target cargo scattering trajectory prediction based on cargo scattering trajectory evolution data, and output the multi-target cargo scattering prediction trajectory.

[0044] In this embodiment of the invention, the system calls the generated multi-target cargo geometric morphology data and extracts the morphological area value and morphological structure curvature value of each cargo target. The extraction process is performed sequentially according to the cargo number order. First, the target coverage area is calculated based on the pixel set of the two-dimensional boundary contour and recorded as morphological area data. Then, the edge curvature sequence is calculated using the arc length method on the same contour sequence, and the morphological structure curvature data is obtained through continuous curvature integration. The morphological area and structural curvature data of all targets are stored sequentially to form a feature matrix, which is used to describe the air force-related characteristics. Subsequently, the environmental wind conditions on the highway are read in real time through the vehicle-mounted barometric pressure sensor and wind speed sensor. The wind condition data is output in the form of wind speed value and wind direction angle, and is bound to the current frame timestamp through time synchronization unit to ensure that the environmental parameters are consistent with the target's state at that time. Finally, the morphological area, morphological structure curvature, and wind condition parameters are stored together on the same time axis to ensure the accuracy of the subsequent trajectory evolution fitting input.

[0045] Initial drop velocity and attitude analysis are performed on the initial parabolic state of multiple targets. The analysis first determines the rate of change of the center point coordinates of each target in consecutive video frames, and then calculates the initial drop velocity vector based on the inter-frame time interval. Subsequently, independent velocity components are calculated in the vertical and horizontal decomposition directions. The vertical component characterizes the amplitude of the object's lift-off, while the horizontal component reflects the scattering direction. By combining the cargo attitude image sequence from the previous time period, the cargo's spatial orientation angle, tilt angle, and rotation angle are extracted, and the orientation in the triangular coordinate system is normalized to obtain the drop attitude data. To reduce measurement fluctuations, the moving average method is used to smooth the velocity and attitude parameters across multiple consecutive frames, ensuring the continuity of initial velocity and attitude changes. After processing, the initial drop velocity value and three-dimensional drop attitude description data for each target are output, providing initial state input for subsequent trajectory evolution calculations.

[0046] Based on environmental wind conditions, the trajectory evolution of scattered cargo is fitted using the area, curvature, initial parabolic velocity, and parabolic attitude of the cargo. The calculation first adjusts the force angle of the cargo's shape and attitude according to the environmental wind direction, applying the wind direction component to the area direction of the cargo surface, and calculating the force projection through force vector decomposition. Then, the force amplitude is linearly scaled up according to the wind speed to form the wind pressure value. Next, the attitude deflection trend of the cargo during the scattering process is calculated based on the curvature of its shape and its parabolic attitude. Numerical integration is then performed on the force change process at a fixed time step to obtain the dynamic attitude change sequence. The force and attitude change data are embedded into the three-dimensional motion sequence, and the cargo's spatial velocity vector is continuously updated to maintain a temporal correspondence between the velocity direction and the wind influence. The spatial velocity distribution is obtained through time accumulation and further processed into intermediate trajectory evolution data. To suppress abrupt change errors, a weighted time-series average is used to smooth the velocity change rate between adjacent time points. After all calculations are completed, a complete data set of cargo scattering trajectory evolution is generated, which describes the trajectory evolution characteristics of cargo over time under the influence of external wind force and geometric forces.

[0047] Multi-target cargo trajectory prediction is performed based on cargo scattering trajectory evolution data. The processing unit first resamples the trajectory evolution data over time, unifying the trajectory evolution points of different goods at different time periods to the same time resolution. Then, it calculates the spatial coordinates for several future time steps based on the spatial position evolution trend of each goods. The prediction employs a discrete iterative method based on continuous-time interpolation, introducing the previously calculated velocity and acceleration directions as basic variables at each time step. Dynamic extension is performed on the spatial motion path of each target, using spatial coordinates as the core and superimposing motion trend differences to obtain the position prediction results for subsequent time steps. For goods with rotation or tumbling tendencies, the aforementioned attitude change sequence is used to maintain a consistent rotation axis direction in the prediction, and the rotation angular velocity is iteratively updated to ensure the trajectory curve remains physically continuous in the time dimension. The final generated multi-target cargo scattering prediction trajectory contains a series of spatial coordinate points and direction data with temporal order, forming a complete three-dimensional prediction trajectory sequence, providing accurate external references for dynamic planning of automatic vehicle avoidance.

[0048] Step S233 includes:

[0049] Calculate the ambient wind speed under the aforementioned environmental wind conditions and determine the ambient wind direction;

[0050] Based on the environmental wind direction, a dynamic analysis of the force posture of the parabolic posture, the area of ​​the shape, and the curvature of the shape structure is performed during the cargo scattering process to obtain dynamic force posture data; wherein the force posture includes the force angle and force area of ​​the cargo affected by the environmental wind direction.

[0051] Based on the environmental wind speed, dynamic data of force posture and initial parabolic velocity of multi-target cargo, a spatial vector coupling calculation of the scattering velocity is performed to obtain the cargo spatial scattering velocity vector.

[0052] The disordered incremental gradient of kinetic energy is identified by spatial tumbling of the velocity vector of cargo scattered in space, and the disordered incremental gradient of kinetic energy is obtained.

[0053] The random process of cargo scattering location is decomposed based on the cargo spatial scattering velocity vector and the disordered incremental gradient of kinetic energy to obtain cargo scattering location decomposition data.

[0054] Based on the cargo scattering location decomposition data, the scattering trajectory evolution is fitted to output cargo scattering trajectory evolution data.

[0055] In this embodiment of the invention, wind speed and wind direction sensors continuously detect the current ambient airflow parameters on the highway. A moving average method is used to filter out high-frequency noise, resulting in a stable ambient wind speed value. The wind direction sensor, using a magnetoresistive structure, measures the airflow direction angle in real time and outputs the direction data through an angle encoder. Based on the longitudinal direction of vehicle travel, the system calculates the wind direction angle into three components: tailwind, headwind, and crosswind. Vector decomposition is used to obtain the lateral and longitudinal components of the wind field, forming environmental wind condition parameters. The final output ambient wind speed and direction are recorded with synchronized timestamps, providing accurate static input conditions for the direction and intensity of atmospheric motion in subsequent dynamic force and attitude analysis.

[0056] The system performs dynamic force-posture analysis on the parabolic attitude, area, and curvature of cargo during its scattering process based on ambient wind direction. First, it reads the parabolic attitude data of each cargo target in three-dimensional space, including pitch, roll, and yaw angles. The system uses ambient wind direction data to calculate the angle between the wind vector and the cargo surface to determine the force angle relative to the cargo's main surface. Then, it quantifies the degree of force influence based on the area and curvature data; a larger area indicates a more significant force per unit area, and a larger curvature indicates stronger airflow characteristics, leading to different aerodynamic force distributions in different directions. To describe the temporal evolution of the force-posture, the system calculates the rate of change of the force angle and the rate of centroid shift of the force area at multiple time points, obtaining a continuous sequence of force-posture changes. Finally, the calculated force angles, force areas, and their changing trends are summarized to form dynamic force-posture data. This data reflects the wind direction, force distribution, and attitude adjustment patterns of each cargo in a high-speed airflow field, which is then used for subsequent spatial velocity coupling calculations.

[0057] Based on environmental wind speed, dynamic force-attitude data, and the initial parabolic velocities of multiple cargo targets, a spatial vector coupling calculation of scattering velocity is performed. The initial parabolic velocity of each cargo target is calculated in previous steps and includes three-dimensional velocity components. The system first aligns the environmental wind speed vector with the initial cargo velocity vector and calculates the angle between them. Based on the force angle and force area changes in the dynamic force-attitude data, the influence of the wind speed vector on the contact surface is reconstructed, converting the aerodynamic influence into vector velocity increments. Subsequently, the wind speed increment vector and the initial parabolic velocity vector are progressively superimposed over time to form a velocity space coupling process. To maintain the continuity of velocity changes, the system uses linear interpolation at each time step to ensure the continuity of the coupling vector direction and magnitude. Through multiple iterative accumulations, the velocity components in each direction are finally merged into a complete three-dimensional cargo spatial scattering velocity vector, which reflects the instantaneous velocity direction and intensity of the cargo under the combined action of environmental wind force and its own dynamic attitude. After calculation, the cargo spatial scattering velocity vector is output, providing accurate dynamic input for subsequent trajectory evolution fitting.

[0058] The system performs kinetic energy disorder increment gradient identification for spatial tumbling of cargo scattering velocity vectors. First, it reads the three-dimensional spatial scattering velocity vector data obtained in the previous step, converting the velocity change trajectory of each cargo target in continuous time frames into a time-series vector group. The velocity vector difference between adjacent time points is calculated using time-series analysis, extracting instantaneous energy change rate features, and separating and calibrating the three-dimensional components to distinguish the energy distribution of velocity in the longitudinal, lateral, and vertical directions. Then, the spatial tumbling trend of the cargo during scattering is derived through the relationship between angular velocity and linear velocity, and the velocity gradient change at each time node is calculated using the finite difference method. If the gradient change exceeds a set threshold, it is determined that there is a disordered kinetic energy increment in that stage, i.e., the velocity energy state has transitioned from a stable state to an unstable state. The system clusters the velocity gradient change results of all time nodes, establishing a multi-dimensional vector cluster set. By analyzing the cluster spacing and difference, the temporal and spatial distribution of the disordered energy increment region is confirmed. The entire process is completed using time-frequency characteristic analysis. After identifying the asymmetry and random disturbances in energy distribution, continuous kinetic energy disordered incremental gradient data are output, forming a quantitative description of the degree of cargo tumbling instability and the range of energy distribution.

[0059] The three-dimensional scattered velocity vector is decomposed into a velocity scalar and a unit direction vector. To describe the random fluctuations in the force direction of the object, the disordered incremental gradient of kinetic energy is introduced as a perturbation factor into the calculation of the direction component. The velocity direction at each time point is biased and corrected to simulate the real changes under the influence of air disturbance and tumbling energy. Multiple independent iterative samplings are performed on the vector direction component. By establishing a mapping relationship between the time step and the perturbation amplitude, the motion direction angle is updated in each sampling, generating a series of randomized direction vector sets. To ensure the physical continuity of the scattered azimuth changes, the direction vectors of adjacent time steps are smoothed and filtered to eliminate jump noise. Then, the corrected direction vector and velocity scalar are recombined into an adjusted scattered velocity set, and its spatial distribution azimuth angle is calculated. By calculating the change path of the azimuth angle in the spatial spherical coordinate system, the direction changes after random perturbation are mapped into an azimuth deviation sequence. The system ultimately outputs cargo scattering orientation decomposition data, which reflects the scattering directionality and orientation shift characteristics of each target under the influence of kinetic energy disorder, providing input reference for continuous fitting of trajectory evolution.

[0060] First, a time-series coordinate axis is established. Using the azimuth decomposition data at each time step as the driving force, the displacement increment of the cargo in space at each moment is calculated. The system uses cubic interpolation to smooth the displacement increments between adjacent time points, ensuring the continuity of trajectory changes and curve smoothness. Then, combining the instantaneous values ​​of the aforementioned spatial scattered velocity vectors, the displacement increments are superimposed according to the time scale to generate a cargo position sequence. A moving average is performed on the sequence of each target to obtain a time-smoothed spatial trajectory curve. To further describe the rotational characteristics of the cargo motion, the system extracts the roll angle change rate from the attitude data at each time node and fine-tunes the trajectory curve in the corresponding direction. During trajectory generation, parallel computation is performed on the multi-target dataset to ensure the relative correspondence of different cargo trajectories on the same time axis. Finally, the trajectory point sets of all targets are merged in time to form three-dimensional scattered trajectory evolution data. This data records the movement path of the cargo over time under the influence of environmental wind force, force attitude, and disordered kinetic energy in the form of a spatial coordinate point sequence, providing a high-precision reference for subsequent avoidance trajectory prediction and behavior adjustment.

[0061] Step S24 includes the following steps:

[0062] Step S241: Based on the driving status of the intelligent vehicle and the predicted trajectory of the scattered goods, determine the dynamic relative radius between the current position of the intelligent vehicle and the scattered goods.

[0063] Step S242: Extract the vehicle speed in the driving state of the intelligent vehicle, and calculate the relative distance between the intelligent vehicle and surrounding vehicles using the vehicle's onboard radar and the intelligent vehicle's driving state.

[0064] Step S243: Adjust the dynamic speed of the intelligent vehicle according to the dynamic relative radius and the relative distance to ensure that the intelligent vehicle is at a safe distance from the scattered goods and surrounding vehicles.

[0065] Step S244: Perform avoidance trajectory attitude control based on the dynamic speed of the intelligent vehicle and the predicted trajectory of multi-target cargo scattering to obtain avoidance trajectory attitude control data;

[0066] Step S245: Adjust the automatic avoidance behavior based on the avoidance trajectory attitude control data to obtain automatic avoidance behavior adjustment data.

[0067] In this embodiment of the invention, the dynamic relative radius between the intelligent vehicle's current position and the scattered goods is determined based on the intelligent vehicle's driving status and the predicted trajectory data of the scattered goods. First, parameters such as the vehicle's longitudinal speed, yaw angle, steering wheel angle, and front-rear wheel distance are read from the vehicle's dynamic status dataset, while simultaneously extracting the vehicle's spatial coordinate data provided by the real-time positioning system. Then, these coordinates are matched with the spatial coordinates of each time step in the predicted trajectory of the scattered goods to calculate the three-dimensional Euclidean distance between the vehicle's center point and each predicted position of the goods. To determine the dynamic safety boundary, the system calculates the relative motion trend using the angle between the vehicle's running direction vector and the goods' moving direction vector, thereby obtaining the minimum relative radial distance between the vehicle and the goods. At this point, the rate of change is calculated using the distance sequence of continuous time frames to generate a dynamic relative radius curve. If the rate of decline of the curve exceeds a predetermined threshold, the goods are recorded in real time as potential collision targets. By parallel calculation of the radius data of all targets, a dynamic relative radius matrix for the current frame is formed. This matrix characterizes the relative motion boundary between the vehicle and all scattered goods in space in a time-series manner, providing a basis for subsequent dynamic avoidance adjustments.

[0068] The system extracts the vehicle speed from the intelligent vehicle's driving status and uses millimeter-wave radar devices installed on the front and side wings of the vehicle to acquire real-time position data of surrounding vehicles. The radar devices continuously emit high-frequency electromagnetic waves and receive reflected signals, calculating the distance information between the vehicle and obstacles in front and to the sides through phase difference. The system then synchronizes the vehicle speed data with the radar's returned distance data, calculating the relative distance and relative speed for each detected target. For surrounding vehicles, the system performs azimuth calculation on the returned signals using beam angle to obtain the lateral and longitudinal displacement components of the target vehicles, thereby determining their relative coordinate positions. The system then calculates the relative speed change trend through differential calculation of consecutive frames to assess the stability of distance changes. After spatial reconstruction, the system obtains the real-time relative distance matrix and relative speed matrix between the vehicle and all adjacent vehicles in the front, rear, left, and right directions. This dataset, combined with the dynamic vehicle speed output from the vehicle's driving status, is input into a distance safety assessment system to determine the critical distance threshold for subsequent dynamic speed control strategies.

[0069] The system reads the minimum safe distance data and corresponding time change rates around the vehicle in each direction to define dynamic speed constraint intervals. A longitudinal speed adjustment benchmark is established with the vehicle's direction of travel as the main axis. By comparing the dynamic relative radius change rate, if the distance between the cargo ahead and the vehicle enters the safe boundary, the deceleration ratio is calculated based on the distance descent rate. In the lateral direction, the relative distance to the nearest lateral obstacle is used as a constraint to determine the lateral offset space. To prevent steering instability caused by rapid deceleration, the system simultaneously adjusts the braking force distribution ratio during longitudinal deceleration based on the vehicle's wheel-to-wheel load distribution characteristics, achieving coordinated speed and direction control. An acceleration smoothing curve algorithm is used during speed adjustment to ensure continuous variation in the output throttle and braking control quantities. In each control cycle, the system re-matches the adjusted dynamic speed data with the vehicle's current position, checking in real-time whether the distances between the vehicle and the cargo and surrounding vehicles meet the safe boundary. Upon confirmation, a stable speed control state is output, providing a timing basis for obstacle avoidance attitude calculations.

[0070] The system uses the vehicle's steering angle, lateral acceleration, and dynamic velocity values ​​to determine the vector representation of the vehicle's motion state at the current moment. Then, at the same time step, it reads all predicted trajectory points of scattered goods to form a set of spatial coordinates. Potential intersection points between the vehicle's motion path and the goods' trajectory are calculated using affine geometry mapping. If the intersection points are distributed within 50 meters in the forward direction, the corresponding goods are marked as priority avoidance targets. For each priority target, the system calculates a steering correction angle based on the relative azimuth angle and distance to ensure that the vehicle remains in the center of the lane after trajectory deviation. To make the steering process dynamically smooth, vehicle yaw rate feedback is added to the control algorithm to adjust the steering wheel input amplitude in real time. Subsequently, dynamic velocity data is used to limit the steepness of the steering trajectory; the rate of change of the steering angle is reduced at higher speeds and increased at lower speeds. After multiple trajectory simulation iterations, avoidance trajectory attitude control data is generated. This data describes the vehicle's steering angle, vehicle attitude angle changes, and driving path coordinates, providing the execution basis for the final behavior adjustment.

[0071] Within the control cycle, the aforementioned attitude control commands are synchronously read, and the steering angle, vehicle speed correction value, and acceleration signal are transmitted to the vehicle chassis control system. The wheel angles are adjusted by precisely driving the steering actuator to achieve vehicle steering along the planned trajectory. Simultaneously, the braking system distributes braking torque according to the speed correction value, ensuring stable deceleration during the avoidance process. The control unit calculates new vehicle posture and yaw rate at each time cycle; if there is a deviation from the predicted posture, dynamic fine-tuning is performed based on the deviation signal. The entire process iterates periodically, using real-time closed-loop control to ensure the vehicle's spatial motion trajectory matches the preset avoidance trajectory. Furthermore, the vehicle avoidance effect is verified using cargo scattering prediction trajectory data. If the vehicle trajectory re-enters the safe zone, the system outputs final automatic avoidance behavior adjustment data, recording the speed change, posture change, and path deviation results of this avoidance behavior, providing basic behavioral samples for subsequent strategy learning and optimization stages.

[0072] Step S244 includes the following steps:

[0073] Based on the dynamic speed of intelligent vehicles and the predicted trajectory of multi-objective cargo scattering, the dynamic radius of potential contact points in space is analyzed to obtain the dynamic radius of potential contact points.

[0074] Based on the dynamic radius of the potential contact point, determine the adjustment range of the avoidance yaw rate and sideslip angle;

[0075] Based on the adjustment range and the dynamic radius of the potential contact point, adaptive steering angle constraints are applied to obtain adaptive steering angle adjustment data.

[0076] Linear power assist regression was performed on the adaptive steering angle adjustment data to obtain steering assist fitting data.

[0077] Based on the adjustment range, adaptive adjustment data of steering angle, and steering assist fitting data, avoidance trajectory attitude control is performed to obtain avoidance trajectory attitude control data.

[0078] In this embodiment of the invention, longitudinal velocity, lateral acceleration, and vehicle heading angle data at the current moment are retrieved from the vehicle's operating status set. Simultaneously, the three-dimensional coordinate points of each target in the predicted cargo scattering trajectory are extracted at the same moment. By establishing a unified reference frame between the vehicle's position coordinate system and the cargo trajectory coordinate system, the vehicle's movement path is divided into continuous spatial nodes according to the time step. At each time node, the system calculates the angle between the vehicle's travel direction and the cargo's movement direction, and estimates the minimum geometric intersection distance between the two paths by combining the velocity components of both. The rate of change of the minimum distance sequence within the continuous time period is taken to obtain the dynamic proximity of potential spatial contact points. This rate of change is used for threshold analysis to determine whether a high-risk contact zone has been entered. When the change exceeds a preset safety limit, a dynamic radius value of the potential contact point is generated in real time to describe the size of the intersection area between the vehicle and the scattered cargo. All calculation results are stored in chronological order, forming a continuous dynamic radius sequence of potential contact points, providing geometric measurement parameters for subsequent angle adjustments of the avoidance posture.

[0079] By employing vehicle dynamics equilibrium relationships and using the dynamic radius of the potential contact point as the core input parameter, the rate of change of the vehicle's lateral distance is calculated. When the dynamic radius decreases and falls below a specific safety threshold, the system determines a correction coefficient for the yaw rate using a yaw control algorithm. The next step combines real-time vehicle speed and tire lateral force data to calculate the corresponding trend of the slip angle change, determining the upper limit of the slip angle adjustment based on the contraction rate of the potential contact point radius. By adjusting the yaw rate and slip angle in combination, the vehicle's lateral attitude correction is formed. To ensure the continuity of the adjustment range, piecewise linear interpolation is used to smooth the yaw rate adjustment curve, thereby generating adjustment range data with temporally continuous characteristics. This data records the vehicle's directional response range under specific dynamic radius conditions, as well as the maximum yaw rate limit matched to the speed, providing a basic input for subsequent adaptive constraint calculations.

[0080] Using the front wheel steering angle as the primary control variable, the system reads the yaw rate and sideslip angle adjustment values ​​from the previous cycle to establish an angle constraint relationship. Subsequently, the dynamic radius of the potential contact point is mapped to an angle adjustment range limit factor; when the radius decreases, the allowable steering range is reduced, and when the radius increases, the steering range is widened. The system determines the adaptive constraint curve through a real-time range mapping table and continuously tracks the rate of change of the angle output, ensuring a smooth transition of the actual steering angle within a safe dynamic range. In each update, the system calculates the current angle deviation differentially and corrects the control signal in real time to prevent sudden increases in steering angle from causing vehicle yaw instability. After multiple iterations of calculation, adaptive steering angle adjustment data is output, including instantaneous angle limits, angle change rate, and directional stability weights, ensuring that the vehicle maintains a restricted and safe steering response during obstacle avoidance maneuvers.

[0081] Linear power assist regression fitting is performed on the aforementioned adaptive steering angle adjustment data to obtain steering assist fitting data. The process uses the actual torque response value fed back by the vehicle's steering actuator and the adaptive angle adjustment data as input. First, the relationship between angle change and torque output is calculated. Then, a set of power assist adjustment coefficients is generated using a linear regression algorithm to describe the power assist intensity distribution corresponding to different angle change rates. To ensure smooth power assist output, the regression residuals are filtered to eliminate discrete noise and retain the trend characteristics over continuous time periods. The final steering assist fitting data reflects the matching law between the vehicle's steering wheel rotation force and the angle change rate, ensuring a dynamic balance between the vehicle's steering sensitivity and stability under different road conditions.

[0082] The system comprehensively controls the avoidance trajectory attitude based on the adjustment range, adaptive steering angle adjustment data, and steering assist fitting data. First, all input parameters are read within a unified time frame, and the comprehensive proportional relationship between longitudinal velocity, steering angle, and wheel-end torque distribution is calculated. Then, based on the assist fitting data, the optimal torque is distributed to the front and rear steering mechanisms, enabling the vehicle to achieve a smooth attitude adjustment process within a specified yaw rate and sideslip angle range. Simultaneously, the rate of directional change is corrected in real time based on adaptive angle adjustment data to prevent deviation from the turning path due to excessive angle. During control, the system calculates the offset error by comparing the directional angular velocity with the vehicle's center of gravity trajectory and implements feedback correction for attitude angle changes. Finally, the system outputs avoidance trajectory attitude control data, which fully describes the spatial attitude parameter sequence, wheel-end force adjustment ratio, and steering assist response mode of the vehicle when avoiding cargo, providing precise control basis for subsequent avoidance behavior adjustments.

[0083] Step S3 includes the following steps:

[0084] Step S31: Perform avoidance feature learning on the automatic avoidance behavior adjustment data to obtain avoidance behavior feature learning data;

[0085] Step S32: Perform behavioral feature weighting processing on the avoidance behavior feature learning data to output avoidance behavior feature weighted data;

[0086] Step S33: Construct an automatic avoidance strategy model using weighted data of avoidance behavior features;

[0087] Step S34: Verify and optimize the automatic avoidance strategy model to obtain the optimized automatic avoidance strategy model;

[0088] Step S35: Send the automatic avoidance strategy optimization model to the intelligent vehicle control center to execute the automatic avoidance of the intelligent vehicle.

[0089] In this embodiment of the invention, avoidance feature learning is performed on the automatic avoidance behavior adjustment data. Time-series elements are extracted from the behavior adjustment data recorded during the automatic avoidance process, including basic dynamic parameters such as vehicle longitudinal speed, yaw rate, sideslip angle, steering wheel angular velocity, and vehicle attitude change values. Each set of data is stamped with a timestamp and execution cycle identifier. The data preprocessing unit segments the continuous behavior sequence in the time domain according to a fixed sampling interval to form a windowed sample set. Then, normalization processing is used to map physical variables of different dimensions to a unified scale, eliminating numerical differences. Multidimensional feature extraction is performed on the sample set, and principal component analysis is used to extract key components representing the vehicle's dynamic avoidance characteristics, such as longitudinal acceleration fluctuation rate, steering angle amplitude change rate, and vehicle attitude stability index. The processed feature vectors form a feature trajectory sequence in continuous time. After integration and synchronization, an avoidance behavior feature learning dataset is generated, providing a quantitative expression for subsequent weighted processing and strategy model construction.

[0090] The correlation coefficient is calculated for each feature component in the avoidance feature learning dataset. Correlation analysis is performed to screen the main factors affecting avoidance accuracy by calculating the cross-correlation between velocity change rate, direction change time delay, and attitude variability. Subsequently, corresponding weights are assigned based on the statistical contribution of features and prior empirical coefficients. To prevent any single feature from having an excessively high weight, a standardized weight constraint mechanism is used to redistribute all weights, ensuring the total weight remains within a standardized unit. The system performs numerical synthesis on the weighted feature set, merging multidimensional data within the same time window into a single high-order feature vector, and reducing the impact of local noise through weighted averaging. Finally, weighted data of avoidance behavior features is generated, where each time slice contains feature response parameters reflecting the comprehensive dynamic laws, which can be directly used in the strategy model construction stage.

[0091] An automatic obstacle avoidance strategy model is constructed using weighted data of the aforementioned obstacle avoidance behavior features. The operation is based on a training and validation set partitioning, selecting 80% of the samples in chronological order as training input and the remaining samples as validation input. Multi-layer feature convolution operations are performed on the training input to extract high-order spatiotemporal correlation features, obtaining composite dynamic feedback features during the obstacle avoidance process. The convolutional feature results are then input into a decision tree ensemble structure, where a random forest algorithm is used to achieve clustered classification and policy branching. The algorithm calculates the minimum information gain at each sample node and branches according to the weighted features, continuously building multi-layered decision paths. The system automatically generates policy output nodes in the tree structure, corresponding to the control command arrangement during vehicle dynamic obstacle avoidance, including steering adjustment values, acceleration constraint ratios, and vehicle offset corrections. A comprehensive output is generated through inter-tree voting, resulting in a preliminary automatic obstacle avoidance strategy model, achieving a one-to-one mapping between data input and automatic response rules.

[0092] The automatic obstacle avoidance strategy model is validated and optimized. The optimization process begins by inputting validation set feature data into the constructed strategy model to generate predicted obstacle avoidance control results. Then, the prediction error is calculated by comparing the validation results with the actual obstacle avoidance execution data. Based on the error distribution pattern, the system performs gradient correction, readjusting the feature weight distribution of the decision path nodes with the largest deviations. An iterative correction mechanism is then applied to gradually reduce the error through multiple training and validation iterations. To prevent premature convergence leading to computational stagnation, a momentum update strategy is introduced to maintain a dynamic balance between the strategy convergence speed and the error correction intensity. The final output strategy model, after error iterative training and parameter balancing, forms the optimized automatic obstacle avoidance strategy model. This model retains hierarchical feature weights and dynamic feedback paths in its structure, possesses stable output characteristics, and can directly generate stable and reliable obstacle avoidance control commands. It's important to clarify that the validation optimization here occurs after model building and is part of the dynamic retraining and parameter correction process. It not only detects the deviation between the model output and the actual avoidance behavior but also systematically optimizes the model's internal weights, branch paths, and feature contributions through gradient correction, error distribution analysis, and momentum adjustment mechanisms. This allows the model to converge continuously and enhance its generalization ability through multiple iterations. In contrast, the error validation in step S334 is a simulation test during the model building phase. Its main purpose is to verify the model's basic adaptability and prediction accuracy by inputting an avoidance behavior feature test set after the initial automatic avoidance strategy model is established. This is used to identify basic problems such as classification bias, dynamic response delay, or feature input errors; it is a static model performance test. Therefore, step S334 focuses on basic accuracy validation, while step S34 emphasizes dynamic optimization of overall performance and model self-learning improvement; these two steps are distinct.

[0093] The optimized automatic avoidance strategy model is output to the intelligent vehicle control center for execution. The model's data structure is translated to ensure its parameter format aligns with the vehicle control command format. Subsequently, the dynamic parameters output by the model are mapped in real-time to the vehicle control layer via the main control bus, including wheel steering angle commands, brake pressure setpoints, and acceleration valve opening control signals. The control center issues continuous control commands to the vehicle chassis execution unit according to the input model commands, enabling the vehicle to perform automatic avoidance maneuvers in real-time while sensing the cargo scattering trajectory. All command execution results are then re-input into the data acquisition system via a feedback loop, forming a closed-loop real-time control structure. Simultaneously, execution data is recorded to provide a sample library for subsequent learning and updates. Ultimately, the real-time deployment and execution of the automatic avoidance strategy optimization model are achieved, ensuring the intelligent vehicle can stably complete avoidance operations in sudden cargo scattering environments.

[0094] Step S33 includes the following steps:

[0095] Step S331: Select avoidance behavior features from the weighted data of avoidance behavior features, and then divide the training set and test set into avoidance behavior feature test set and avoidance behavior feature training set respectively;

[0096] Step S332: Perform convolution processing on the avoidance behavior feature training set to obtain the avoidance feature convolution training set;

[0097] Step S333: Construct an initial automatic obstacle avoidance strategy model using the random forest algorithm on the obstacle avoidance feature convolution training set to obtain the initial automatic obstacle avoidance strategy model;

[0098] Step S334: Input the avoidance behavior feature test set into the initial automatic avoidance strategy model to perform simulation testing and obtain the automatic avoidance strategy model.

[0099] In this embodiment of the invention, avoidance behavior features are selected from weighted data, and training and testing sets are divided. The process first identifies the core attribute columns in the weighted data, including key feature parameters such as longitudinal acceleration rate of change, lateral acceleration rate of change, steering wheel angular velocity, vehicle yaw rate, vehicle body slip angle, and speed fluctuation intensity. Then, each feature is standardized according to the principle of temporal continuity, mapping the numerical range of each data category to the same distribution range. For the multi-dimensional feature vectors, correlation coefficient analysis is used to calculate the correlation between each feature parameter and the avoidance response result. The top few sets of feature variables with the highest correlation are selected in descending order as the training feature set to ensure the representativeness of the input data. After feature selection, all weighted samples are divided according to the time axis proportion, with 80% used for training and 20% for testing, generating avoidance behavior feature training and test sets. After division, cross-window segmentation is performed on the sample boundaries to retain a certain overlap at time transitions between adjacent data intervals, ensuring that the model can recognize the temporal extension characteristics of avoidance behavior during training.

[0100] A convolutional processing method is applied to the obstacle avoidance behavior feature training set to obtain an obstacle avoidance feature convolutional training set. The convolution operation takes the feature time series as input and performs multi-layer local convolution operations within a fixed time window. First, a feature matrix is ​​established for each sample time period, using physical features such as vertical, horizontal, and directional changes as row vectors and time sampling points as column vectors to form a two-dimensional data structure. Then, a two-dimensional convolution kernel is applied sequentially to this matrix, extracting the coupling pattern between time and features through local weighted superposition, thereby obtaining the local response representing the instantaneous obstacle avoidance dynamics of the vehicle. After convolution, normalization is performed to eliminate amplitude differences between feature values ​​in different channels. For the multi-channel convolution output, pooling is performed to reduce redundant dimensions and retain representative response regions with concentrated energy. This convolutional processing extracts the temporal feature correlations in obstacle avoidance behavior, structurally forming a feature representation with temporal continuity. The output obstacle avoidance feature convolutional training set is stored for subsequent model construction.

[0101] An initial automatic obstacle avoidance strategy model is constructed using the random forest algorithm on the convolutional training set of obstacle avoidance features. This process begins by using each feature sample in the convolutional training set as input to determine the probability distribution between each feature and the control decision. During construction, the random forest algorithm randomly selects a subset of the feature set and a subset of the sample set to generate multiple decision trees. Each decision tree calculates a feature threshold through node splitting and determines the optimal splitting path based on the Shannon information gain criterion to distinguish different obstacle avoidance behavior patterns. After each decision tree is trained independently, it forms a mapping path from features to control results, outputting discrete labels such as steering angle adjustment, speed control amount, and attitude correction coefficients. After training all trees, a voting mechanism is used to perform weighted aggregation of the outputs of each tree to obtain the average probability distribution of the comprehensive prediction results. Finally, all decision paths constitute a complete initial automatic obstacle avoidance strategy model, reflecting the obstacle avoidance control rules and action command configurations that the vehicle should execute under different dynamic feature input conditions.

[0102] The obstacle avoidance behavior feature test set is input into the initial automatic obstacle avoidance strategy model to perform simulation testing and generate the automatic obstacle avoidance strategy model results. The simulation testing process first reads the trained forest structure parameters, then inputs the test set into each decision tree to calculate the classification probability. Each test sample outputs a predicted behavioral response after passing through multiple split nodes. The measured obstacle avoidance results of the test samples are compared with the model output results, and the prediction deviation distribution is calculated using an error function. Static feature errors and dynamic response errors are statistically analyzed separately. Through multiple iterations of input comparison, the stability of the model under different vehicle speeds, yaw rates, and angular accelerations is analyzed. For samples with large errors, their corresponding feature weights are re-evaluated to determine whether model parameters need to be adjusted in subsequent optimization stages. The test results are output in a statistical table format, including indicators such as classification accuracy, average deviation value, and decision time delay. Once all results are structured and stored, a simulation-validated automatic obstacle avoidance strategy model is formed, providing a reliable reference basis for entering the model validation and optimization stage.

[0103] The present invention also provides an automatic obstacle avoidance system for an intelligent vehicle, for performing the automatic obstacle avoidance method for an intelligent vehicle as described above, the automatic obstacle avoidance system for the intelligent vehicle comprising:

[0104] The scattered cargo marking module is used to acquire the driving status of the intelligent vehicle at high speed and monitor video frames of scattered cargo in the lane ahead through the vehicle-mounted camera; based on the video frames, multi-target marking of the scattered cargo is performed to obtain multi-target marking data of the scattered cargo.

[0105] The obstacle avoidance behavior adjustment module is used to predict the trajectory of scattered goods based on the multi-target marking data of scattered goods, and then adjust the automatic obstacle avoidance behavior according to the driving status of the intelligent vehicle to obtain automatic obstacle avoidance behavior adjustment data.

[0106] The obstacle avoidance model building module is used to build an automatic obstacle avoidance strategy model through automatic obstacle avoidance behavior adjustment data, and send the automatic obstacle avoidance strategy model to the intelligent vehicle control center to execute the automatic obstacle avoidance of the intelligent vehicle.

[0107] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. An automatic obstacle avoidance method for intelligent vehicles, characterized in that, Includes the following steps: Step S1: Obtain the driving status of the intelligent vehicle at high speed and monitor video frames of scattered goods in the lane ahead using the vehicle-mounted camera; perform multi-target marking of the scattered goods based on the video frames to obtain multi-target marking data of the scattered goods. Step S2: Based on the multi-target marking data of scattered goods, predict the trajectory of scattered goods, and then adjust the automatic avoidance behavior according to the driving status of the intelligent vehicle to obtain the automatic avoidance behavior adjustment data. Step S3: Construct an automatic avoidance strategy model using automatic avoidance behavior adjustment data, and send the automatic avoidance strategy model to the intelligent vehicle control center to execute the automatic avoidance of the intelligent vehicle; Step S2 includes: Step S21: Perform geometric morphology analysis on the scattered cargo multi-target marking data to obtain multi-target cargo geometric morphology data; Step S22: Perform initial parabolic state analysis on the scattered cargo multi-target marking data to obtain the initial parabolic state of the multi-target cargo; Step S23: Based on the geometric shape data of the multi-objective cargo and the initial parabolic state of the multi-objective cargo, predict the scattering trajectory of the multi-objective cargo to output the predicted scattering trajectory of the multi-objective cargo. Step S24: Adjust the automatic avoidance behavior based on the driving status of the intelligent vehicle and the predicted trajectory of multi-target cargo scattering to obtain automatic avoidance behavior adjustment data; Step S23 includes: Step S231: Extract the morphological area and morphological curvature from the geometric morphology data of multi-target cargo; obtain the current environmental wind conditions on the highway; Step S232: Analyze the initial drop velocity and attitude of the multi-target cargo in its initial parabolic state to obtain the initial parabolic velocity and parabolic attitude of the multi-target cargo; Step S233: Based on the environmental wind conditions, fit the scattering trajectory evolution of the shape area, shape structure curvature, initial parabolic velocity and parabolic attitude of the multi-target cargo to output cargo scattering trajectory evolution data; Step S234: Perform multi-target cargo scattering trajectory prediction based on cargo scattering trajectory evolution data, and output multi-target cargo scattering prediction trajectory; Step S24 includes: Step S241: Based on the driving status of the intelligent vehicle and the predicted trajectory of the scattered goods, determine the dynamic relative radius between the current position of the intelligent vehicle and the scattered goods. Step S242: Extract the vehicle speed in the driving state of the intelligent vehicle, and calculate the relative distance between the intelligent vehicle and surrounding vehicles using the vehicle's onboard radar and the intelligent vehicle's driving state. Step S243: Adjust the dynamic speed of the intelligent vehicle according to the dynamic relative radius and the relative distance to ensure that the intelligent vehicle is at a safe distance from the scattered goods and surrounding vehicles. Step S244: Perform avoidance trajectory attitude control based on the dynamic speed of the intelligent vehicle and the predicted trajectory of multi-target cargo scattering to obtain avoidance trajectory attitude control data; Step S245: Adjust the automatic avoidance behavior based on the avoidance trajectory attitude control data to obtain automatic avoidance behavior adjustment data; Step S244 includes: Based on the dynamic speed of intelligent vehicles and the predicted trajectory of multi-objective cargo scattering, the dynamic radius of potential contact points in space is analyzed to obtain the dynamic radius of potential contact points. Based on the dynamic radius of the potential contact point, determine the adjustment range of the avoidance yaw rate and sideslip angle; Based on the adjustment range and the dynamic radius of the potential contact point, adaptive steering angle constraints are applied to obtain adaptive steering angle adjustment data. Linear power assist regression was performed on the adaptive steering angle adjustment data to obtain steering assist fitting data. Based on the adjustment range, adaptive adjustment data of steering angle, and steering assist fitting data, avoidance trajectory attitude control is performed to obtain avoidance trajectory attitude control data.

2. The automatic obstacle avoidance method for intelligent vehicles according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Obtain the driving status of the intelligent vehicle at high speed and monitor video frames of goods scattered in the lane ahead using the onboard camera; Step S12: Sharpen the video frames to obtain video sharpening data; Step S13: Divide the video sharpening data into a regional grid to obtain video region division data; Step S14: Based on the video region segmentation data, perform multi-target marking of scattered goods to obtain multi-target marking data of scattered goods.

3. The automatic obstacle avoidance method for intelligent vehicles according to claim 1, characterized in that, Step S233 includes: Calculate the ambient wind speed under the aforementioned environmental wind conditions and determine the ambient wind direction; Based on the environmental wind direction, a dynamic analysis of the force posture of the parabolic posture, the area of ​​the shape, and the curvature of the shape structure is performed during the cargo scattering process to obtain dynamic force posture data; wherein the force posture includes the force angle and force area of ​​the cargo affected by the environmental wind direction. Based on the environmental wind speed, dynamic data of force posture and initial parabolic velocity of multi-target cargo, a spatial vector coupling calculation of the scattering velocity is performed to obtain the cargo spatial scattering velocity vector. The disordered incremental gradient of kinetic energy is identified by spatial tumbling of the velocity vector of cargo scattered in space, and the disordered incremental gradient of kinetic energy is obtained. The random process of cargo scattering location is decomposed based on the cargo spatial scattering velocity vector and the disordered incremental gradient of kinetic energy to obtain cargo scattering location decomposition data. Based on the cargo scattering location decomposition data, the scattering trajectory evolution is fitted to output cargo scattering trajectory evolution data.

4. The automatic obstacle avoidance method for intelligent vehicles according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Perform avoidance feature learning on the automatic avoidance behavior adjustment data to obtain avoidance behavior feature learning data; Step S32: Perform behavioral feature weighting processing on the avoidance behavior feature learning data to output avoidance behavior feature weighted data; Step S33: Construct an automatic avoidance strategy model using weighted data of avoidance behavior features; Step S34: Verify and optimize the automatic avoidance strategy model to obtain the optimized automatic avoidance strategy model; Step S35: Send the automatic avoidance strategy optimization model to the intelligent vehicle control center to execute the automatic avoidance of the intelligent vehicle.

5. The automatic obstacle avoidance method for intelligent vehicles according to claim 4, characterized in that, Step S33 includes the following steps: Step S331: Select avoidance behavior features from the weighted data of avoidance behavior features, and then divide the training set and test set into avoidance behavior feature test set and avoidance behavior feature training set respectively; Step S332: Perform convolution processing on the avoidance behavior feature training set to obtain the avoidance feature convolution training set; Step S333: Construct an initial automatic obstacle avoidance strategy model using the random forest algorithm on the obstacle avoidance feature convolution training set to obtain the initial automatic obstacle avoidance strategy model; Step S334: Input the avoidance behavior feature test set into the initial automatic avoidance strategy model to perform simulation testing and obtain the automatic avoidance strategy model.

6. An automatic obstacle avoidance system for an intelligent vehicle, characterized in that, For performing the automatic obstacle avoidance method for an intelligent vehicle as described in claim 1, the automatic obstacle avoidance system of the intelligent vehicle includes: The scattered cargo marking module is used to acquire the driving status of the intelligent vehicle at high speed and monitor video frames of scattered cargo in the lane ahead through the vehicle-mounted camera; based on the video frames, multi-target marking of the scattered cargo is performed to obtain multi-target marking data of the scattered cargo. The obstacle avoidance behavior adjustment module is used to predict the trajectory of scattered goods based on the multi-target marking data of scattered goods, and then adjust the automatic obstacle avoidance behavior according to the driving status of the intelligent vehicle to obtain automatic obstacle avoidance behavior adjustment data. The obstacle avoidance model building module is used to build an automatic obstacle avoidance strategy model through automatic obstacle avoidance behavior adjustment data, and send the automatic obstacle avoidance strategy model to the intelligent vehicle control center to execute the automatic obstacle avoidance of the intelligent vehicle.

Images