A method and system for obstacle risk level assessment based on three-dimensional voxel occupancy

Abstract

This invention relates to the field of autonomous driving and intelligent vehicle environmental perception technology, and provides a method and system for obstacle risk level assessment based on three-dimensional voxel occupancy. The system includes: constructing a three-dimensional voxel occupancy map to simultaneously describe the spatial distribution of static and dynamic obstacles; obtaining a set of reachable trajectories within a preset future time window; performing spatial correlation analysis on the three-dimensional voxel occupancy map and the set of reachable trajectories to determine whether a voxel is located within the influence area of ​​the vehicle's trajectory; for voxels located within the influence area of ​​the vehicle's trajectory, comprehensively quantifying the risk level from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, to obtain a voxel-level comprehensive risk score; and achieving obstacle risk level assessment based on the voxel-level comprehensive risk score. This invention can directly quantify risk in three-dimensional voxel space, characterizing the risk distribution in continuous space, and is applicable to obstacles of any shape.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and intelligent vehicle environmental perception technology, and more specifically, to a method and system for assessing obstacle risk levels based on three-dimensional voxel occupancy. Background Technology

[0002] In existing autonomous driving systems, environmental risk assessment is usually based on target-level perception results, that is, obstacle risk is judged based on obstacle bounding boxes, categories and trajectory information output by target detection and tracking, combined with indicators such as time-to-collision (TTC) and minimum safe distance.

[0003] After searching, the applicant found some typical existing technologies, such as Chinese invention patent application number CN202511547539.3 "Real-time Environmental Perception and Obstacle Trajectory Prediction Method for Autonomous Driving", Chinese invention patent application number CN202311526521.6 "A Method for Automatic Obstacle Avoidance in Assisted Driving", and Chinese invention patent application number CN202110843823.0 "A Lane Changing and Obstacle Avoidance Method and System Based on Autonomous Driving". These autonomous driving obstacle avoidance methods have the same or similar defects. They mostly make risk judgments on a target-by-target basis, which makes it difficult to describe the different risk levels of local areas of obstacles to the passage of the vehicle, and lacks the ability to express the continuous spatial risk distribution.

[0004] Therefore, there is an urgent need for an obstacle risk level assessment method that can directly characterize obstacle risks at the spatial level. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention proposes a method and system for assessing obstacle risk levels based on three-dimensional voxel occupancy, the specific technical solution of which is as follows:

[0006] An obstacle risk level assessment method based on three-dimensional voxel occupancy includes: Acquire information about the vehicle's surrounding environment, perform voxel modeling of the environment space through a 3D occupancy prediction network, and construct a 3D voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space. Obtain the current motion status information of the vehicle and, based on the motion status information, obtain a set of reachable trajectories within a preset future time window; Spatial correlation analysis is performed on the 3D voxel occupancy map and the set of reachable trajectories to determine whether the voxels are located within the influence area of ​​the vehicle trajectory. For voxels located within the influence area of ​​the vehicle trajectory, the risk level is comprehensively quantified from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, to obtain a voxel-level comprehensive risk score. Obstacle risk level assessment is achieved based on voxel-level comprehensive risk scoring.

[0007] The obstacle risk level assessment method based on three-dimensional voxel occupancy uses a three-dimensional occupancy prediction network to perform voxel modeling of the environmental space, and constructs a three-dimensional voxel occupancy map to simultaneously describe the distribution of static and dynamic obstacles in space. It breaks through the traditional target-level risk assessment method, can directly quantify risk in three-dimensional voxel space, and characterize the risk distribution in continuous space, and is applicable to obstacles of any shape.

[0008] Furthermore, by performing spatial correlation analysis on the three-dimensional voxel occupancy map and the set of reachable trajectories, it is determined whether the voxel is located within the influence area of ​​the vehicle's trajectory. Then, for the voxels located within the influence area of ​​the vehicle's trajectory, the risk level is comprehensively quantified from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, to obtain a voxel-level comprehensive risk score. The obstacle risk level assessment method based on three-dimensional voxel occupancy described in this invention can achieve 2. a risk discrimination mechanism based on vehicle trajectory driving, so that the risk assessment no longer relies solely on the obstacle's own attributes, but rather on its influence on the future trajectory of the vehicle as the core basis, thereby making the risk assessment results more in line with driving decision-making needs.

[0009] Preferably, the specific method for constructing a three-dimensional voxel occupancy map includes: The pre-defined space around the vehicle is divided into a regular three-dimensional voxel grid; For each voxel, record the probability value of the voxel being occupied by an obstacle, and construct a three-dimensional voxel occupancy map based on the probability value.

[0010] Preferably, the specific methods for obtaining the set of reachable trajectories include: Construct a vehicle kinematics model and input the motion state information as the initial condition for trajectory generation into the vehicle kinematics model; The vehicle pose at several future moments is recursively calculated using discrete time steps, and physical constraints are introduced to ensure that the generated reachable trajectory conforms to the actual feasible range of vehicle movement. Based on the vehicle's current speed, heading angle, and steering status, the possible driving behaviors within a future preset time window are enumerated and modeled, and corresponding candidate motion trajectories are generated respectively. Based on the candidate motion trajectories, obtain the set of reachable trajectories.

[0011] Preferably, the specific method for spatial correlation analysis of the three-dimensional voxel occupancy map and the set of reachable trajectories includes: Obtain the occupancy probability distribution of each voxel at multiple discrete time points in the future to characterize the potential occupancy of voxels over time. Under a unified coordinate system, for each predicted time step, the predicted position of the voxel at the corresponding future time is matched with the trajectory points in the set of reachable trajectories of the vehicle with the same or similar arrival times, and the spatial distance between the voxel center point and the trajectory point is calculated. By aligning the voxel prediction occupancy state with the arrival time of the vehicle's trajectory in the time dimension, it is determined whether the area is likely to be occupied by an obstacle when the vehicle is expected to pass through it.

[0012] Preferably, the specific method for determining whether a voxel is located within the influence area of ​​the vehicle trajectory includes: When the probability of a voxel occupying a position within a future preset time window is higher than a first preset probability threshold, and the spatial distance between the predicted position and the trajectory point at the corresponding time is less than a preset influence range, the voxel is determined to be located within the influence area of ​​the vehicle trajectory. For voxels that consistently move away from the vehicle's potential movement path or have a probability of occupying a position less than a second preset probability threshold within a future preset time window, it is determined that the voxel is not located within the vehicle's trajectory influence area.

[0013] Preferably, specific methods for comprehensively quantifying the degree of risk include: Obtain the minimum spatial distance from the voxel to the trajectory within a preset time window, normalize the minimum spatial distance, and obtain the spatial risk factor that is negatively correlated with the minimum spatial distance. Obtain the estimated arrival time corresponding to the trajectory point, normalize the estimated arrival time, and obtain the time risk factor that is negatively correlated with the estimated arrival time; Obtain the voxel occupancy probability, which reflects the credibility of a voxel being actually occupied by an obstacle within a future preset time window, and obtain the occupancy risk factor that is positively correlated with the voxel occupancy probability. Obtain the relative approach speed of voxels along the direction of the vehicle's movement, and obtain the dynamic risk factor that is positively correlated with the relative approach speed.

[0014] An obstacle risk level assessment system based on three-dimensional voxel occupancy, used to implement the aforementioned obstacle risk level assessment method, includes: The 3D voxel occupancy map construction module is used to acquire information about the environment around the vehicle, perform voxel modeling of the environmental space through a 3D occupancy prediction network, and construct a 3D voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space. The reachable trajectory set acquisition module is used to acquire the current motion state information of the vehicle and, based on the motion state information, acquire the set of reachable trajectories within a preset future time window; The trajectory influence area determination module performs spatial correlation analysis on the 3D voxel occupancy map and the set of reachable trajectories to determine whether the voxel is located within the trajectory influence area of ​​the vehicle. The comprehensive risk score acquisition module is used to comprehensively quantify the risk level of voxels located within the influence area of ​​the vehicle trajectory from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, and obtain a voxel-level comprehensive risk score. The obstacle risk level assessment module is used to assess risk levels based on voxel-level comprehensive risk scores.

[0015] Preferably, the reachable trajectory set acquisition module includes: The vehicle kinematics model is used to generate trajectories based on motion state information. In the trajectory prediction process, discrete time steps are used to recursively calculate the vehicle pose at several future moments, and physical constraints are introduced to ensure that the generated reachable trajectory conforms to the actual feasible motion range of the vehicle. The candidate motion trajectory acquisition unit is used to enumerate and model the driving behaviors that may occur within a preset time window based on the vehicle's current speed, heading angle and steering state, and generate corresponding candidate motion trajectories respectively. The reachable trajectory set acquisition unit is used to acquire a set of reachable trajectories based on candidate motion trajectories.

[0016] Preferably, the trajectory influence area determination module includes: The occupancy analysis unit is used to obtain the occupancy probability distribution of each voxel at multiple discrete time points in the future, and to characterize the potential occupancy status of the voxel as it evolves over time. The spatial distance acquisition unit is used to match the predicted position of the voxel at the corresponding future time with the trajectory points with the same or similar arrival times in the set of reachable trajectories of the vehicle at each prediction time step in a unified coordinate system, and calculate the spatial distance between the voxel center point and the trajectory point. The obstacle occupancy determination unit is used to determine whether an area is likely to be occupied by an obstacle when the vehicle is expected to pass through it by aligning the voxel prediction occupancy state with the arrival time of the vehicle's trajectory in the time dimension.

[0017] Preferably, the comprehensive risk score acquisition module includes: The spatial risk factor acquisition unit is used to obtain the minimum spatial distance from the voxel to the trajectory within a preset time window, normalize the minimum spatial distance, and obtain the spatial risk factor that is negatively correlated with the minimum spatial distance. The time risk factor acquisition unit is used to obtain the estimated arrival time corresponding to the trajectory point, normalize the estimated arrival time, and obtain the time risk factor that is negatively correlated with the estimated arrival time. The occupation risk factor acquisition unit is used to acquire the voxel occupation probability, which reflects the credibility of a voxel being actually occupied by an obstacle within a future preset time window, and to acquire the occupation risk factor that is positively correlated with the voxel occupation probability. The dynamic risk factor acquisition unit is used to acquire the relative approach speed of voxels along the direction of the vehicle's movement, and to acquire dynamic risk factors that are positively correlated with the relative approach speed. Attached Figure Description

[0018] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0019] Figure 1 This is a schematic diagram of the overall process of an obstacle risk level assessment method based on three-dimensional voxel occupancy in one embodiment of the present invention; Figure 2 This is a flowchart illustrating a specific method for constructing a three-dimensional voxel occupancy map in one embodiment of the present invention; Figure 3 This is a flowchart illustrating a specific method for obtaining a set of reachable trajectories in one embodiment of the present invention; Figure 4 This is a flowchart illustrating a specific method for spatial correlation analysis of a three-dimensional voxel occupancy map and a set of reachable trajectories in one embodiment of the present invention. Figure 5 This is a flowchart illustrating a specific method for determining whether a voxel is located within the influence area of ​​the vehicle trajectory in one embodiment of the present invention; Figure 6 This is a flowchart illustrating a specific method for comprehensively quantifying the degree of risk in one embodiment of the present invention; Figure 7 This is a system flowchart according to an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the relationship between voxel risk level and vehicle trajectory space in one embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to its embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0021] 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. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] In this invention, "first" and "second" do not represent a specific quantity or order, but are merely used to distinguish names.

[0023] Before describing the specific embodiments of the present invention, a brief introduction to the prior art will be given first.

[0024] In existing autonomous driving systems, environmental risk assessment is usually based on target-level perception results, that is, obstacle risk is judged based on obstacle bounding boxes, categories and trajectory information output by target detection and tracking, combined with indicators such as time-to-collision (TTC) and minimum safe distance.

[0025] However, this type of method has the following shortcomings: 1. Over-reliance on object detection results Obstacles must first be classified and modeled at the target level. It is difficult to effectively handle obstacles with irregular shapes, unknown categories, or ambiguous boundaries, such as curbs, construction facilities, and irregularly shaped obstacles.

[0026] 2. Coarse-grained risk assessment Existing methods mostly assess risk based on "targets," which makes it difficult to describe the different levels of risk posed by obstacles to vehicle passage in local areas and lacks the ability to express the distribution of continuous risks in space.

[0027] 3. Low coupling with decision-making and planning Risk assessment results are usually still output in the form of a target list. The planning module needs to perform spatial collision detection and feasibility assessment again, which leads to computational redundancy and increases system latency.

[0028] 4. Insufficient robustness in scenarios with weak maps or without high-precision maps. Target-level risk assessment often relies on road boundaries and semantic priors provided by high-precision maps. In scenarios where maps are missing, offset, or invalid, the stability of risk assessment results is poor.

[0029] Therefore, there is an urgent need for a risk assessment method that does not rely on target semantics, can directly characterize obstacle risks at the spatial level, and is highly compatible with the construction of decision-feasible domains.

[0030] One objective of this invention is to directly characterize obstacle risks at the spatial level, refining the coarse-grained nature of risk assessment. To this end, such as... Figure 1 As shown, an embodiment of the present invention provides a method for assessing obstacle risk level based on three-dimensional voxel occupancy, comprising the following steps: S1: Obtain information about the vehicle's surrounding environment, and use a 3D occupancy prediction network to perform voxel modeling of the environmental space, constructing a 3D voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space.

[0031] As a preferred technical solution, such as Figure 2 As shown, the specific methods for constructing a 3D voxel occupancy map include: S11 divides the preset space around the vehicle into a regular three-dimensional voxel grid.

[0032] S12, for each voxel, record the probability value of the voxel being occupied by an obstacle, and construct a three-dimensional voxel occupancy map based on the probability value.

[0033] Specifically, information about the vehicle's surrounding environment can be acquired using multi-camera visual sensors, and a 3D occupancy prediction network can be used to create a voxel-based model of the environment. The pre-defined space around the vehicle is divided into a regular 3D voxel grid, with each voxel recording its probability of being occupied by an obstacle, forming a 3D voxel occupancy map. This occupancy map can simultaneously describe the spatial distribution of both static and dynamic obstacles.

[0034] S2, obtain the current motion state information of the vehicle and, based on the motion state information, obtain the set of reachable trajectories within a preset future time window.

[0035] As a preferred technical solution, such as Figure 3 As shown, the specific methods for obtaining the set of reachable trajectories include: S21, Construct a vehicle kinematics model, using motion state information as the initial condition for trajectory generation and inputting it into the vehicle kinematics model.

[0036] S22 uses discrete time steps to recursively calculate the vehicle pose at several future moments and introduces physical constraints to ensure that the generated reachable trajectory conforms to the actual feasible range of vehicle movement.

[0037] S23, based on the vehicle's current speed, heading angle and steering state, enumerate and model the possible driving behaviors that may occur within the future preset time window, and generate corresponding candidate motion trajectories.

[0038] S24. Obtain the set of reachable trajectories based on the candidate motion trajectories.

[0039] Specifically, based on the vehicle's current motion status information, a set of reachable trajectories for the vehicle within a preset future time window can be constructed.

[0040] The system acquires real-time state parameters of the vehicle, including longitudinal velocity, longitudinal acceleration, heading angle, yaw rate, and front wheel steering angle, and inputs these parameters into the vehicle's kinematic model as initial conditions for trajectory generation. During trajectory prediction, the system recursively calculates the vehicle's pose at several future moments using discrete time steps. Simultaneously, physical constraints such as wheelbase, maximum steering angle, minimum turning radius, and maximum permissible lateral acceleration are introduced to ensure that the generated trajectory conforms to the vehicle's actual feasible range of motion.

[0041] Based on this, the system enumerates and models possible driving behaviors within a preset time window according to the vehicle's current speed, heading angle, and steering status, and generates corresponding candidate motion trajectories. The candidate trajectories include at least a straight trajectory that maintains the current heading, left-turn and right-turn curve trajectories generated based on different steering angle change rates, and lane-changing trajectories formed within a preset lateral displacement range, to cover various movement modes that the vehicle may take in the short term.

[0042] Each trajectory branch consists of a series of discrete trajectory points arranged in chronological order. Each trajectory point contains its corresponding spatial location and estimated arrival time information, thus describing the area that the vehicle may traverse in the future in both spatial and temporal dimensions. By generating this set of reachable trajectories, a clear reference path is provided for subsequent voxel-based risk assessment, enabling risk calculation to focus on the vehicle's potential range of motion, rather than performing an indiscriminate assessment of the entire scenario.

[0043] S3 performs spatial correlation analysis on the 3D voxel occupancy map and the set of reachable trajectories to determine whether the voxels are located within the influence area of ​​the vehicle trajectory.

[0044] As a preferred technical solution, such as Figure 4 As shown, the specific methods for spatial correlation analysis of the 3D voxel occupancy map and the set of reachable trajectories include: S31 obtains the occupancy probability distribution of each voxel at multiple discrete time points in the future, characterizing the potential occupancy of voxels as they evolve over time.

[0045] S32, under a unified coordinate system, for each predicted time step, the predicted position of the voxel at the corresponding future time is matched with the trajectory points in the set of reachable trajectories of the vehicle that have the same or similar arrival times, and the spatial distance between the voxel center point and the trajectory point is calculated.

[0046] S33, by aligning the voxel prediction occupancy state with the arrival time of the vehicle trajectory in the time dimension, it is determined whether the area is likely to be occupied by an obstacle when the vehicle is expected to pass through a certain area.

[0047] Specifically, after generating the set of future reachable trajectories for the vehicle, the system further performs spatial correlation analysis on the three-dimensional voxel occupancy map and the set of reachable trajectories based on the voxel occupancy prediction results within a preset future time window.

[0048] It should be noted that the system does not directly use the static voxel occupancy state at the current moment, but uses multi-frame occupancy prediction or temporal occupancy model to obtain the occupancy probability distribution of each voxel at multiple discrete time points in the future, thereby characterizing the potential occupancy situation of voxels as they evolve over time.

[0049] In a unified coordinate system, for each predicted time step, the system matches the predicted occupancy position of a voxel at the corresponding future time with trajectory points in the vehicle's reachable trajectory set that have the same or similar arrival times, and calculates the spatial distance between the voxel's center point and the trajectory point. By aligning the predicted voxel occupancy state with the vehicle's trajectory arrival time in the time dimension, the system can determine whether an area is likely to be occupied by an obstacle when the vehicle is expected to pass through it.

[0050] As a preferred technical solution, such as Figure 5 As shown, the specific methods for determining whether a voxel is located within the influence area of ​​the vehicle's trajectory include: S34, when the probability of a voxel occupying a position within a future preset time window is higher than a first preset probability threshold, and the spatial distance between the predicted position and the trajectory point at the corresponding time is less than a preset influence range, it is determined that the voxel is located within the influence area of ​​the vehicle trajectory. S35, for a voxel that is always far away from the potential movement path of the vehicle or whose probability of occupying the path is less than the second preset probability threshold within a future preset time window, it is determined that the voxel is not located in the influence area of ​​the vehicle trajectory.

[0051] Specifically, when the probability of a voxel occupying a position within a future prediction time window is higher than a first preset probability threshold, and the spatial distance between its predicted position and the trajectory point at the corresponding time is less than a preset influence range, the voxel is determined to be within the influence area of ​​the vehicle's trajectory. Conversely, for voxels that are consistently far from the vehicle's potential path or have a low probability of occupying a position in future predictions (i.e., consistently far from the vehicle's potential path or have a probability of occupying a position less than a second preset probability threshold), they can be considered not to constitute an effective constraint on the vehicle's driving decision within the current time window. The first preset probability threshold is greater than the second preset probability threshold.

[0052] The system only performs subsequent risk scoring and classification processes on voxels that are determined to be located within the trajectory influence area, thereby significantly reducing the overall computational scale and real-time computational burden while ensuring the accuracy of risk assessment.

[0053] S4 comprehensively quantifies the risk level of voxels located within the influence area of ​​the vehicle's trajectory from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, to obtain a voxel-level comprehensive risk score.

[0054] Specifically, for each voxel to be evaluated within the influence range of the vehicle's trajectory, the system comprehensively quantifies its risk level from four aspects: spatial proximity, time urgency, occupancy credibility, and dynamic change trend, forming a unified voxel risk score.

[0055] As a preferred technical solution, such as Figure 6 As shown, specific methods for comprehensively quantifying the degree of risk include: S41, obtain the minimum spatial distance from the voxel to the trajectory within a preset time window, normalize the minimum spatial distance, and obtain the spatial risk factor that is negatively correlated with the minimum spatial distance.

[0056] Specifically, based on the predicted occupancy of voxels at multiple discrete time steps in the future, the spatial proximity between voxels and the future reachable trajectory of the vehicle is calculated.

[0057] Assuming voxels are predicted at future time steps The coordinates of the center point are The set of reachable trajectories for a vehicle consists of several discrete trajectory points. Composition, in which Indicates the first... The reachable trajectory is in the 1st The spatial coordinate vectors corresponding to each discrete time step. Then, the minimum spatial distance from a voxel to the trajectory within the prediction time window can be defined as:

[0058] in, The smaller the value, the closer the voxel is to the vehicle's potential driving path, and the higher its spatial risk.

[0059] To facilitate subsequent integration, the minimum spatial distance can be... Mapped to normalized spatial risk factors:

[0060] in, The distance threshold scale parameter is used to characterize the spatial range in which voxels have a significant impact on the vehicle's driving path.

[0061] S42, obtain the estimated arrival time corresponding to the trajectory point, normalize the estimated arrival time, and obtain the time risk factor that is negatively correlated with the estimated arrival time.

[0062] After determining the trajectory point corresponding to the minimum spatial distance between the predicted voxel location and the vehicle's trajectory, the system obtains the estimated arrival time of that trajectory point. This time is used to characterize the urgency of the vehicle's approach to the predicted voxel location during future travel. Let this estimated arrival time be... Then time risk factor It can be represented as: in This is a time-scale parameter.

[0063] S43, obtain the voxel occupancy probability, which reflects the credibility of a voxel being actually occupied by an obstacle within a future preset time window, and obtain the occupancy risk factor that is positively correlated with the voxel occupancy probability.

[0064] The voxel occupancy probability can also be understood as occupancy confidence, which reflects the degree of confidence that the voxel will actually be occupied by an obstacle within a future time window.

[0065] Let the occupancy probability of a voxel be... Then it occupies the risk factor It can be directly expressed as:

[0066] When the voxel occupancy probability is low, its risk contribution decreases accordingly to avoid making overly conservative judgments on low-confidence perceived results.

[0067] S44, obtain the relative approach speed of the voxel along the direction of the vehicle's movement, and obtain the dynamic risk factor that is positively correlated with the relative approach speed.

[0068] For dynamic voxels exhibiting time-varying characteristics, we further analyze their changing trends relative to the vehicle's motion direction. Let the estimated velocity vector of the voxel be... The velocity vector of the vehicle at the corresponding trajectory point is The unit vector pointing the voxel to the direction of the vehicle is The relative approach velocity of the voxels along that direction Defined as: When voxels show a trend of converging towards the vehicle's driving path, their dynamic risk factors... The corresponding increase can be expressed as:

[0069] in For velocity scale parameters.

[0070] S5 assesses obstacle risk levels based on voxel-level comprehensive risk scores.

[0071] Specifically, the aforementioned risk factors can be fused using a preset risk scoring function to obtain a voxel-level comprehensive risk score. .

[0072] For example, a weighted fusion method can be used:

[0073] in , which are weighting coefficients used to adjust the degree of influence of each risk factor on the final score.

[0074] This voxel-level comprehensive risk score serves as the basis for subsequent voxel risk level determination and feasible region construction. In this way, by directly using the risk results to construct the feasible region, redundant collision detection calculations in the planning module can be reduced, thus lowering system computational redundancy.

[0075] For example, given a threshold Then the risk level can be determined. Define it as follows:

[0076] Based on the acquired risk level, a three-dimensional voxel risk level distribution map is output, and this three-dimensional voxel risk level distribution map is used as input to the pre-decision module to generate the safe and feasible domain of the vehicle in the future time.

[0077] The obstacle risk level assessment method based on three-dimensional voxel occupancy uses a three-dimensional occupancy prediction network to perform voxel modeling of the environmental space, constructing a three-dimensional voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space. This method breaks through the traditional target-level risk assessment method, allowing for direct risk quantification in three-dimensional voxel space, characterizing the risk distribution in continuous space, and is applicable to obstacles of any shape. This enables the transformation from target-level to voxel-level risk assessment, which is beneficial for improving the accuracy of risk assessment.

[0078] Furthermore, by performing spatial correlation analysis on the three-dimensional voxel occupancy map and the set of reachable trajectories, it is determined whether the voxel is located within the influence area of ​​the vehicle's trajectory. Then, for the voxels located within the influence area of ​​the vehicle's trajectory, the risk level is comprehensively quantified from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, to obtain a voxel-level comprehensive risk score. The obstacle risk level assessment method based on three-dimensional voxel occupancy described in this invention can realize a risk discrimination mechanism driven by the vehicle's trajectory, so that the risk assessment no longer relies solely on the obstacle's own attributes, but takes its influence on the future trajectory of the vehicle as the core basis, thereby making the risk assessment results more in line with the needs of driving decisions. It has stronger robustness to unknown obstacles, irregular structures, and occluded scenes, and enhances the adaptability to complex scenes.

[0079] Because spatial and temporal risk factors employ exponential decay functions, risk values ​​may decay too rapidly over long distances, neglecting potential long-tail risks. As a preferred technical solution, piecewise nonlinear functions or adaptive decay mechanisms can be introduced to more accurately capture risk changes at different distances and time scales.

[0080] For example, the space risk factor is expressed as .in, These represent the distance threshold (e.g., 5 meters) and the proximity scale parameter, respectively. An adjustable index (e.g., set to 0.8 and 1.5 respectively) is used to control the gradual decay of long-distance risks. In this way, by introducing a piecewise nonlinear function, the granularity of risk assessment can be improved, especially in edge scenarios (such as slowly approaching obstacles at a distance) to reduce misjudgments.

[0081] The time risk factor is expressed as .in, These represent the slope parameter and the critical time (e.g., the vehicle's reaction time), respectively. Thus, based on the time risk factor function... This allows the risk to rise sharply near the critical time, making it more suitable for emergency driving scenarios.

[0082] An embodiment of the present invention also provides an obstacle risk level assessment system based on three-dimensional voxel occupancy, used to implement the obstacle risk level assessment method, which includes a three-dimensional voxel occupancy map construction module, a reachable trajectory set acquisition module, a trajectory influence area determination module, a comprehensive risk score acquisition module, and an obstacle risk level assessment module.

[0083] The 3D voxel occupancy map construction module is used to acquire information about the environment around the vehicle. It performs voxel modeling of the environmental space through a 3D occupancy prediction network to construct a 3D voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space.

[0084] The reachable trajectory set acquisition module is used to acquire the current motion state information of the vehicle and, based on the motion state information, acquire the reachable trajectory set within a preset time window in the future; the trajectory influence area determination module performs spatial correlation analysis on the three-dimensional voxel occupancy map and the reachable trajectory set to determine whether the voxel is located within the trajectory influence area of ​​the vehicle.

[0085] The comprehensive risk score acquisition module is used to comprehensively quantify the risk level of voxels located within the influence area of ​​the vehicle trajectory from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, and obtain a voxel-level comprehensive risk score; the obstacle risk level assessment module is used to assess the risk level based on the voxel-level comprehensive risk score.

[0086] like Figure 7 As shown, the system performs refined risk assessment of environmental obstacles based on three-dimensional voxel occupancy representation. First, it uses a multi-camera visual sensor to acquire information about the surrounding environment, constructs a three-dimensional voxel occupancy map containing spatial location and occupancy probability through a three-dimensional occupancy prediction network, and predicts its evolution in future time series.

[0087] Simultaneously, the system combines the vehicle's current speed, acceleration, heading angle, and steering angle to generate multiple reachable trajectories within a preset future time window under the vehicle's kinematic model and physical constraints, and then discretizes these trajectories. Subsequently, the predicted trajectory points are spatiotemporally aligned with the future voxel occupancy results to filter out voxel regions that may affect the vehicle. For these voxels, risk factors are calculated from multiple dimensions, including spatial proximity, time urgency, occupancy probability reliability, and dynamic approach trend, and a weighted fusion is used to obtain a voxel-level comprehensive risk score. Finally, the risks are classified according to preset rules to form a three-dimensional voxel risk distribution, which is used to constrain the planning and decision-making process.

[0088] Specifically, the reachable trajectory set acquisition module includes a vehicle kinematics model, a candidate motion trajectory acquisition unit, and a reachable trajectory set acquisition unit.

[0089] The vehicle kinematics model is used to generate trajectories based on motion state information. In the trajectory prediction process, discrete time steps are used to recursively calculate the vehicle pose at several future moments, and physical constraints are introduced to ensure that the generated reachable trajectory conforms to the actual feasible motion range of the vehicle.

[0090] The candidate motion trajectory acquisition unit is used to enumerate and model possible driving behaviors within a preset time window based on the vehicle's current speed, heading angle, and steering state, and generate corresponding candidate motion trajectories; the reachable trajectory set acquisition unit is used to acquire a reachable trajectory set based on the candidate motion trajectories.

[0091] The trajectory influence area determination module includes an occupancy analysis unit, a spatial distance acquisition unit, and an obstacle occupancy determination unit.

[0092] The occupancy analysis unit is used to obtain the occupancy probability distribution of each voxel at multiple discrete time points in the future, and to characterize the potential occupancy status of the voxel as it evolves over time.

[0093] The spatial distance acquisition unit is used to match the predicted position of the voxel at the corresponding future time with the trajectory points with the same or similar arrival times in the set of reachable trajectories of the vehicle at each prediction time step in a unified coordinate system, and calculate the spatial distance between the voxel center point and the trajectory point.

[0094] The obstacle occupancy determination unit is used to determine whether an area is likely to be occupied by an obstacle when the vehicle is expected to pass through it by aligning the voxel prediction occupancy state with the arrival time of the vehicle's trajectory in the time dimension.

[0095] The comprehensive risk score acquisition module includes a spatial risk factor acquisition unit, a temporal risk factor acquisition unit, a location risk factor acquisition unit, and a dynamic risk factor acquisition unit.

[0096] The spatial risk factor acquisition unit is used to obtain the minimum spatial distance from the voxel to the trajectory within a preset time window, normalize the minimum spatial distance, and obtain the spatial risk factor negatively correlated with the minimum spatial distance; the temporal risk factor acquisition unit is used to obtain the expected arrival time corresponding to the trajectory point, normalize the expected arrival time, and obtain the temporal risk factor negatively correlated with the expected arrival time.

[0097] The occupancy risk factor acquisition unit is used to acquire the voxel occupancy probability, which reflects the credibility of the voxel being actually occupied by an obstacle within a preset time window in the future, and to acquire the occupancy risk factor that is positively correlated with the voxel occupancy probability; the dynamic risk factor acquisition unit is used to acquire the relative approach speed of the voxel along the direction of the vehicle's movement, and to acquire the dynamic risk factor that is positively correlated with the relative approach speed.

[0098] like Figure 8 As shown, the system uses the vehicle's current position as the starting point and generates a predicted driving trajectory for the vehicle within a preset future time window based on the vehicle's current motion state. The environmental space is discretized into regular three-dimensional voxel units, with each voxel describing the state of the corresponding spatial region being occupied by obstacles at the current moment.

[0099] Under a unified coordinate system, and combining the voxel occupancy prediction results within the future time window, the spatial distance and temporal correlation between each voxel and the predicted trajectory of the vehicle are analyzed, and the risk level of the voxels is assessed accordingly. Voxels located near the vehicle's future trajectory, with a high probability of occupancy and posing a direct threat to the vehicle's driving safety are assessed as high-risk voxels; voxels with a certain spatial distance from the vehicle's trajectory but still potentially influencing driving decisions are assessed as medium-risk voxels; and voxels far from the vehicle's trajectory are assessed as medium-risk voxels. Voxels that deviate from the predicted trajectory of the vehicle and have little impact on the current driving decision are assessed as low-risk voxels.

[0100] The obstacle risk level assessment system based on three-dimensional voxel occupancy uses a three-dimensional occupancy prediction network to perform voxel modeling of the environmental space, constructing a three-dimensional voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space. It breaks through the traditional target-level risk assessment method, can directly quantify risk in three-dimensional voxel space, and characterize the risk distribution in continuous space, and is applicable to obstacles of any shape.

[0101] Furthermore, by performing spatial correlation analysis on the three-dimensional voxel occupancy map and the set of reachable trajectories, it is determined whether the voxel is located within the influence area of ​​the vehicle's trajectory. Then, for the voxels located within the influence area of ​​the vehicle's trajectory, the risk level is comprehensively quantified from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, to obtain a voxel-level comprehensive risk score. The obstacle risk level assessment method based on three-dimensional voxel occupancy described in this invention can achieve 2. a risk discrimination mechanism based on vehicle trajectory driving, so that the risk assessment no longer relies solely on the obstacle's own attributes, but rather on its influence on the future trajectory of the vehicle as the core basis, thereby making the risk assessment results more in line with driving decision-making needs.

[0102] In summary, the present invention has the following advantages: 1. Propose a voxel-level obstacle risk assessment mechanism: This invention breaks through the traditional target-level risk assessment method. By directly quantifying the risk in three-dimensional voxel space, it can characterize the risk distribution in continuous space and is applicable to obstacles of any shape.

[0103] 2. Risk assessment method based on vehicle trajectory: Risk assessment no longer relies solely on the properties of the obstacle itself, but uses its impact on the future trajectory of the vehicle as the core basis, making the risk assessment results more in line with driving decision-making needs.

[0104] 3. Reduce reliance on high-precision maps and improve system robustness: The risk assessment process of this invention does not rely on the semantic information of high-precision maps, and can run stably under weak map or no map conditions, making it suitable for complex dynamic scenarios.

[0105] 4. Naturally adapted feasible region generation: Voxel-level risk level results can be directly used to eliminate high-risk spatial areas, constructing safe and feasible regions for subsequent planning modules and reducing computational redundancy.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for assessing obstacle risk level based on three-dimensional voxel occupancy, characterized in that, include: Acquire information about the vehicle's surrounding environment, perform voxel modeling of the environment space through a 3D occupancy prediction network, and construct a 3D voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space. Obtain the current motion status information of the vehicle and, based on the motion status information, obtain a set of reachable trajectories within a preset future time window; Spatial correlation analysis is performed on the 3D voxel occupancy map and the set of reachable trajectories to determine whether the voxels are located within the influence area of ​​the vehicle trajectory. For voxels located within the influence area of ​​the vehicle trajectory, the risk level is comprehensively quantified from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, to obtain a voxel-level comprehensive risk score. Obstacle risk level assessment is achieved based on voxel-level comprehensive risk scoring.

2. The obstacle risk level assessment method based on three-dimensional voxel occupancy as described in claim 1, characterized in that, Specific methods for constructing a 3D voxel occupancy map include: The pre-defined space around the vehicle is divided into a regular three-dimensional voxel grid; For each voxel, record the probability value of the voxel being occupied by an obstacle, and construct a three-dimensional voxel occupancy map based on the probability value.

3. The obstacle risk level assessment method based on three-dimensional voxel occupancy as described in claim 1, characterized in that, Specific methods for obtaining the set of reachable trajectories include: Construct a vehicle kinematics model and input the motion state information as the initial condition for trajectory generation into the vehicle kinematics model; The vehicle pose at several future moments is recursively calculated using discrete time steps, and physical constraints are introduced to ensure that the generated reachable trajectory conforms to the actual feasible range of vehicle movement. Based on the vehicle's current speed, heading angle, and steering status, the possible driving behaviors within a future preset time window are enumerated and modeled, and corresponding candidate motion trajectories are generated respectively. Based on the candidate motion trajectories, obtain the set of reachable trajectories.

4. The obstacle risk level assessment method based on three-dimensional voxel occupancy as described in claim 1, characterized in that, Specific methods for spatial correlation analysis of 3D voxel occupancy maps and reachable trajectory sets include: Obtain the occupancy probability distribution of each voxel at multiple discrete time points in the future to characterize the potential occupancy of voxels over time. Under a unified coordinate system, for each predicted time step, the predicted position of the voxel at the corresponding future time is matched with the trajectory points in the set of reachable trajectories of the vehicle with the same or similar arrival times, and the spatial distance between the voxel center point and the trajectory point is calculated. By aligning the voxel prediction occupancy state with the arrival time of the vehicle's trajectory in the time dimension, it is determined whether the area is likely to be occupied by an obstacle when the vehicle is expected to pass through it.

5. The obstacle risk level assessment method based on three-dimensional voxel occupancy as described in claim 4, characterized in that, Specific methods for determining whether a voxel is located within the influence area of ​​the vehicle's trajectory include: When the probability of a voxel occupying a position within a future preset time window is higher than a first preset probability threshold, and the spatial distance between the predicted position and the trajectory point at the corresponding time is less than a preset influence range, the voxel is determined to be located within the influence area of ​​the vehicle trajectory. For voxels that consistently move away from the vehicle's potential movement path or have a probability of occupying a position less than a second preset probability threshold within a future preset time window, it is determined that the voxel is not located within the vehicle's trajectory influence area.

6. The obstacle risk level assessment method based on three-dimensional voxel occupancy as described in claim 5, characterized in that, Specific methods for comprehensively quantifying the degree of risk include: Obtain the minimum spatial distance from the voxel to the trajectory within a preset time window, normalize the minimum spatial distance, and obtain the spatial risk factor that is negatively correlated with the minimum spatial distance. Obtain the estimated arrival time corresponding to the trajectory point, normalize the estimated arrival time, and obtain the time risk factor that is negatively correlated with the estimated arrival time; Obtain the voxel occupancy probability, which reflects the credibility of a voxel being actually occupied by an obstacle within a future preset time window, and obtain the occupancy risk factor that is positively correlated with the voxel occupancy probability. Obtain the relative approach speed of voxels along the direction of the vehicle's movement, and obtain the dynamic risk factor that is positively correlated with the relative approach speed.

7. An obstacle risk level assessment system based on three-dimensional voxel occupancy, used to implement the obstacle risk level assessment method as described in any one of claims 1-6, characterized in that, include: The 3D voxel occupancy map construction module is used to acquire information about the environment around the vehicle, perform voxel modeling of the environmental space through a 3D occupancy prediction network, and construct a 3D voxel occupancy map that simultaneously describes the distribution of static and dynamic obstacles in space. The reachable trajectory set acquisition module is used to acquire the current motion state information of the vehicle and, based on the motion state information, acquire the set of reachable trajectories within a preset future time window; The trajectory influence area determination module performs spatial correlation analysis on the 3D voxel occupancy map and the set of reachable trajectories to determine whether the voxel is located within the trajectory influence area of ​​the vehicle. The comprehensive risk score acquisition module is used to comprehensively quantify the risk level of voxels located within the influence area of ​​the vehicle trajectory from four dimensions: spatial proximity, time urgency, occupancy credibility, and dynamic approach trend, and obtain a voxel-level comprehensive risk score. The obstacle risk level assessment module is used to assess risk levels based on voxel-level comprehensive risk scores.

8. The obstacle risk level assessment system based on three-dimensional voxel occupancy as described in claim 7, characterized in that, The module for obtaining the reachable trajectory set includes: The vehicle kinematics model is used to generate trajectories based on motion state information. In the trajectory prediction process, discrete time steps are used to recursively calculate the vehicle pose at several future moments, and physical constraints are introduced to ensure that the generated reachable trajectory conforms to the actual feasible motion range of the vehicle. The candidate motion trajectory acquisition unit is used to enumerate and model the driving behaviors that may occur within a preset time window based on the vehicle's current speed, heading angle and steering state, and generate corresponding candidate motion trajectories respectively. The reachable trajectory set acquisition unit is used to acquire a set of reachable trajectories based on candidate motion trajectories.

9. The obstacle risk level assessment system based on three-dimensional voxel occupancy as described in claim 8, characterized in that, The trajectory influence area determination module includes: The occupancy analysis unit is used to obtain the occupancy probability distribution of each voxel at multiple discrete time points in the future, and to characterize the potential occupancy status of the voxel as it evolves over time. The spatial distance acquisition unit is used to match the predicted position of the voxel at the corresponding future time with the trajectory points with the same or similar arrival times in the set of reachable trajectories of the vehicle at each prediction time step in a unified coordinate system, and calculate the spatial distance between the voxel center point and the trajectory point. The obstacle occupancy determination unit is used to determine whether an area is likely to be occupied by an obstacle when the vehicle is expected to pass through it by aligning the voxel prediction occupancy state with the arrival time of the vehicle's trajectory in the time dimension.

10. The obstacle risk level assessment system based on three-dimensional voxel occupancy as described in claim 9, characterized in that, The comprehensive risk score acquisition module includes: The spatial risk factor acquisition unit is used to obtain the minimum spatial distance from the voxel to the trajectory within a preset time window, normalize the minimum spatial distance, and obtain the spatial risk factor that is negatively correlated with the minimum spatial distance. The time risk factor acquisition unit is used to obtain the estimated arrival time corresponding to the trajectory point, normalize the estimated arrival time, and obtain the time risk factor that is negatively correlated with the estimated arrival time. The occupation risk factor acquisition unit is used to acquire the voxel occupation probability, which reflects the credibility of a voxel being actually occupied by an obstacle within a future preset time window, and to acquire the occupation risk factor that is positively correlated with the voxel occupation probability. The dynamic risk factor acquisition unit is used to acquire the relative approach speed of voxels along the direction of the vehicle's movement, and to acquire dynamic risk factors that are positively correlated with the relative approach speed.

Images