A smart driving-based vehicle obstacle avoidance and path planning method

Abstract

The application relates to the technical field of intelligent driving, in particular to a vehicle obstacle avoidance and path planning method based on intelligent driving. The application proposes the following scheme: an environment data is collected through a multi-modal sensor, a dynamic target probability space-time confidence field is constructed, a target category and a geometric state probability distribution are output by using a Bayesian deep neural network, and a multi-modal future trajectory prediction set is generated in combination with a historical trajectory. Further, an evolution scene is constructed through trajectory sampling, a forward simulation is performed on a candidate self trajectory, an optimal driving path is generated based on a conditional risk value evaluation, and the optimal driving path is converted into a control instruction to drive the vehicle. The application has uncertainty modeling capability and high risk identification capability, and the safety and robustness of path planning are improved.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to a vehicle obstacle avoidance and path planning method based on intelligent driving. Background Technology

[0002] While various deep learning and rule-based fusion schemes have been widely applied in existing intelligent driving path planning technologies, they still have many shortcomings when facing traffic scenarios with high dynamic target density, strong trajectory uncertainty, and complex environmental interference. Existing technologies often rely on a single predicted trajectory or static risk assessment when generating paths, leading to problems such as insufficient safety boundaries, frequent conflicts with dynamic targets, and poor path stability in actual execution. Furthermore, insufficient modeling of trajectory perturbations and evolution trends in multi-target interaction scenarios results in a lack of robustness to sudden behavioral changes, easily causing obstacle avoidance failures or control command malfunctions. Existing methods cannot adequately address the decision-making complexity brought about by multimodal prediction results, ultimately affecting the safety and adaptability of path planning in real-world vehicle environments.

[0003] To address the above issues, this application proposes a vehicle obstacle avoidance and path planning method based on intelligent driving. Summary of the Invention

[0004] The technical problem this application aims to solve is to address the shortcomings of existing technologies by providing a vehicle obstacle avoidance and path planning method based on intelligent driving. The method collects environmental data through multimodal sensors, constructs a probabilistic spatiotemporal confidence field for dynamic targets, and utilizes a Bayesian deep neural network to output the probability distribution of target categories and geometric states. This, combined with historical trajectories, generates a multimodal future trajectory prediction set. Furthermore, it constructs an evolutionary scenario through trajectory sampling, performs forward simulation on candidate self-trajectories, generates an optimal driving path based on conditional risk-value assessment, and converts this into control commands to drive the vehicle.

[0005] To achieve the above objectives, this application provides the following technical solution:

[0006] A vehicle obstacle avoidance and path planning method based on intelligent driving is applied to a vehicle equipped with a processing system and a multimodal sensor array. The method includes:

[0007] Acquire sensor data collected by the multimodal sensor array, the sensor data including environmental data of the vehicle's location, the environmental data including at least one dynamic target;

[0008] Based on the sensor data, a probabilistic spatiotemporal confidence field is generated by the processing system. The probabilistic spatiotemporal confidence field is a structured data space used to describe the motion of the dynamic target. For each dynamic target, the description includes at least a multimodal future trajectory prediction set representing the corresponding dynamic target. The multimodal future trajectory prediction set includes multiple future trajectory hypotheses, which are all different and have corresponding probability labels.

[0009] Based on the probabilistic spatiotemporal confidence field, multiple evolution scenarios are generated by sampling the multimodal future trajectory prediction set, vehicle trajectories are generated, and control commands are generated based on the vehicle trajectories to drive the vehicle.

[0010] The processing system generates a probabilistic spatiotemporal confidence field, including:

[0011] The sensor data is input into a preset Bayesian deep neural network model to process the dynamic target. The Bayesian deep neural network model outputs a discrete probability distribution and a continuous multidimensional probability distribution for each dynamic target. The discrete probability distribution is used to represent the category of the dynamic target, and the continuous multidimensional probability distribution is used to represent the geometric state of the dynamic target.

[0012] Using the discrete probability distribution and the continuous multidimensional probability distribution as observations, the dynamic target is state tracked, and the trajectory probability distribution is output based on the state tracking result, wherein the trajectory probability distribution is used to represent the probability distribution of the current kinematic state of the corresponding dynamic target;

[0013] Based on the trajectory probability distribution and combined with the historical kinematic state sequence of the corresponding dynamic target, a multimodal future trajectory prediction set is generated.

[0014] The main structure of the Bayesian deep neural network model is a Bayesian convolutional neural network. The weights of the Bayesian convolutional neural network are random variables following a specific distribution, which is set according to the distance between the dynamic target and the vehicle, wherein:

[0015] During the inference phase, a set of different prediction results are obtained by performing multiple forward propagations on the Bayesian convolutional neural network, wherein the forward propagation is determined by Monte Carlo random deactivation;

[0016] The prediction results are statistically analyzed, and discrete probability distributions and continuous multidimensional probability distributions are obtained based on the weights.

[0017] Based on the trajectory probability distribution and combined with the historical kinematic state sequence of the corresponding dynamic target, a multimodal future trajectory prediction set is generated, including:

[0018] The historical kinematic state sequence of the dynamic target is input into the encoder of a preset conditional variational autoencoder to compress the historical kinematic state sequence into a low-dimensional latent space.

[0019] Multiple independent random samples are performed on the low-dimensional latent space, and the trajectory probability distribution is used as a conditional input. A set of future trajectory hypotheses is generated by the decoder of the conditional variational autoencoder to obtain a multimodal future trajectory prediction set.

[0020] Based on the probabilistic spatiotemporal confidence field, multiple evolution scenarios are generated by sampling the multimodal future trajectory prediction set to generate vehicle trajectories, including:

[0021] Based on the probabilistic spatiotemporal confidence field, the terminal state point is obtained by sampling the terminal state of the dynamic target in a preset coordinate system;

[0022] A smooth candidate self-trajectory is generated for each terminal state point by a preset function, connecting the current state of the vehicle with the corresponding terminal state point. The candidate self-trajectory set is obtained by summarizing the results. The candidate self-trajectory set includes multiple different driving trajectories.

[0023] For each smooth candidate self-trajectory in the candidate self-trajectory set, an evolutionary scenario is constructed by sampling from the probabilistic spatiotemporal confidence field within a preset number of simulation cycles. The candidate self-trajectory is then subjected to forward simulation in the evolutionary scenario to calculate the comprehensive cost. Based on the comprehensive cost, a cost distribution list of smooth candidate self-trajectories is obtained, wherein the dimension of the cost distribution list is the same as the number of simulation cycles.

[0024] Based on the cost distribution list, the conditional risk value of each smooth candidate self-trajectory is calculated as the final risk score. The final risk scores are sorted in ascending order, and the smooth candidate self-trajectory with the highest score is selected as the vehicle trajectory.

[0025] Constructing an evolutionary scenario by sampling from the probabilistic spatiotemporal confidence field includes:

[0026] In each simulation cycle, each dynamic target recorded in the probabilistic spatiotemporal confidence field is traversed, and a future trajectory hypothesis is randomly selected based on the correlation probability of each future trajectory hypothesis in the corresponding multimodal future trajectory prediction set.

[0027] The future trajectories of all dynamic targets are assumed to constitute an evolutionary scenario.

[0028] The overall cost is calculated by weighting multiple cost items, which at least include:

[0029] Collision cost is a preset first penalty value assigned when the bounding box of the vehicle overlaps with the bounding box of any dynamic target at any point in time during the simulation.

[0030] The comfort cost is calculated based on the longitudinal and lateral accelerations of the candidate ego trajectory;

[0031] The task progress cost is calculated based on the degree of deviation between the candidate self-trajectories and the preset reference paths, as well as the difference between the candidate self-trajectories and the expected travel speeds.

[0032] Rule cost: When the candidate self-trajectory violates a preset traffic rule during the simulation, a preset second penalty value is assigned to the rule cost.

[0033] Based on the cost distribution list, the conditional value of risk for each smoothed candidate self-trajectory is calculated as the final risk score, including:

[0034] Sort all the comprehensive costs in the cost distribution list in ascending order;

[0035] Based on a preset confidence level factor, the risk value of the sorted cost distribution list is determined, wherein the risk value is the composite cost located in the sorted cost distribution list that reflects the quantile position of the confidence level factor;

[0036] Calculate the arithmetic mean of all aggregate costs in the cost distribution list that are greater than or equal to the risk value, and use the arithmetic mean as the conditional risk value of the candidate self-trajectories.

[0037] Before acquiring the sensor data collected by the multimodal sensor array, the method further includes:

[0038] Raw data streams are acquired through the multimodal sensor array, which includes at least one lidar, a forward-looking camera, and a millimeter-wave radar.

[0039] The original data stream is time-synchronized according to a unified system timestamp, and all time-synchronized original data streams are uniformly transformed to the carrier coordinate system using a pre-calibrated external parameter matrix of each sensor relative to the carrier coordinate system to obtain sensor data.

[0040] Generating control commands based on the vehicle trajectory to drive the vehicle includes:

[0041] An initial time segment is extracted from the vehicle trajectory, the initial time segment including the vehicle's desired path curvature and velocity profile;

[0042] The initial time segment is input to the vehicle's underlying control module, which includes a lateral controller and a longitudinal controller. The lateral controller calculates the front wheel steering angle command required to achieve path tracking based on the geometric relationship between the vehicle's current pose and the desired path. The longitudinal controller calculates the throttle opening or braking pressure command required to achieve speed tracking based on the error between the desired speed profile and the vehicle's current actual speed.

[0043] The front wheel steering angle command and the throttle opening or braking pressure command are sent to the corresponding actuators via the vehicle's controller area network bus to drive the vehicle's movement.

[0044] Compared with the prior art, the beneficial effects of this application are:

[0045] This application introduces a probabilistic spatiotemporal confidence field to perform multimodal modeling of the future behavior of dynamic targets, and combines a conditional variational autoencoder to generate multiple future trajectory hypotheses with probabilistic labels, effectively improving the ability to predict the behavior of highly uncertain traffic targets. Simultaneously, it utilizes an evolutionary scenario simulation mechanism to perform multiple rounds of extrapolation on candidate self-trajectories and performs path selection based on conditional risk value, resulting in the final selected trajectory having higher safety margins and behavioral stability in complex scenarios. Compared to traditional path planning methods that rely on single trajectory prediction or static evaluation, this application exhibits stronger decision robustness and practical feasibility in scenarios with variable target behavior and severe environmental perception interference, thereby improving the driving safety and path planning performance of the vehicle in real traffic environments. Attached Figure Description

[0046] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0047] Figure 1 This is a schematic diagram illustrating an exemplary application scenario of an embodiment of this application;

[0048] Figure 2 This is a schematic diagram of the vehicle system architecture according to an embodiment of this application;

[0049] Figure 3 This is a flowchart illustrating a vehicle obstacle avoidance and path planning method based on intelligent driving, according to an embodiment of this application.

[0050] Figure 4 This is a flowchart of the multimodal trajectory prediction process of the conditional variational autoencoder in an embodiment of this application. Detailed Implementation

[0051] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0052] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0053] The intelligent driving obstacle avoidance and path planning method described in this application is applicable to autonomous driving vehicle systems with complex dynamic environmental interactions and uncertainties in multi-objective behavior. Its core challenge lies in how to achieve robust prediction and evolutionary modeling of dynamic target behavior when there are delays, occlusions or errors in multimodal perception information, and thereby generate the optimal driving trajectory under safe, stable and rule-constrained conditions.

[0054] The application scenarios of this application include, but are not limited to:

[0055] Dynamic targets on highways or urban roads with high-density traffic flow and sudden lane changes or crossing behaviors;

[0056] In scenarios such as urban intersections and roundabouts, target behavior exhibits a clear trend of multimodal movement.

[0057] Multi-sensor perception systems are limited by viewing angle, occlusion, specular reflection or low light conditions, which can lead to blurring or reconstruction errors in the local target state;

[0058] The control system requires high-frequency path replanning and forward simulation under safety constraints in a dynamic environment.

[0059] The selection of application scenarios is based on the common challenges of perception and decision-making in real-world complex environments for autonomous driving systems. Representative characteristics include, but are not limited to, one of the following:

[0060] Dynamic targets exist in mixed traffic scenarios with high movement frequency and strong uncertainty in behavior patterns;

[0061] There are interactions among multiple objectives, and the behavior of these objectives is significantly affected by external disturbances (such as courtesy, following, and obstacle avoidance behaviors).

[0062] Sensor observation data may be affected by viewing angle bias, multipath reflections (such as glass curtain walls and slippery roads), or occlusion artifacts.

[0063] The optimal path requires a comprehensive consideration of multiple objectives, including risk avoidance, driving comfort, and traffic rules.

[0064] It should be noted that the method proposed in this application does not rely on predefined structural rules, static maps, or standard traffic layouts as prerequisites. Instead, it constructs a probabilistic spatiotemporal confidence field model for target motion prediction, combined with multimodal trajectory assumptions and scenario evolution simulation, to complete end-to-end path generation and risk assessment. This application has good versatility and scalability, and can be widely applied to different types of intelligent driving systems.

[0065] It is worth noting that the evolutionary scenario generation strategy and conditional risk value assessment method in this application are not designed for a single target or static obstacle, but are applicable to:

[0066] Scenarios where motion behavior mirror propagation occurs (e.g., similar vehicles exhibiting similar trajectory distributions from different viewpoints);

[0067] The model space exhibits diverse multi-objective behaviors and long-tailed predictive distributions;

[0068] In the path generation process, there are scenarios where the sensitivity of the target trajectory to planning decisions cannot be effectively covered by traditional deterministic modeling.

[0069] Please see Figure 1 This figure is a schematic diagram of an exemplary application scenario provided by an embodiment of this application.

[0070] Figure 1 The scenario shown includes an intelligent driving vehicle equipped with a multimodal sensor array, in which there are various types of dynamic targets in the surrounding environment, including pedestrians, motor vehicles, and non-motorized vehicles.

[0071] It is understood that the multimodal sensor array includes, but is not limited to, sensing units such as lidar, forward-looking cameras, and millimeter-wave radar, used to collect information such as three-dimensional geometry, image texture, speed, and orientation of the vehicle's surrounding environment. The collected sensor data will be synchronously transmitted to the processing system.

[0072] It is understandable that the functions of the processing system include:

[0073] Perform time synchronization and spatial unification transformation on the raw sensor data;

[0074] Identify dynamic targets and predict their multimodal future motion trajectories based on a Bayesian deep neural network model;

[0075] Construct a spatiotemporal confidence field containing the probability distribution of target trajectory prediction;

[0076] Based on the vehicle's target planning mission, sample and extrapolate future evolution scenarios to generate candidate driving trajectories;

[0077] Taking into account multiple cost indicators such as collision risk, comfort, rule compliance, and mission progress, the optimal path is output.

[0078] Furthermore, the generated driving trajectory is translated into low-level control commands and executed by the vehicle's low-level control module. This module includes lateral and longitudinal controllers to ensure that the vehicle can run smoothly and accurately along the optimal path.

[0079] In one specific implementation, the method described in this application is deployed on a vehicle with a high level of autonomous driving capability, such as... Figure 2 The diagram shows the vehicle system architecture.

[0080] The vehicle's hardware is based on a highly integrated processing system, centered around a high-performance heterogeneous computing system-on-a-chip (SoC). This system integrates multiple ARM-based central processing unit (CPU) cores for system scheduling, state machine management, sensor data synchronization, and traditional algorithm logic. Simultaneously, it integrates an Ampere-based graphics processing unit (GPU) with a floating-point performance, particularly its tensor core performance in half-precision (FP16), requiring at least 100 TFLOPS. This GPU is specifically designed to support the parallelized, high-throughput inference computation of models such as the Bayesian deep neural network and conditional variational autoencoder described in this application. The entire processing system is powered by an independent onboard power module and equipped with an active cooling system to ensure stable operation over a wide temperature range.

[0081] Figure 2 The processing system is further shown to be connected to a multimodal sensor array mounted on the vehicle via a high-speed data bus.

[0082] Specifically, the multimodal sensor array includes a 64-line lidar mounted in the center of the vehicle's roof, which is connected to the processing system via an onboard Ethernet interface.

[0083] In one example, the lidar performs a mechanical rotating scan with a 360-degree horizontal field of view at a fixed frequency of 10 Hz, emitting hundreds of thousands of laser beams in each scan cycle and receiving their return signals, thereby generating high-density point cloud data containing accurate three-dimensional spatial geometry information.

[0084] Furthermore, the multimodal sensor array includes a forward-facing camera mounted on the inner side of the upper center of the windshield. The forward-facing camera is connected to the processing system via a GMSL2 (Gigabit Multimedia Serial Link) serial interface, acquiring a stream of color image data of the road environment ahead at a rate of 30 frames per second.

[0085] Furthermore, the multimodal sensor array includes a 77GHz long-range millimeter-wave radar installed in the center of the vehicle's front bumper. It communicates with the processing system via CAN-FD (CAN with Flexible Data-Rate) or an in-vehicle Ethernet interface, and can effectively detect targets at distances exceeding 250 meters. It also directly outputs a pre-processed list of targets, which includes information such as the distance, azimuth, and radial velocity of each target.

[0086] Furthermore, to ensure the spatiotemporal consistency of data required for subsequent algorithms, the processing system also includes a hardware time synchronization scheme based on the Precision Time Protocol (PTP). The processing system, acting as the Grandmaster Clock, broadcasts a unified global system time to all sensor nodes through a clock synchronization network supporting the IEEE 802.1AS protocol. Each sensor, when generating each frame of data, obtains its current precise global timestamp from its local PTP client and appends it to the data packet. Regarding spatial alignment, all sensors undergo a rigorous offline calibration process before the vehicle leaves the factory. Camera intrinsic parameters are determined using the Zhang Zhengyou checkerboard calibration method, while the extrinsic parameters of each sensor relative to the vehicle's reference coordinate system (a right-handed coordinate system defined at the ground projection point of the vehicle's rear axle center, with the X-axis pointing forward, the Y-axis pointing to the left, and the Z-axis vertically upward), i.e., a 4x4 homogeneous transformation matrix, are obtained through a joint optimization algorithm. The joint optimization algorithm utilizes multi-sensor data collected in the calibration field, which includes specific geometric calibration objects, to minimize the reprojection error of the same feature point observed by different sensors in a unified coordinate system.

[0087] Understandably, when the vehicle is in motion, the first step of this application is the acquisition and preprocessing of sensor data. The CPU in the processing system continuously receives raw data streams with high-precision timestamps from the aforementioned multimodal sensor array. The system aligns the data streams according to the timestamps, treating the closest temporally connected frame of the LiDAR point cloud, a frame of the camera image, and a frame of the millimeter-wave radar target list as a single processing unit. Subsequently, using a pre-calibrated extrinsic parameter matrix, through standard homogeneous coordinate transformation, every 3D point in the LiDAR point cloud, every target point detected by the millimeter-wave radar, and the image feature points projected into 3D space by the camera's intrinsic parameters are all uniformly transformed to the vehicle's reference coordinate system. This results in a spatiotemporally aligned fusion sensor data frame containing rich geometric and semantic information about the environment, providing a solid data foundation for subsequent perception and prediction steps.

[0088] Next, with reference to the accompanying drawings, a vehicle obstacle avoidance and path planning method based on intelligent driving provided by an embodiment of this application will be introduced. Figure 3 The method shown includes the following steps S1-S3:

[0089] S1: Acquire the sensor data collected by the multimodal sensor array;

[0090] In this embodiment, the multimodal sensor array includes a lidar, a forward-looking camera, and a millimeter-wave radar. Each sensor collects raw data streams based on a unified system timestamp, and these data streams are uniformly mapped to the carrier coordinate system after extrinsic parameter calibration, forming a standardized 3D spatial dataset. This standardized 3D spatial dataset covers surrounding environmental information including various types of dynamic targets. The data fusion achieved in this step ensures spatial alignment and temporal consistency, thus providing an accurate basis for subsequent state estimation and trajectory prediction. In urban road scenarios with complex structures and multiple occlusions or specular reflection interference, this data synchronization mechanism can significantly reduce the risk of misidentification caused by artifacts and multipath reflections, improving perception reliability.

[0091] S2: Based on the sensor data, a probabilistic spatiotemporal confidence field is generated by the processing system;

[0092] In this embodiment, the synchronized sensor data is processed using a Bayesian convolutional neural network to output a discrete probability distribution for the category and a continuous multidimensional probability distribution for the geometric state of each dynamic target. Based on these probability outputs, and combined with the historical state sequences of each dynamic target, a conditional variational autoencoder is used to perform multimodal modeling of their future trajectories, forming a set of multiple trajectory hypotheses with probability labels. This set, together with the current state of the target, constitutes a spatiotemporal confidence field, describing the potential evolutionary trend of the dynamic target in the time dimension. This step does not rely on predefined structural models or traffic semantic maps, but actively adapts to highly uncertain traffic environments through probabilistic modeling, which is particularly effective for targets with viewpoint occlusion, severe structural reflection, and drastic trajectory changes.

[0093] It should be noted that the weight parameters in the Bayesian network in this application are no longer deterministic values, but random variables that follow a specific distribution. The specific distribution can be adjusted according to the relative distance between the target and the vehicle, allowing the model to be aware of input uncertainty and improving the prediction robustness in high-noise perception scenarios.

[0094] S3: Based on the probabilistic spatiotemporal confidence field, multiple evolution scenarios are generated by sampling the multimodal future trajectory prediction set, vehicle trajectories are generated, and control commands are generated based on the vehicle trajectories to drive the vehicle.

[0095] In this embodiment, for each dynamic target's multimodal trajectory prediction set in the probabilistic spatiotemporal confidence field, a random sampling strategy is used to generate multiple feasible evolutionary scenarios. Under each evolutionary scenario, multiple smooth candidate trajectories are constructed from the vehicle's current position to different terminal state points. Forward evolutionary reasoning is performed on each trajectory based on a simulation loop to calculate a multidimensional comprehensive cost function, including collision risk, comfort, mission progress, and traffic rules. Furthermore, by setting a confidence level factor, the conditional risk value of each candidate trajectory is calculated as a path scoring criterion. Finally, the trajectory with the lowest score is selected as the driving decision output.

[0096] Before elaborating on the specific technical details of this embodiment, it is necessary to further emphasize the following points.

[0097] In typical intelligent driving applications, the core challenges of obstacle avoidance and path planning for multiple dynamic targets do not stem from insufficient perception accuracy or control response speed, but rather from the combined effects of multiple systemic factors. These factors are particularly pronounced in open road environments, especially in typical scenarios such as: multi-lane intersections in urban areas with obstructed visibility, areas near crosswalks where pedestrians and non-motorized vehicles share the road, and reflective areas of metallic vehicle bodies under rain, snow, or strong sunlight.

[0098] Understandably, in such scenarios, the environmental data acquired by sensors often contains structural biases or multi-source path interference, manifesting as anomalies such as mirror image misalignment of target positions, reversed trajectory directions, and abrupt changes in contour boundaries. These biases do not fundamentally originate from changes in the target's own state, but rather from the spatial transformation effect of the environmental geometry on the perception path. Traditional processing strategies often rely on image filtering, point cloud sparsity detection, or confidence filtering mechanisms based on model matching. However, these methods cannot cover the uncertainties inherent in the future motion trends of dynamic targets, especially when the target state is undergoing drastic changes or is in the occlusion leading-edge stage, easily leading to state estimation biases and further misleading path decisions.

[0099] In this embodiment, instead of directly improving the identification accuracy of a single step, a probabilistic spatiotemporal confidence field is constructed as an intermediate representation structure within the joint space-time dimension. This is achieved through a set of uncertain scenarios describing the future movement trend of the target, and a multimodal trajectory sampling and conditional risk value assessment mechanism to systematically screen potential outcomes of path generation. In practice, decision-making depends not only on the target's location and category but also on its probability weight distribution along its future evolutionary path, thereby enabling dynamic adjustment of path planning and risk avoidance.

[0100] Furthermore, in complex scenarios where point cloud information distribution exhibits asymmetric perturbations, spatial distortion due to reflections, and multiple dynamic targets occluding or overlapping each other, traditional methods, lacking trajectory evolution simulation mechanisms, often rely solely on static indicators for judgments, lacking the ability to perceive potential future conflicts. In contrast, the evolutionary scenario simulation mechanism employed in this application constructs multiple possibilities for target motion through multiple independent samplings and performs collaborative matching during the candidate path generation stage. This ensures that the final selected path not only minimizes the expected cost but also controls tail risks within an acceptable range.

[0101] Therefore, this application is not intended for targeted optimization design of a specific fixed scenario or transportation facility, but rather takes the uncertain modeling of the target trajectory as its starting point to systematically solve the problem of path planning instability in highly dynamic, high-density, and multi-sensory interference environments. The proposed method has stronger versatility, stability, and engineering feasibility, and is particularly suitable for the following situations: the dynamic target motion behavior is difficult to predict accurately, and traditional single-trajectory models fail; sensor data is subject to dual interference from structural occlusion and specular reflection, and point cloud information contains symmetrical artifacts; the path planning task needs to simultaneously meet multiple dimensions of concurrent constraints such as safety, regularity, and comfort.

[0102] Next, we will further elaborate on the part of the method in this application regarding the probabilistic spatiotemporal confidence field.

[0103] Understandably, the probabilistic spatiotemporal confidence field is implemented as a hash table in terms of data structure, used to dynamically store and update the state information of all tracked dynamic targets in the environment. The key of this hash table is a globally unique identifier assigned to each dynamic target, and its corresponding value is a complex custom data structure object. The object encapsulates the complete probabilistic description of the target, and its core content is a multimodal future trajectory prediction set. This prediction set is not a single trajectory, but consists of multiple future trajectory assumptions that differ in both spatial path and velocity planning, and each trajectory assumption is accompanied by a corresponding probability label. The sum of all probability labels is 1, fully expressing the uncertainty of the target's future motion intention.

[0104] In one example, the processing system generates a probabilistic spatiotemporal confidence field, including:

[0105] S2.1: Input the sensor data into a preset Bayesian deep neural network model to process the dynamic target. The Bayesian deep neural network model outputs a discrete probability distribution and a continuous multidimensional probability distribution for each dynamic target. The discrete probability distribution is used to represent the category of the dynamic target, and the continuous multidimensional probability distribution is used to represent the geometric state of the dynamic target.

[0106] Specifically, in actual perception data, due to sensor noise, limited acquisition angles, and frequent target occlusion, directly using traditional neural networks to output fixed category labels and geometric position estimates can easily lead to unstable target state estimations due to error accumulation. This is particularly serious in areas with dense multi-target interaction or blurred boundaries, where misidentification is more severe. Therefore, this embodiment introduces a Bayesian deep neural network model to address the lack of uncertainty characterization in the target state output results and enhance the model's adaptability to complex real-world road environments.

[0107] In this embodiment, the Bayesian deep neural network model adopts a Bayesian convolutional neural network structure. Its key feature is that all neural network weights in the model are no longer deterministic parameters, but are expressed as random variables following a prior distribution (such as a normal distribution). During network training, variational inference optimizes the posterior distribution parameters in the network. During the inference phase, multiple Monte Carlo samplings are performed on the network weights to generate a set of statistically significant output results. For each dynamic target, the model outputs two parts of probabilistic information:

[0108] One is the discrete probability distribution, which describes the probability distribution value of the target belonging to semantic categories such as "motor vehicle", "pedestrian", and "non-motor vehicle".

[0109] The second type is the continuous multidimensional probability distribution, which adopts the form of Gaussian distribution or mixture Gaussian distribution to characterize the geometric state probability distribution of the target at various dimensions in space.

[0110] In one example, the weights of the Bayesian convolutional neural network are random variables following a specific distribution, which is set according to the distance between the dynamic target and the vehicle, wherein:

[0111] During the inference phase, a set of different prediction results are obtained by performing multiple forward propagations on the Bayesian convolutional neural network, wherein the forward propagation is determined by Monte Carlo random deactivation;

[0112] The prediction results are statistically analyzed, and discrete probability distributions and continuous multidimensional probability distributions are obtained based on the weights.

[0113] S2.2: Using the discrete probability distribution and the continuous multidimensional probability distribution as observations, perform state tracking on the dynamic target, and output the trajectory probability distribution based on the state tracking results, wherein the trajectory probability distribution is used to represent the probability distribution of the current kinematic state of the corresponding dynamic target;

[0114] Specifically, in dynamic traffic scenarios, the target state is continuous, but due to perceptual interference and data latency, single-frame data often cannot accurately reconstruct the target's true physical state. Therefore, after obtaining the observation data output by the Bayesian neural network, it is necessary to model and filter the target's state evolution process in a continuous time series through state tracking, thereby deriving its true current kinematic state. The main purpose of this step is to transform the aforementioned probability output from static observation into a time-dependent dynamic estimate, enabling trajectory prediction to be built upon the synergistic support of a dynamic physical model and continuous observation.

[0115] In this embodiment, a state tracking method combining joint probabilistic data association and extended Kalman filtering is employed, specifically including:

[0116] First, the class probability and geometric state probability output by the Bayesian neural network are used as the input observation vector;

[0117] Secondly, a state vector containing elements such as position, velocity, acceleration, and orientation is defined in the target state space. Then, the state from the previous time step is predicted forward using a state prediction function, and jointly updated with the current observation to obtain the posterior state estimate and covariance matrix. To enhance target separation capability, a joint similarity function based on category distribution is introduced during data association to improve target retention in scenarios with a large number of targets or close proximity.

[0118] Furthermore, the above processing yields the trajectory probability distribution of each target at the current moment, which is the weighted probability result of its trajectory across multiple feasible states. This trajectory probability distribution can be used to measure the state stability of the target and the evolutionary trend of its current behavior, serving as a key input parameter for subsequent multimodal trajectory prediction modeling.

[0119] S2.3: Based on the trajectory probability distribution and combined with the historical kinematic state sequence of the corresponding dynamic target, generate a multimodal future trajectory prediction set;

[0120] Reference Figure 4 , Figure 4 This is a flowchart illustrating the multimodal trajectory prediction process of a conditional variational autoencoder according to an embodiment of this application. This step aims to generate predictions of the target's various possible future intentions based on its current motion state and historical behavior.

[0121] Specifically, for each tracked dynamic target, the system first extracts a historical kinematic state sequence from its tracker over a past period (e.g., 2 seconds, with 21 points at 0.1-second intervals). This sequence contains the target's position, velocity, acceleration, and heading angle at each moment. This time-series data is then fed into a pre-trained conditional variational autoencoder (CVAE) model.

[0122] Understandably, the historical state sequence is first fed into the CVAE encoder. The encoder consists of a gated recurrent unit (GRU) network with two layers, each containing 256 hidden units. The GRU network effectively captures long-term dependencies in the time-series data and compresses and encodes the entire input sequence into parameters of a Gaussian distribution in a low-dimensional latent space, namely the mean and log-variance. This latent variable z captures the inherent patterns, driving styles, or behavioral intentions contained in the target's historical trajectory.

[0123] Furthermore, to generate diverse future trajectories, this application does not directly use the mean of the encoder output, but instead performs M independent random samplings in this low-dimensional latent space. Each sampling draws a latent variable sample from a Gaussian distribution N defined by the mean and log-variance.

[0124] Furthermore, based on the latent variable samples and the trajectory probability distribution (specifically its mean vector, used as conditional input), these are fed into the CVAE decoder. The decoder is also a GRU network structure; it uses the latent variables and the trajectory probability distribution corresponding to the current state as its initial hidden state, and then, within a preset future time domain, autoregressively generates the position and velocity of future trajectory points step by step. Since the M input latent variable samples are all different, the decoder naturally generates M future trajectory hypotheses that differ in path shape, velocity profile, and other aspects. For example, for a vehicle waiting at an intersection, different latent variable samples might correspond to different intentions such as "continue waiting," "go straight," or "turn left," thus generating trajectories with vastly different shapes.

[0125] Furthermore, to assign a confidence level to each generated trajectory hypothesis, each latent variable sample is fed into a lightweight multilayer perceptron (MLP) running in parallel with the decoder. The MLP analyzes the input latent variables and outputs a scalar value. These M scalar values ​​are collected and normalized using a predefined softmax function to obtain probability labels corresponding to the M trajectory hypotheses. These M trajectory hypotheses and their corresponding probability labels together constitute the multimodal future trajectory prediction set for the dynamic target. Finally, this prediction set, rich in uncertainty information, is fully stored in a probabilistic spatiotemporal confidence field as the value of the corresponding target ID.

[0126] In one example, based on the trajectory probability distribution and combined with the historical kinematic state sequence of the corresponding dynamic target, a multimodal future trajectory prediction set is generated, including:

[0127] The historical kinematic state sequence of the dynamic target is input into the encoder of a preset conditional variational autoencoder to compress the historical kinematic state sequence into a low-dimensional latent space.

[0128] Multiple independent random samples are performed on the low-dimensional latent space, and the trajectory probability distribution is used as a conditional input. A set of future trajectory hypotheses is generated by the decoder of the conditional variational autoencoder to obtain a multimodal future trajectory prediction set.

[0129] Next, we will further elaborate on the part of the method in this application regarding the generation of vehicle trajectories.

[0130] Understandably, the core idea is to make decisions based on the principle of risk aversion, taking full account of all possible future environmental possibilities.

[0131] In one example, based on the probabilistic spatiotemporal confidence field, multiple evolutionary scenarios are generated by sampling the multimodal future trajectory prediction set to generate vehicle trajectories, including:

[0132] S3.1: Based on the probabilistic spatiotemporal confidence field, the terminal state point is obtained by sampling the terminal state of the dynamic target in a preset coordinate system;

[0133] Specifically, the selection of terminal state points is a crucial prerequisite for constructing future evolution scenarios. This is because the spatial positions that a dynamic target may occupy at different times have probabilistic distribution characteristics and cannot be described by a single definite value. Ignoring this uncertainty will cause the planned trajectory to deviate from the real environment, thereby increasing the risk of collision or failure. To accurately model the uncertainty distribution of the future state of a dynamic target, this embodiment performs conditional sampling on the predicted trajectory sets of each target in the probabilistic spatiotemporal confidence field, obtaining multiple terminal state points under a uniformly constructed local Cartesian coordinate system or road constraint coordinate system.

[0134] In this embodiment, when sampling the multimodal trajectory prediction set for each dynamic target, the peak probability density distribution at the trajectory endpoint is given priority. Simultaneously, the final state point is corrected by incorporating the target's current velocity and acceleration direction, ensuring that the generated terminal state point is not only representative but also covers the main trajectory branches in the prediction space. This sampling process supports the use of importance sampling, introducing a path confidence weight function to weight high-probability trajectory endpoints, avoiding overfitting to extremely low-probability paths, thereby controlling the convergence boundary of the evolution scenario and reducing the simulation dimensionality.

[0135] S3.2: Generate a smooth candidate self-trajectory connecting the current state of the vehicle and the corresponding terminal state point for each terminal state point through a preset function, and summarize them to obtain a candidate self-trajectory set, wherein the candidate self-trajectory set includes multiple different driving trajectories;

[0136] Specifically, to adapt to the possible future states of each dynamic target, it is necessary to construct feasible paths for the vehicle to extend in various future traffic evolution directions at the current moment. Therefore, for the terminal state points obtained above, a series of candidate self-trajectories that can cover different spatial behavior trends need to be generated, and these trajectories need to be ensured to have controllable curvature changes and smoothness before they can be used for execution by the downstream controller. The main purpose of this step is to form a set of candidate trajectories that are representative, controllable, and satisfy dynamic constraints for subsequent cost calculation and path selection.

[0137] In this embodiment, the preset trajectory generation function can be in the form of a polynomial spline. It uses the current state of the vehicle as the starting point and the target terminal state point as the ending point, while simultaneously setting a set of intermediate control points to adjust the shape curvature and acceleration changes of the trajectory, thus constructing a continuous and differentiable two-dimensional or three-dimensional path segment. During trajectory generation, this function also generates the velocity profile and heading rate of change for each trajectory segment, ensuring that the generated trajectory meets physical feasibility requirements.

[0138] Furthermore, to enhance trajectory diversity, multiple sampling combinations are set for intermediate control point positions, path lengths, and endpoint speeds during the generation process. Each combination generates a trajectory curve, forming multiple alternative paths with different spatial layouts and motion patterns. This results in a candidate trajectory set encompassing strategies such as turning, following, detouring, and deceleration / waiting. The candidate trajectory set provides sufficient strategic flexibility to adapt to obstacle avoidance requirements in different evolutionary scenarios, providing a foundation of a path library for evaluation during the simulation phase.

[0139] S3.3: Based on each smooth candidate self-trajectory in the candidate self-trajectory set, within a preset number of simulation cycles, an evolutionary scenario is constructed by sampling from the probabilistic spatiotemporal confidence field, and a forward simulation is performed on the candidate self-trajectory under the evolutionary scenario to calculate the comprehensive cost. Based on the comprehensive cost, a cost distribution list of smooth candidate self-trajectories is obtained, wherein the dimension of the cost distribution list is the same as the number of simulation cycles.

[0140] Specifically, the optimality of a trajectory cannot be determined solely by its structure; its robustness and risk performance must be verified under specific traffic evolution scenarios. Due to the uncertainty of the target behavior, this embodiment employs a multi-round simulation loop to construct a stochastic evolution scenario. Forward inference simulation is performed on each candidate trajectory to simulate potential collisions, comfort fluctuations, rule violations, and other issues encountered during actual driving, ultimately generating cost statistics for each trajectory.

[0141] In this embodiment, in each simulation cycle, a trajectory hypothesis is sampled for all dynamic targets from the probabilistic spatiotemporal confidence field, causing them to move in the form of this trajectory during the simulation. Then, the candidate trajectory is superimposed with the target trajectory set, and the vehicle's position is forward-evolved using a discrete time step. At the same time, the spatial relationship and dynamic interaction state between the vehicle and surrounding dynamic targets are detected, and indicators such as collision risk, speed change, and out-of-line behavior at critical moments are recorded. The cost of a single simulation is calculated by summing the data according to weights.

[0142] Furthermore, the preset number of simulation cycles can be hundreds. Each cycle constructs a possible traffic evolution instance and records the performance of candidate trajectories in that scenario, forming a set of cost distributions. This distribution comprehensively reflects the stability and fault tolerance of the trajectory under uncertain environments. Compared with static cost evaluation models, it can more realistically simulate complex factors such as target behavior disturbances, path interference, and implicit rule violations, thereby improving the safety and reliability of path selection.

[0143] In one example, in each simulation cycle, each dynamic target recorded in the probabilistic spatiotemporal confidence field is traversed, and a future trajectory hypothesis is randomly selected based on the correlation probability of each future trajectory hypothesis in the corresponding multimodal future trajectory prediction set.

[0144] The future trajectories of all dynamic targets are assumed to constitute an evolutionary scenario.

[0145] In one example, the cost item includes at least:

[0146] Collision cost is a preset first penalty value assigned when the bounding box of the vehicle overlaps with the bounding box of any dynamic target at any point in time during the simulation.

[0147] The comfort cost is calculated based on the longitudinal and lateral accelerations of the candidate ego trajectory;

[0148] The task progress cost is calculated based on the degree of deviation between the candidate self-trajectories and the preset reference paths, as well as the difference between the candidate self-trajectories and the expected travel speeds.

[0149] Rule cost: When the candidate self-trajectory violates a preset traffic rule during the simulation, a preset second penalty value is assigned to the rule cost.

[0150] In some optional embodiments, the system first generates a set of candidate self-trajectories in a coordinate system based on a pre-planned global reference path (e.g., a lane centerline provided by a high-precision map). The system generates diverse candidate trajectories by discretizing future terminal states in both the lateral and longitudinal dimensions. For example, in the longitudinal dimension, different target times T are sampled, along with the target speed differences relative to the reference path speed expected at these times. In the lateral dimension, the expected lateral offset d relative to the reference path centerline at target time T is sampled. For each sampled combination of terminal states, a fifth-order polynomial and a fourth- or fifth-order polynomial connecting the vehicle's current state to that terminal state are solved. These polynomials ensure that the derivatives of the generated trajectory are continuous at both the start and end points, thus ensuring the smoothness and kinematic feasibility of the trajectory. By traversing and solving all sampled points, a set of K candidate self-trajectories covering various driving strategies such as following, lane changing, and deceleration avoidance can be generated.

[0151] Furthermore, a risk assessment based on Monte Carlo simulation is performed on each candidate self-trajectories.

[0152] Understandably, for any trajectory in the candidate self-trajectory set, a simulation loop consisting of S iterations is initiated. At the beginning of each iteration, a specific future world evolution scenario is first constructed by sampling from the probabilistic spatiotemporal confidence field. The construction process is as follows: traverse each dynamic target recorded in the probabilistic spatiotemporal confidence field, and based on the correlation probabilities of the hypothetical trajectories of each trajectory in its corresponding multimodal future trajectory prediction set, randomly select a future trajectory using roulette wheel selection or alias sampling. For example, if target A has two predicted trajectories with probabilities of 0.9 and 0.1 respectively, then in this simulation, there is a 90% probability of choosing the first trajectory and a 10% probability of choosing the second trajectory. By aggregating the trajectories extracted from all dynamic targets, a logically self-consistent single scenario of how the world evolves in the next 5 seconds is formed.

[0153] In the constructed scenario, forward simulation is performed on the candidate self-trajectories currently being evaluated to calculate their comprehensive cost in this specific scenario.

[0154] After performing all S simulation cycles on the same candidate trajectory, the system obtains a list containing S comprehensive cost values. This cost distribution list intuitively reflects the performance of the candidate trajectory under various possibilities in the future, especially including the costs corresponding to "black swan" events, which, although low in probability, are extremely costly.

[0155] S3.4: Based on the cost distribution list, calculate the conditional risk value of each smooth candidate self-trajectory as the final risk score, sort the final risk scores in ascending order, and select the smooth candidate self-trajectory with the highest score as the vehicle trajectory.

[0156] Specifically, in the cost distribution generated by the simulation, sorting by average value alone cannot reflect the performance of the trajectory in the high-risk tail. Therefore, risk control indicators need to be introduced. Conditional Value at Risk (CVaR) is a commonly used tail distribution measurement method in financial risk management. This embodiment introduces this mechanism to quantify the performance of the trajectory in extreme scenarios, ensuring that the selected path has robust safety guarantees.

[0157] In this embodiment, the cost distribution list corresponding to each candidate trajectory is sorted in ascending order. A confidence level is set, and the average of all cost values ​​above the confidence level in the sorted results is calculated as the conditional value of risk (CVaR) of that trajectory. The CVaR values ​​of all candidate trajectories form a risk score result, which is then sorted in ascending order, and the path with the lowest score is selected as the final execution path.

[0158] In one example, based on the cost distribution list, the conditional value of risk for each smoothed candidate self-trajectory is calculated as the final risk score, including:

[0159] Sort all the comprehensive costs in the cost distribution list in ascending order;

[0160] Based on a preset confidence level factor, the risk value of the sorted cost distribution list is determined, wherein the risk value is the composite cost located in the sorted cost distribution list that reflects the quantile position of the confidence level factor;

[0161] Calculate the arithmetic mean of all aggregate costs in the cost distribution list that are greater than or equal to the risk value, and use the arithmetic mean as the conditional risk value of the candidate self-trajectories.

[0162] In one example, generating control commands based on the vehicle trajectory to drive the vehicle includes:

[0163] An initial time segment is extracted from the vehicle trajectory, the initial time segment including the vehicle's desired path curvature and velocity profile;

[0164] The initial time segment is input to the vehicle's underlying control module, which includes a lateral controller and a longitudinal controller. The lateral controller calculates the front wheel steering angle command required to achieve path tracking based on the geometric relationship between the vehicle's current pose and the desired path. The longitudinal controller calculates the throttle opening or braking pressure command required to achieve speed tracking based on the error between the desired speed profile and the vehicle's current actual speed.

[0165] The front wheel steering angle command and the throttle opening or braking pressure command are sent to the corresponding actuators via the vehicle's controller area network bus to drive the vehicle's movement.

[0166] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A vehicle obstacle avoidance and path planning method based on intelligent driving, applied to a vehicle equipped with a processing system and a multimodal sensor array, characterized in that, The method includes: Acquire sensor data collected by the multimodal sensor array, the sensor data including environmental data of the vehicle's location, the environmental data including at least one dynamic target; Based on the sensor data, a probabilistic spatiotemporal confidence field is generated by the processing system. The probabilistic spatiotemporal confidence field is a structured data space used to describe the motion of the dynamic target. For each dynamic target, the description includes at least a multimodal future trajectory prediction set representing the corresponding dynamic target. The multimodal future trajectory prediction set includes multiple future trajectory hypotheses, which are all different and have corresponding probability labels. Based on the probabilistic spatiotemporal confidence field, multiple evolution scenarios are generated by sampling the multimodal future trajectory prediction set, vehicle trajectories are generated, and control commands are generated based on the vehicle trajectories to drive the vehicle. The processing system generates a probabilistic spatiotemporal confidence field, including: The sensor data is input into a preset Bayesian deep neural network model to process the dynamic target. The Bayesian deep neural network model outputs a discrete probability distribution and a continuous multidimensional probability distribution for each dynamic target. The discrete probability distribution is used to represent the category of the dynamic target, and the continuous multidimensional probability distribution is used to represent the geometric state of the dynamic target. Using the discrete probability distribution and the continuous multidimensional probability distribution as observations, the dynamic target is state tracked, and the trajectory probability distribution is output based on the state tracking result, wherein the trajectory probability distribution is used to represent the probability distribution of the current kinematic state of the corresponding dynamic target; Based on the trajectory probability distribution, and combined with the historical kinematic state sequence of the corresponding dynamic target, a multimodal future trajectory prediction set is generated. Based on the trajectory probability distribution and combined with the historical kinematic state sequence of the corresponding dynamic target, a multimodal future trajectory prediction set is generated, including: The historical kinematic state sequence of the dynamic target is input into the encoder of a preset conditional variational autoencoder to compress the historical kinematic state sequence into a low-dimensional latent space. Multiple independent random samples are performed on the low-dimensional latent space, and the trajectory probability distribution is used as a conditional input. A set of future trajectory hypotheses is generated by the decoder of the conditional variational autoencoder to obtain a multimodal future trajectory prediction set. Based on the probabilistic spatiotemporal confidence field, multiple evolution scenarios are generated by sampling the multimodal future trajectory prediction set to generate vehicle trajectories, including: Based on the probabilistic spatiotemporal confidence field, the terminal state point is obtained by sampling the terminal state of the dynamic target in a preset coordinate system; A smooth candidate self-trajectory is generated for each terminal state point by a preset function, connecting the current state of the vehicle with the corresponding terminal state point. The candidate self-trajectory set is obtained by summarizing the results. The candidate self-trajectory set includes multiple different driving trajectories. For each smooth candidate self-trajectory in the candidate self-trajectory set, an evolutionary scenario is constructed by sampling from the probabilistic spatiotemporal confidence field within a preset number of simulation cycles. The candidate self-trajectory is then subjected to forward simulation in the evolutionary scenario to calculate the comprehensive cost. Based on the comprehensive cost, a cost distribution list of smooth candidate self-trajectories is obtained, wherein the dimension of the cost distribution list is the same as the number of simulation cycles. Based on the cost distribution list, the conditional risk value of each smooth candidate self-trajectory is calculated as the final risk score. The final risk scores are sorted in ascending order, and the smooth candidate self-trajectory with the highest score is selected as the vehicle trajectory.

2. The vehicle obstacle avoidance and path planning method based on intelligent driving according to claim 1, characterized in that, The main structure of the Bayesian deep neural network model is a Bayesian convolutional neural network. The weights of the Bayesian convolutional neural network are random variables following a specific distribution, which is set according to the distance between the dynamic target and the vehicle, wherein: During the inference phase, a set of different prediction results are obtained by performing multiple forward propagations on the Bayesian convolutional neural network, wherein the forward propagation is determined by Monte Carlo random deactivation; The prediction results are statistically analyzed, and discrete probability distributions and continuous multidimensional probability distributions are obtained based on the weights.

3. The vehicle obstacle avoidance and path planning method based on intelligent driving according to claim 1, characterized in that, Constructing an evolutionary scenario by sampling from the probabilistic spatiotemporal confidence field includes: In each simulation cycle, each dynamic target recorded in the probabilistic spatiotemporal confidence field is traversed, and a future trajectory hypothesis is randomly selected based on the correlation probability of each future trajectory hypothesis in the corresponding multimodal future trajectory prediction set. The future trajectories of all dynamic targets are assumed to constitute an evolutionary scenario.

4. The vehicle obstacle avoidance and path planning method based on intelligent driving according to claim 1, characterized in that, The overall cost is calculated by weighting multiple cost items, which at least include: Collision cost is a preset first penalty value assigned when the bounding box of the vehicle overlaps with the bounding box of any dynamic target at any point in time during the simulation. The comfort cost is calculated based on the longitudinal and lateral accelerations of the candidate ego trajectory; The task progress cost is calculated based on the degree of deviation between the candidate self-trajectories and the preset reference paths, as well as the difference between the candidate self-trajectories and the expected travel speeds. Rule cost: When the candidate self-trajectory violates a preset traffic rule during the simulation, a preset second penalty value is assigned to the rule cost.

5. The vehicle obstacle avoidance and path planning method based on intelligent driving according to claim 1, characterized in that, Based on the cost distribution list, the conditional value of risk for each smoothed candidate self-trajectory is calculated as the final risk score, including: Sort all the comprehensive costs in the cost distribution list in ascending order; Based on a preset confidence level factor, the risk value of the sorted cost distribution list is determined, wherein the risk value is the composite cost located in the sorted cost distribution list that reflects the quantile position of the confidence level factor; Calculate the arithmetic mean of all aggregate costs in the cost distribution list that are greater than or equal to the risk value, and use the arithmetic mean as the conditional risk value of the candidate self-trajectories.

6. The vehicle obstacle avoidance and path planning method based on intelligent driving according to claim 1, characterized in that, Before acquiring the sensor data collected by the multimodal sensor array, the method further includes: Raw data streams are acquired through the multimodal sensor array, which includes at least one lidar, a forward-looking camera, and a millimeter-wave radar. The original data stream is time-synchronized according to a unified system timestamp, and all time-synchronized original data streams are uniformly transformed to the carrier coordinate system using a pre-calibrated external parameter matrix of each sensor relative to the carrier coordinate system to obtain sensor data.

7. The vehicle obstacle avoidance and path planning method based on intelligent driving according to claim 1, characterized in that, Generating control commands based on the vehicle trajectory to drive the vehicle includes: An initial time segment is extracted from the vehicle trajectory, the initial time segment including the vehicle's desired path curvature and velocity profile; The initial time segment is input to the vehicle's underlying control module, which includes a lateral controller and a longitudinal controller. The lateral controller calculates the front wheel steering angle command required to achieve path tracking based on the geometric relationship between the vehicle's current pose and the desired path. The longitudinal controller calculates the throttle opening or braking pressure command required to achieve speed tracking based on the error between the desired speed profile and the vehicle's current actual speed. The front wheel steering angle command and the throttle opening or braking pressure command are sent to the corresponding actuators via the vehicle's controller area network bus to drive the vehicle's movement.

Images