Intelligent vehicle multi-modal perception collaborative obstacle avoidance control method

Abstract

The present application relates to the technical field of obstacle avoidance control, and particularly relates to a multi-modal perception cooperative obstacle avoidance control method for intelligent vehicles. The method comprises the following steps: collecting and preprocessing perception data, extracting features from the preprocessed perception data, and then uniformly processing the features; introducing a high-dimensional spatiotemporal manifold embedding algorithm to map the uniformly processed feature data, obtain a high-dimensional manifold embedding vector, and construct a global perception vector; based on the global perception vector, dividing a candidate obstacle region and calculating a disturbance risk score thereof; based on the disturbance risk score of the candidate obstacle region, generating an obstacle avoidance trajectory in real time, generating a control decision signal, and realizing cooperative obstacle avoidance control. The method solves the problem that, in the process of intelligent vehicle operation, multi-source sensor data is inconsistent in time delay, spatial coordinates, data dimension and distribution structure, leading to inaccurate obstacle identification, slow path obstacle avoidance response and frequent control instruction jumps.

Description

Technical Field

[0001] This invention relates to the field of obstacle avoidance control technology, and in particular to a multimodal perception and collaborative obstacle avoidance control method for intelligent vehicles. Background Technology

[0002] With the rapid development of autonomous driving and intelligent robot technologies, intelligent vehicles are increasingly being used in unstructured scenarios such as warehousing and logistics, industrial automation, and urban delivery. In actual operation, the environment in which intelligent vehicles operate typically has unstructured characteristics such as complex obstacle types, irregular spatial structures, and frequent crossings by pedestrians and other moving targets. Therefore, they need to have strong environmental perception capabilities and dynamic obstacle avoidance capabilities.

[0003] However, existing intelligent vehicle perception and obstacle avoidance control solutions still face the following bottlenecks in practical applications: Currently, most mainstream platforms rely on single or low-fusion sensor systems such as LiDAR, cameras, or ultrasonic sensors, lacking deep fusion processing of multimodal sensor data. This makes it impossible to achieve precise modeling of obstacle attributes, behaviors, and motion trends. Especially when multi-source data exhibits time asynchrony, scale inconsistency, sampling density differences, and spatial reference frame conflicts, it often leads to the accumulation of perception errors and blurred target recognition, thereby affecting the accuracy of obstacle avoidance strategies and the stability of the control system. In addition, existing path planning methods often employ static path perturbation or simple obstacle avoidance strategies, failing to adjust in real time according to the dynamic behavior changes and potential risks of obstacles. They lack local harmonization and path reconstruction capabilities, easily leading to path oscillations, trajectory abrupt changes, or obstacle avoidance failures. At the same time, traditional control systems often use fixed-structure PID or single-channel closed-loop models, failing to fully utilize high-dimensional perception information for flexible control adjustment. This results in limitations on the driving stability, control continuity, and behavioral interpretability of intelligent vehicles in complex dynamic scenarios. Summary of the Invention

[0004] This invention provides a multimodal perception and collaborative obstacle avoidance control method for intelligent vehicles to solve the technical problems of inconsistencies in time delay, spatial coordinates, data dimensions and distribution structure of multi-source sensor data during intelligent vehicle operation, which lead to inaccurate obstacle recognition, slow path obstacle avoidance response and frequent control command jumps.

[0005] The present invention provides a multimodal perception and cooperative obstacle avoidance control method for intelligent vehicles, specifically including the following technical solutions: A multimodal perception and cooperative obstacle avoidance control method for intelligent vehicles includes the following steps: S1. Collect and preprocess the sensing data to obtain preprocessed sensing data; extract features from the preprocessed sensing data to obtain preliminary feature data, and perform unified processing to obtain unified feature data; introduce a high-dimensional spatiotemporal manifold embedding algorithm to map the unified feature data to obtain a high-dimensional manifold embedding vector, and construct a global sensing vector. S2. Based on the global perception vector, spatial regions are divided to obtain candidate obstacle regions, and disturbance risk scores of candidate obstacle regions are calculated; based on the disturbance risk scores of candidate obstacle regions, obstacle avoidance trajectories are generated in real time to obtain the target trajectory function; based on the target trajectory function, control decision signals are generated and mapped into control commands to achieve cooperative obstacle avoidance control.

[0006] Preferably, S1 specifically includes: The sensing data includes lidar point cloud data, RGB image data, infrared thermal imaging data, and ultrasonic distance measurement data. Feature engineering techniques are used to extract features from the preprocessed sensing data to obtain preliminary feature data. Multimodal feature alignment and distribution regularization techniques are then used to uniformly process the preliminary feature data to obtain uniformly processed feature data.

[0007] Preferably, S1 specifically includes: In the implementation of the high-dimensional spatiotemporal manifold embedding algorithm, based on the uniformly processed feature data, a sliding integral principal term and a local gradient projection correction term are generated. All uniformly processed feature data are then projected onto the joint non-Euclidean manifold space to obtain the high-dimensional manifold embedding vector.

[0008] Preferably, S1 specifically includes: The high-dimensional manifold embedding vectors of all sensors are concatenated to obtain the global sensing vector.

[0009] Preferably, S2 specifically includes: Multidimensional physical attributes are extracted and normalized from the subset point group of each candidate obstacle region. Then, composite nonlinear fusion is performed on the multidimensional physical attributes to obtain the disturbance risk score of the candidate obstacle region.

[0010] Preferably, S2 specifically includes: Based on the disturbance risk score of the candidate obstacle region, combined with the Gaussian attenuation term, a disturbance amplitude coefficient is constructed, and a frequency harmonic disturbance term is generated.

[0011] Preferably, S2 specifically includes: Define an initial path function and construct a curvature contraction term. Combine this with a frequency harmonic perturbation term to perform local trajectory reconstruction and obtain the target trajectory function.

[0012] Preferably, S2 specifically includes: Based on the target trajectory function, the expected velocity and tangent direction angle of the target path are calculated, and velocity error terms and angle error terms are constructed. Combined with obstacle avoidance and braking terms, a nonlinear control function is constructed to generate control decision signals. The control decision signals are then mapped to obtain control commands.

[0013] The beneficial effects of the technical solution of the present invention are: 1. By introducing multimodal sensors to acquire sensing data, and through preprocessing methods such as noise reduction, synchronization, standardization and normalization, the consistency of different data sources in terms of time, space and statistical characteristics is significantly improved, avoiding the problem of sensing error accumulation. By introducing a high-dimensional spatiotemporal manifold embedding algorithm, the feature data of different sensors after unified processing are projected onto a non-Euclidean manifold space, and a sensing manifold mapping model based on sliding integral and local gradient projection is constructed. This effectively eliminates the differences between sensors in terms of dimension, scale, sampling frequency and spatial reference frame, thereby enhancing the robustness and generalization ability of target recognition and obstacle perception in complex scenes.

[0014] 2. A disturbance risk scoring formula is introduced, which integrates the obstacle's spatial occupancy rate, boundary ambiguity, orientation, dynamic change rate, and boundary fluctuation. It is weighted and evaluated in the form of a composite nonlinear function. This not only accurately identifies high-threat obstacles directly in front, but also identifies potential moving targets (such as pedestrians, animals, or other intelligent vehicles) through boundary uncertainty and dynamic change rate. It achieves efficient risk estimation without relying on a classifier, providing a theoretical basis for dynamic path avoidance.

[0015] 3. The constructed target trajectory function introduces a trajectory deformation mechanism that combines frequency harmonic perturbation terms and curvature contraction terms. In scenarios where obstacles intervene in the main path, it can generate asymmetric, multi-frequency, and dynamic perturbation paths in real time. This not only enables flexible obstacle avoidance actions but also effectively suppresses violent steering in high curvature regions through curvature contraction terms, preventing vehicles from slipping, turning sharply, and other unstable behaviors, thereby improving the controllability and smoothness of the path. Attached Figure Description

[0016] Figure 1 This is a flowchart of a multimodal perception and collaborative obstacle avoidance control method for intelligent vehicles according to the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a multimodal perception and collaborative obstacle avoidance control method for intelligent vehicles provided by this invention.

[0020] See attached document Figure 1 The diagram illustrates a flowchart of a multimodal perception and cooperative obstacle avoidance control method for intelligent vehicles according to an embodiment of the present invention. The method includes the following steps: S1. Collect and preprocess the sensing data to obtain preprocessed sensing data; extract features from the preprocessed sensing data to obtain preliminary feature data, and perform unified processing to obtain unified feature data; introduce a high-dimensional spatiotemporal manifold embedding algorithm to map the unified feature data to obtain a high-dimensional manifold embedding vector, and construct a global sensing vector. Multimodal sensors, such as laser sensors, vision sensors, infrared sensors, and ultrasonic sensors, are used to collect perception data from the working environment, including lidar point cloud data, RGB image data, infrared thermal imaging data, and ultrasonic distance measurement data. The perception data undergoes preprocessing such as noise reduction, synchronization, standardization, and normalization to obtain preprocessed perception data. The preprocessing processes all employ techniques well-known to those skilled in the art and will not be elaborated upon here. The preprocessed perception data is then used to extract features using existing feature engineering techniques to obtain preliminary feature data, such as FastPoint Feature Extraction. Histograms (FPFH) and region clustering methods are used to extract local density, normal vector distribution, and spatial geometric features from preprocessed LiDAR point cloud data, and these features are combined into a structural vector describing the obstacle morphology, serving as preliminary point cloud feature data. For preprocessed infrared thermal imaging data, the average temperature gradient, hotspot area, and edge intensity of each heat source region are statistically analyzed using a sliding window to construct its thermal imaging description vector, serving as preliminary infrared feature data. For preprocessed ultrasonic distance measurement data, a distance change curve is constructed based on the continuous frame distance change rate, and then combined with historical velocity changes to form acceleration and predicted position vectors, serving as preliminary ultrasonic feature data. For preprocessed RGB image data, a lightweight convolutional neural network (such as MobileNet or a simplified ResNet) is used to extract its multi-level texture and edge direction features, serving as preliminary RGB feature data. Furthermore, existing multimodal feature alignment and distribution normalization techniques are used to uniformly process the preliminary feature data to avoid inconsistencies in dimensionality, scale, statistical distribution, and feature space structure. Specifically, Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are used for feature dimensionality reduction and orientation alignment, the Maximum Mean Difference (MMD) method is used for modal distribution alignment, and canonical correlation analysis (CCA) or kernel canonical correlation analysis (KCCA) is used for cross-modal feature correlation matching. Simultaneously, z-score normalization or whitening transformation is used for variance normalization, ensuring that all preliminary feature data are comparable on a uniform distribution scale, resulting in uniformly processed feature data. Furthermore, to avoid uncontrollable lag in the time dimension of the uniformly processed feature data, and the inability to directly superimpose the data due to its spatial distribution not being in a unified coordinate system, a high-dimensional spatiotemporal manifold embedding algorithm is introduced to project all uniformly processed feature data onto a joint non-Euclidean manifold space. This high-dimensional spatiotemporal manifold embedding algorithm is implemented by constructing a perceptual manifold mapping model based on sliding integrals and local gradient projection, as shown in the following formula: , in, It is in time No. The high-dimensional manifold embedding vector of each sensor, i.e., the projection vector of each sensor in the unified manifold space; It is in time No. The characteristic data of each sensor after unified processing; It is a time integral variable; It is the half-width of the sliding time window, which is selected based on the lowest sampling frequency and minimum fusion requirements of the multimodal sensor. The reference value is ; This is a time decay factor used to attenuate the influence of data from non-current time periods. It is determined based on a Bayesian optimization method, and its reference value range is [value missing]. Bayesian optimization is a well-known technique in the field and will not be elaborated here. It is the first The confidence weight factors for each sensor are determined based on a confidence inference model using a Bayesian confidence mechanism, with a reference value range of [value missing]. The confidence reasoning model of the Bayesian confidence mechanism is a well-known technique in the art, and will not be elaborated here. It is in time No. The local gradient of the feature data after unified processing of each sensor is calculated by the difference method, which is a well-known technique in the art and will not be described in detail here. It is a spatial projection matrix that can project local features onto a unified joint non-Euclidean manifold space. It is constructed by a neural network model. Specifically, preprocessed perceptual data of different modalities are used as input to the neural network model. The goal is to learn to project these modal data onto a unified joint non-Euclidean manifold space. During the training process, the projection relationship is optimized by minimizing the projection error, and finally the spatial projection matrix is ​​obtained. It is a time-attention-weighted feature term used to obtain a more reliable estimate of the current moment from the local time context; It is the integration kernel, in the denominator This is used to suppress noise amplification to prevent high-value sensing terms from dominating the integral result; It is the main term of the sliding integral, which is the core term of the above formula. It is used to perform weighted integration of the feature data after unified processing of a single modal sensor in the time dimension. Regarding time No. The local gradient of the feature data after unified processing of each sensor is multiplied by the spatial projection matrix to extract the spatial rate of change. It is a local gradient projection correction term that enhances spatial consistency, enables uniform projection in spatial dimensions, and strengthens the ability to express local change trends.

[0021] For all The high-dimensional manifold embedding vectors of each sensor are concatenated to obtain the global sensing vector. , Indicates time The global perception vector, It is not a purely abstract mathematical vector (such as a one-dimensional neural network input vector), but a structured multimodal perception projection combination that retains the meaning of physical space; the global perception vector serves as a global state input vector for multidimensional attribute modeling of obstacles.

[0022] S2. Based on the global perception vector, spatial regions are divided to obtain candidate obstacle regions, and disturbance risk scores of candidate obstacle regions are calculated; based on the disturbance risk scores of candidate obstacle regions, obstacle avoidance trajectories are generated in real time to obtain the target trajectory function; based on the target trajectory function, control decision signals are generated and mapped into control commands to achieve cooperative obstacle avoidance control.

[0023] Based on global perception vectors, existing techniques such as neighborhood density-based DBSCAN, KMeans++, or tensor projection domain clustering are employed to perform clustering or semantic segmentation operations, thereby achieving spatial region division and obtaining... A subset of points in the candidate obstacle region For each candidate obstacle region, a subset of points is extracted and normalized to obtain multidimensional physical properties, including obstacle area. This is the projected area on the ground plane obtained from preprocessed LiDAR point cloud data, normalized by combining it with the expected maximum obstacle area to eliminate the influence of dimensions. It is used to measure the size of the space occupied by the obstacle. The expected maximum obstacle area is determined according to the specific application environment; obstacle boundary fluctuation rate. Based on the preprocessed lidar point cloud data, boundary points are determined through existing α-shape boundaries or normal vectors, and the edge stability is represented by the variance of boundary points between adjacent frames; azimuth angle. Based on preprocessed LiDAR point cloud data and preprocessed RGB image data, in radians (rad), the angle between the target centroid motion vector determined by the existing obstacle centroid trajectory method and the vehicle's forward direction; obstacle dynamic change rate. It is obtained based on the normalized difference of the projected area of ​​the preprocessed lidar point cloud data on the ground plane; fuzzy boundary entropy The edge sharpness is determined based on preprocessed RGB image data or preprocessed infrared thermal imaging data, using existing edge sharpness assessment methods based on gradient amplitude distribution. The methods used in the extraction of multidimensional physical attributes are well-known to those skilled in the art and will not be elaborated upon here. These multidimensional physical attributes are used as basic elements to construct a tensor. ; Indicates transpose; Furthermore, a composite nonlinear fusion of multidimensional physical attributes is performed to calculate a disturbance risk score for each candidate obstacle region, jointly reflecting the static threat level and dynamic unpredictability, as defined below: , in, It is the first Disturbance risk score for each candidate obstacle region; This reflects that larger objects are more dangerous, and the square root makes the response growth tend to be gradual, so as to avoid the risk of small objects being exaggerated; This reflects that the more blurred the boundary (i.e., the larger the entropy value of the blurred boundary), the greater the identification error. Therefore, the perturbation risk score will decrease due to poor judgment. The +2 can avoid taking the denominator for the higher value. ; It is the direction projection factor, which indicates the direction when the target's center of mass motion vector approaches directly in front of the vehicle's direction of travel. When the value approaches 1, the risk weighting is the highest; when it is located on the side or opposite side, the risk decreases. This reflects the volume occupancy of obstacles and boundary uncertainty; the ratio of these two factors multiplied by the direction projection factor. It can assign higher weight to large obstacles directly in front, and is used to assess the degree of interference of static obstacles with the vehicle's driving direction. This represents a nonlinear boundary fluctuation penalty term, used to reduce confidence in drastically fluctuating boundaries and to quantify boundary stability. An exponent of 1.5 can rapidly amplify drastic boundary changes. The coupled score representing the dynamic attributes of an obstacle and its boundary stability can balance dynamic threat and perceived trust, and is used to measure the potential interference of obstacles with obstacle avoidance paths. Based on the Disturbance risk score for each candidate obstacle region Generate obstacle avoidance trajectories in real time. Define the initial path function. This represents the ideal trajectory when there are no obstacles, such as a straight path or an existing map navigation path. However, due to high-risk obstacles intervening in the path, local trajectory reconstruction is required within the current time window to generate the target trajectory function. The specific formula for the target trajectory function is: , in, The final generated target path is in the trajectory parameters The projected position value is the target trajectory function of the actual control, which is provided to the underlying controller for angle and velocity calculation; It is a trajectory parameter, representing the cumulative distance traveled along the path trajectory, that is, the distance covered by the vehicle traveling along the predetermined trajectory direction from the starting point of the path; In trajectory parameters The disturbance amplitude coefficient is given below, with a reference range of values. , ,in, This is a global disturbance adjustment factor used to control the overall disturbance intensity. It is determined based on expert experience and its value range is [value range missing]. ; The obstacle influence attenuation radius represents the effective range of an obstacle's effect on path disturbances. It is determined based on the specific application scenario, with a reference range of values. , It is the first The projected position of each obstacle on the path It is a Gaussian decay term, which makes the path position disturbance smaller when it is far from the obstacle and larger when it is close to the obstacle; This is the primary perturbation frequency, used to control the basic fluctuation rate of the path in space. Based on preprocessed lidar point cloud data, the lidar obstacle density is estimated using a local clustering algorithm (such as DBSCAN or Euclidean clustering), and then further estimated using existing nonlinear mapping. The reference value range is... ; It is the harmonic disturbance frequency, derived by back-calculating the historical tracking error between the actual trajectory collected by the IMU and wheel speed encoder and the SLAM path, combined with error feedback control methods (such as sliding mode control or PID controller). The reference value range is... ; The initial phase of the perturbation function is determined by a back-calculation method based on path obstacle points in the preprocessed lidar point cloud data, with a reference value range of [value missing]. ; This is a curvature penalty factor used to forcibly contract and suppress disturbance terms in regions where the initial path curvature is too high. It is estimated based on the dynamic limit parameter of the vehicle's minimum turning radius, combined with the second-order curvature of the initial path function, through vehicle dynamics modeling and path geometry analysis methods. The reference value range is... ; It is the second derivative of the initial path, that is, the path at point... The curvature; Path curvature penalty term Smoothing nonlinear processing can ensure the continuity of the value range and the existence of derivatives; It is a frequency harmonic disturbance term that achieves dynamic disturbance and asymmetric bypass of the path through frequency superposition; It is a curvature contraction term used to suppress disturbances at high curves, thereby improving controllability and driving smoothness.

[0024] Based on target trajectory function Construct a nonlinear control function to generate control decision signals. To ensure the response speed, stability, and interpretability of obstacle avoidance behavior; the specific formula for the nonlinear control function is as follows: , in, It is a control decision signal; This is the current vehicle speed, determined by the wheel speed encoder / IMU, in meters per second (m / s). Yes, the target path expected speed, based on First derivative calculation, unit: meters per second (m / s); This is the vehicle's maximum design speed, used for normalization, and is measured in meters per second (m / s). It is the current vehicle orientation angle, determined by the wheel speed encoder / IMU, in radians (rad). It is the path tangent direction angle, based on The first derivative is used for inverse calculation to determine the unit: radians (rad). It is the straight-line distance between the current object and the nearest obstacle, determined by LiDAR / depth camera, in meters (m). It is the safe obstacle avoidance judgment distance, used to normalize the distance, and the unit is meters (m); This is the speed error feedback gain, used to control the speed of vehicle acceleration and deceleration response. The reference value range is... ; This is the angle error response gain, used to control steering sensitivity. The reference value range is... ; This is the obstacle distance suppression strength, used to control the degree of speed reduction during obstacle avoidance. The reference value range is... ; , , All of these are determined using gain tuning methods in linear control theory (such as the Ziegler-Nichols method); It is the speed error term, used to control the vehicle's acceleration and deceleration response. By comparing the normalized deviation between the current actual speed of the vehicle and the expected speed of the path, it realizes the dynamic consistency adjustment of the vehicle's speed on the target trajectory, thereby ensuring the continuity, accessibility and robustness of path tracking. It is the angle error term, used to control the consistency between the vehicle's steering direction and the path direction; This is the obstacle avoidance braking function. When an obstacle approaches, this function increases to suppress the vehicle. This is a distance division protection term to avoid division by zero; its possible values ​​are... The minimum turning radius, maximum design speed, and safe obstacle avoidance distance of the vehicle mentioned above are all from the vehicle manual and are not specified here. Finally, control decision signals The existing instruction solver is used for mapping processing to obtain control commands, such as angle and speed; further, the intelligent vehicle executes the control commands to achieve multimodal perception and collaborative obstacle avoidance control.

[0025] In summary, a multimodal perception and collaborative obstacle avoidance control method for intelligent vehicles has been developed.

[0026] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0027] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0028] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A multimodal perception and cooperative obstacle avoidance control method for intelligent vehicles, characterized in that, Includes the following steps: S1. Acquire and preprocess sensing data to obtain preprocessed sensing data; the sensing data includes lidar point cloud data, RGB image data, infrared thermal imaging data, and ultrasonic distance measurement data; extract features from the preprocessed sensing data to obtain preliminary feature data, and perform unified processing to obtain unified feature data; introduce a high-dimensional spatiotemporal manifold embedding algorithm to map the unified feature data to obtain a high-dimensional manifold embedding vector, and construct a global sensing vector; in the implementation of the high-dimensional spatiotemporal manifold embedding algorithm, a sliding product is generated based on the unified feature data. The process involves dividing the data into a principal term and a local gradient projection correction term. All uniformly processed feature data are projected onto a joint non-Euclidean manifold space to obtain a high-dimensional manifold embedding vector. The local gradient projection correction term is generated based on the local gradients and spatial projection matrix of the uniformly processed feature data. This spatial projection matrix projects local features onto a unified joint non-Euclidean manifold space. The spatial projection matrix is ​​constructed using a neural network model. Preprocessed perceptual data from different modalities are used as input to the neural network model. During training, the projection relationship is optimized by minimizing the projection error to obtain the spatial projection matrix. S2. Based on the global perception vector, spatial regions are divided to obtain candidate obstacle regions, and disturbance risk scores for these regions are calculated. Specifically, a subset of points in each candidate obstacle region is extracted and normalized to obtain multi-dimensional physical attributes, including obstacle area, obstacle boundary fluctuation rate, azimuth angle, obstacle dynamic change rate, and fuzzy boundary entropy. The obstacle area is the projected area on the ground plane obtained from preprocessed LiDAR point cloud data, normalized in conjunction with the expected maximum obstacle area. The obstacle boundary fluctuation rate is calculated from the variance of boundary points between adjacent frames after determining boundary points using α-shape boundaries or normal vectors based on preprocessed LiDAR point cloud data. The azimuth angle is calculated from the variance of boundary points between adjacent frames. The point cloud data and preprocessed RGB image data represent the angle between the target centroid motion vector determined by the obstacle centroid trajectory method and the vehicle's forward direction. The obstacle dynamic change rate is obtained based on the normalized difference of the projected area of ​​the preprocessed lidar point cloud data on the ground plane. The fuzzy boundary entropy is determined based on the preprocessed RGB image data or preprocessed infrared thermal imaging data through edge sharpness evaluation based on gradient amplitude distribution. Multidimensional physical properties are then subjected to composite nonlinear fusion to obtain a disturbance risk score for the candidate obstacle region. Based on the disturbance risk score of the candidate obstacle region, an obstacle avoidance trajectory is generated in real time to obtain the target trajectory function. Based on the target trajectory function, a control decision signal is generated and mapped into a control command to achieve cooperative obstacle avoidance control.

2. The intelligent vehicle multimodal perception cooperative obstacle avoidance control method according to claim 1, characterized in that, S1 specifically includes: Feature engineering techniques are used to extract features from the preprocessed sensory data to obtain preliminary feature data. Then, multimodal feature alignment and distribution normalization techniques are used to uniformly process the preliminary feature data to obtain uniformly processed feature data.

3. The intelligent vehicle multimodal perception and cooperative obstacle avoidance control method according to claim 1, characterized in that, S1 specifically includes: The high-dimensional manifold embedding vectors of all sensors are concatenated to obtain the global sensing vector.

4. The intelligent vehicle multimodal perception cooperative obstacle avoidance control method according to claim 1, characterized in that, S2 specifically includes: Based on the disturbance risk score of the candidate obstacle region, a disturbance amplitude coefficient is constructed by combining a Gaussian attenuation term, and a frequency harmonic disturbance term is generated. The main disturbance frequency in the frequency harmonic disturbance term is determined by estimating the LiDAR obstacle density through a local clustering algorithm based on preprocessed LiDAR point cloud data, and then by estimating it through nonlinear mapping. The harmonic disturbance frequency in the frequency harmonic disturbance term is obtained by back-calculation based on the historical tracking error between the actual trajectory collected by the IMU and wheel speed encoder and the SLAM path, combined with the error feedback control method.

5. The intelligent vehicle multimodal perception cooperative obstacle avoidance control method according to claim 4, characterized in that, S2 specifically includes: Define an initial path function and construct a curvature contraction term. Combine this with a frequency harmonic perturbation term to perform local trajectory reconstruction and obtain the target trajectory function.

6. The intelligent vehicle multimodal perception cooperative obstacle avoidance control method according to claim 5, characterized in that, S2 specifically includes: Based on the target trajectory function, the expected velocity and tangent direction angle of the target path are calculated, and velocity error terms and angle error terms are constructed. Combined with obstacle avoidance and braking terms, a nonlinear control function is constructed to generate control decision signals. The control decision signals are then mapped to obtain control commands.

Images