An intelligent driving trajectory planning method, device and equipment

Abstract

The application relates to a trajectory planning method and device for intelligent driving and a trajectory planning equipment, and belongs to the technical field of intelligent driving. The method comprises the following steps: acquiring multi-sensor data of a target vehicle; taking the target vehicle as the center, mapping the multi-sensor data to a target hash table through sparse voxel hashing coding based on a preset voxel size; projecting the target hash table to a two-dimensional BEV space to generate a plurality of grids; determining the confidence score of each grid based on the independent occupancy confidence of each sensor for each grid; determining the static intrinsic risk, dynamic interaction risk and uncertainty risk of each grid; determining the total risk value of each grid, and constructing a trajectory risk evaluation function based on the total risk value of each grid; and performing trajectory planning of the target vehicle with the minimum trajectory risk evaluation function as the target. The application improves the accuracy of intelligent driving perception, and further improves the safety of trajectory planning.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, and in particular to a trajectory planning method, apparatus, and trajectory planning device for intelligent driving. Background Technology

[0002] Current mainstream autonomous driving perception and planning systems generally adopt a serial processing architecture of "object detection-classification-tracking-prediction-planning". The core logic of this architecture is to first identify and classify target objects in the environment (vehicles, pedestrians, bicycles, etc.), then predict their trajectories, and finally plan paths with the target object as the avoidance center. It focuses on identifying risk factors (target category, motion state, etc.) and lacks the ability to quantify implicit risks such as perception uncertainty (e.g., low confidence in occluded areas, missing local sensor data).

[0003] Existing main perception methods include target-centered perception, occupancy grid perception, and risk field / potential field modeling. Target-centered perception only identifies targets within a preset category set, and obstacles outside the preset category or with abnormal shapes are easily missed or misclassified. Occupancy grid perception has limited information expression, mostly showing "occupied / idle" states, and cannot express the difference in risk level of the same grid area under different directions and interaction distances. Risk field / potential field modeling has limitations in fixed weights and dimensions, using static or regular weights for multi-dimensional risk superposition, and lacks a dynamic weight adjustment mechanism that changes continuously over time. These shortcomings lead to a lack of accuracy in existing intelligent driving perception solutions, which in turn results in insufficient safety in trajectory planning.

[0004] Therefore, improving the accuracy of intelligent driving perception, and thus enhancing the safety of trajectory planning, has become an urgent technical problem to be solved. Summary of the Invention

[0005] In view of this, it is necessary to provide a trajectory planning method, device, and equipment for intelligent driving to improve the accuracy of intelligent driving perception and thus improve the safety of trajectory planning.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a trajectory planning method for intelligent driving, comprising: Acquire multi-sensor data of the target vehicle, and map the multi-sensor data to the target hash table using sparse voxel hash encoding, with the target vehicle as the center and based on the preset voxel size. The target hash table is projected onto the two-dimensional BEV space to generate multiple grids. Based on the independent occupancy confidence of each sensor for each grid, the confidence score of each grid is determined. The static intrinsic risk of each grid is determined based on the distance from each grid to the road boundary and the distance from each grid to the lane centerline. The dynamic interaction risk of each grid is determined based on the collision risk of dynamic points within each grid. The uncertainty risk of each grid is determined based on the confidence score of each grid. Based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, the total risk value of each grid is determined, and a trajectory risk evaluation function is constructed based on the total risk value of each grid. The trajectory planning of the target vehicle is carried out with the goal of minimizing the trajectory risk evaluation function.

[0007] In one possible implementation, determining the confidence score for each grid cell based on the independent occupancy confidence of each sensor for each grid cell includes: Based on the independent occupancy confidence of each sensor for each grid, the confidence detection result of each sensor in each grid is determined; Based on the confidence detection results of each sensor in each grid, determine the cross-validation effective sensor count for each grid. The confidence score for each grid is determined based on the cross-validation valid sensor count for each grid and the independent occupancy confidence of each sensor for each grid.

[0008] In one possible implementation, the independent occupancy confidence of each sensor for each grid is determined based on the output confidence of each sensor within each grid.

[0009] In one possible implementation, the confidence score for each grid cell is determined based on the following formula:

[0010] in, Represents grid Confidence score, Indicates sensor For grid Independent occupancy confidence level, Represents grid Cross-validation effective sensor count, , , This is the preset confidence level coefficient.

[0011] In one possible implementation, determining the total risk value for each grid cell based on its static intrinsic risk, dynamic interaction risk, and uncertainty risk includes: The first weight, second weight, and third weight are determined based on the collision time of each grid cell; The static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid are weighted and fused based on the first weight, the second weight, and the third weight to determine the total risk value of each grid.

[0012] In one possible implementation, the collision time of each grid is determined based on the distance from each grid to the edge of the target vehicle and the projected velocity of the target vehicle in the direction from the center of each grid.

[0013] In one possible implementation, the first weight, the second weight, and the third weight are determined based on the following formula:

[0014]

[0015]

[0016] in, Indicates the first weight. Indicates the second weight. Indicates the third weight. Indicates the first basic weight. Indicates the second basic weight. Indicates the third basic weight. This represents the preset collision time coefficient. Represents grid Collision time, Represents grid Distance to the edge of the target vehicle Indicates the target vehicle to the grid. Projection velocity along the central direction.

[0017] In one possible implementation, the trajectory risk assessment function is determined based on the following formula:

[0018] in, This represents the trajectory risk assessment function. This represents the total risk value of the raster. Indicates acceleration. Indicates the degree of impact. Represents the trajectory. The preset security weighting coefficient, This is the preset comfort level.

[0019] On the other hand, the present invention also provides a trajectory planning device for intelligent driving, comprising: The acquisition module is used to acquire multi-sensor data of the target vehicle. Centered on the target vehicle, based on a preset voxel size, the multi-sensor data is mapped to the target hash table through sparse voxel hash encoding. The first determining module is used to project the target hash table onto the two-dimensional BEV space to generate multiple grids, and determine the confidence score of each grid based on the independent occupancy confidence of each sensor for each grid. The second determination module is used to determine the static intrinsic risk of each grid cell based on the distance from each grid cell to the road boundary and the distance from each grid cell to the lane centerline, determine the dynamic interaction risk of each grid cell based on the collision risk of dynamic points within each grid cell, and determine the uncertainty risk of each grid cell based on the confidence score of each grid cell. The planning module is used to determine the total risk value of each grid based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, and to construct a trajectory risk evaluation function based on the total risk value of each grid. The trajectory planning of the target vehicle is carried out with the goal of minimizing the trajectory risk evaluation function.

[0020] Secondly, the present invention also provides a trajectory planning device, including a data acquisition unit, a memory, and a processor, wherein, The data acquisition device is communicatively connected to the processor and is used to acquire multi-sensor data of the target vehicle; The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the trajectory planning method for intelligent driving described in any of the above implementations.

[0021] The beneficial effects of this invention are as follows: The trajectory planning method, device, and equipment for intelligent driving provided by this invention map multi-sensor data to a target hash table through sparse voxel hash encoding, project the target hash table onto a two-dimensional BEV space to generate multiple grids, compressing the high-density point cloud from three dimensions to two dimensions, reducing the data scale of subsequent risk field calculations. Then, the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid are determined, and the total risk value of each grid is determined. Then, a trajectory risk evaluation function is constructed based on the total risk value of each grid, and trajectory planning for the target vehicle is performed with the goal of minimizing the trajectory risk evaluation function. The entire trajectory planning process no longer focuses on target detection, but instead constructs grids through sensor data and performs the final trajectory planning by calculating the risk value of each grid, thereby ensuring the reliability of perception and improving the accuracy of perception and the safety of trajectory planning. In other words, this invention improves the accuracy of intelligent driving perception, thereby improving the safety of trajectory planning. Attached Figure Description

[0022] Figure 1 A schematic flowchart of an embodiment of the trajectory planning method for intelligent driving provided by the present invention; Figure 2 A schematic flowchart of an embodiment of the trajectory planning process for intelligent driving provided by the present invention; Figure 3 A schematic diagram of an embodiment of the trajectory planning device for intelligent driving provided by the present invention; Figure 4 A schematic diagram of an embodiment of the trajectory planning device provided by the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0025] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0026] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. 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.

[0027] In various specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. Additionally, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0028] This invention provides a trajectory planning method, device, and equipment for intelligent driving, which will be described below.

[0029] Figure 1 A schematic flowchart of an embodiment of the trajectory planning method for intelligent driving provided by the present invention is shown below. Figure 1 As shown, the trajectory planning methods for intelligent driving include: S101. Acquire multi-sensor data of the target vehicle. Centered on the target vehicle, map the multi-sensor data to the target hash table using sparse voxel hash encoding based on a preset voxel size.

[0030] It should be noted that the trajectory planning for intelligent driving provided by this invention can be applied to trajectory planning scenarios for intelligent driving vehicles, especially for driving vehicles on urban roads.

[0031] During trajectory planning, the trajectory planning device (such as a portable terminal installed on a vehicle or an existing in-vehicle terminal) first acquires multi-sensor data (such as perception data from LiDAR and cameras) of the target vehicle. Then, centered on the target vehicle, it maps the multi-sensor data to a target hash table using sparse voxel hashing based on a preset voxel size. The data processing does not focus on target detection, thus improving the accuracy of perception.

[0032] S102. Project the target hash table onto the two-dimensional BEV space to generate multiple grids. Based on the independent occupancy confidence of each sensor for each grid, determine the confidence score of each grid.

[0033] It should be noted that after constructing the target hash table, it can be projected onto a two-dimensional BEV space to generate multiple grids. Then, based on the independent occupancy confidence of each sensor for each grid, the confidence score of each grid is determined. By compressing the high-density point cloud from three dimensions to two dimensions, the data scale of subsequent risk field calculations is reduced.

[0034] S103. Determine the static intrinsic risk of each grid cell based on the distance from each grid cell to the road boundary and the distance from each grid cell to the lane centerline; determine the dynamic interaction risk of each grid cell based on the collision risk of dynamic points within each grid cell; and determine the uncertainty risk of each grid cell based on the confidence score of each grid cell.

[0035] It should be noted that after grid generation is completed, the static intrinsic risk of each grid can be determined based on the distance of each grid to the road boundary and the distance of each grid to the lane centerline; the dynamic interaction risk of each grid can be determined based on the collision risk of points within each grid; and the uncertainty risk of each grid can be determined based on the confidence score of each grid. This provides a basis for subsequent trajectory risk assessment.

[0036] S104. Based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, determine the total risk value of each grid, and construct a trajectory risk evaluation function based on the total risk value of each grid. With the goal of minimizing the trajectory risk evaluation function, perform trajectory planning for the target vehicle.

[0037] It should be noted that: Finally, based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, the total risk value of each grid can be determined. Then, a trajectory risk evaluation function is constructed using the total risk value of each grid. Finally, trajectory planning for the target vehicle is performed with the goal of minimizing the trajectory risk evaluation function. The entire trajectory planning process no longer focuses on target detection, but instead constructs grids using sensor data and calculates the risk value of each grid to perform the final trajectory planning. This ensures the reliability of perception, thereby improving the accuracy of perception and the safety of trajectory planning.

[0038] In summary, the trajectory planning method for intelligent driving provided by this invention maps multi-sensor data to a target hash table using sparse voxel hash encoding, projects the target hash table onto a two-dimensional BEV space to generate multiple grids, compressing the high-density point cloud from three dimensions to two dimensions, reducing the data scale of subsequent risk field calculations. Then, it determines the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, and determines the total risk value of each grid. A trajectory risk evaluation function is then constructed using the total risk value of each grid. The trajectory planning of the target vehicle is then performed with the goal of minimizing the trajectory risk evaluation function. The entire trajectory planning process no longer focuses on target detection, but instead constructs grids using sensor data and calculates the risk value of each grid to perform the final trajectory planning. This ensures the reliability of perception, thereby improving the accuracy of perception and the safety of trajectory planning. In other words, this invention improves the accuracy of intelligent driving perception, and thus improves the safety of trajectory planning.

[0039] In some embodiments of the present invention, determining the confidence score of each grid cell based on the independent occupancy confidence of each sensor for each grid cell includes: Based on the independent occupancy confidence of each sensor for each grid, the confidence detection result of each sensor in each grid is determined; Based on the confidence detection results of each sensor in each grid, determine the cross-validation effective sensor count for each grid. The confidence score for each grid is determined based on the cross-validation valid sensor count for each grid and the independent occupancy confidence of each sensor for each grid.

[0040] It should be noted that: when determining the confidence score of each grid, we can first determine the confidence detection result of each sensor in each grid based on the independent occupancy confidence of each sensor in each grid, then determine the cross-validation effective sensor count of each grid based on the confidence detection result of each sensor in each grid, and finally determine the confidence score of each grid based on the cross-validation effective sensor count of each grid and the independent occupancy confidence of each sensor in each grid.

[0041] In some embodiments of the invention, the independent occupancy confidence of each sensor for each grid is determined based on the output confidence of each sensor within each grid.

[0042] It should be noted that the independent occupancy confidence of each sensor for each grid can be determined based on the output confidence of each sensor within each grid.

[0043] In some embodiments of the present invention, the confidence score of each grid is determined based on the following formula:

[0044] in, Represents grid Confidence score, Indicates sensor For grid Independent occupancy confidence level, Represents grid Cross-validation effective sensor count, , , This is the preset confidence level coefficient.

[0045] It should be noted that the confidence score for each grid cell can be calculated using the formula above.

[0046] In some embodiments of the present invention, determining the total risk value of each grid cell based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid cell includes: The first weight, second weight, and third weight are determined based on the collision time of each grid cell; The static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid are weighted and fused based on the first weight, the second weight, and the third weight to determine the total risk value of each grid.

[0047] It should be noted that when determining the total risk value of each grid, the first weight, the second weight, and the third weight can be determined based on the collision time of each grid. Then, the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid are weighted and fused using the first weight, the second weight, and the third weight to determine the total risk value of each grid.

[0048] In some embodiments of the present invention, the collision time of each grid is determined based on the distance from each grid to the edge of the target vehicle and the projection velocity of the target vehicle toward the center of each grid.

[0049] It should be noted that the collision time of each grid can be determined based on the distance from each grid to the edge of the target vehicle and the projected velocity of the target vehicle in the direction from the center of each grid.

[0050] In some embodiments of the present invention, the first weight, the second weight, and the third weight are determined based on the following formula:

[0051]

[0052]

[0053] in, Indicates the first weight. Indicates the second weight. Indicates the third weight. Indicates the first basic weight. Indicates the second basic weight. Indicates the third basic weight. This represents the preset collision time coefficient. Represents grid Collision time, Represents grid Distance to the edge of the target vehicle Indicates the target vehicle to the grid. Projection velocity along the central direction.

[0054] It should be noted that the first weight, second weight, and third weight can be calculated using the formula above.

[0055] In some embodiments of the present invention, the trajectory risk assessment function is determined based on the following formula:

[0056] in, This represents the trajectory risk assessment function. This represents the total risk value of the raster. Indicates acceleration. Indicates the degree of impact. Represents the trajectory. The preset security weighting coefficient, This is the preset comfort level.

[0057] It should be noted that the trajectory risk assessment function can be calculated based on the above formula.

[0058] The trajectory planning method for intelligent driving provided by the present invention will be described below through specific embodiments.

[0059] This invention does not perform target detection, classification, or tracking operations. Instead, it directly generates a dynamic risk field from raw multi-sensor data and searches for the optimal driving trajectory along the gradient descent direction of the risk field, forming an end-to-end closed loop of "sensor data - raster encoding - risk field - trajectory". Combined with... Figure 2 The process specifically includes the following steps: 1. Perception: Multi-source fusion raster coding without target detection.

[0060] Instead of performing object detection and classification, the multi-sensor data is processed into a unified three-layer raster map, specifically including the following operations: 3D Voxelization and Sparse Hash Compression: A 3D voxel space is established centered on the vehicle, with voxel dimensions preset to N1cm×N1cm×N1cm. Sparse voxel hash encoding is used to map the 3D coordinates to a fixed-size hash table, allocating storage space and computing resources only for non-empty voxels, achieving a data compression rate of no less than E1%. Based on sparsity analysis of different traffic scenarios, voxels are divided into three scale levels: the macro level captures the large-scale layout in the distance, the meso level focuses on the overall changes in the nearby road conditions, and the micro level focuses on local details around the vehicle body, ensuring differentiated processing of sparse and dense areas while taking into account both local details and the global scene.

[0061] Hierarchical grid generation: after being processed by the three-dimensional sparse convolutional network, the voxel features are projected onto the two-dimensional BEV space to generate hierarchical grids. The high-resolution bottom grid (N2m×N2m, covering a range of N3m×N3m) is used for refined perception of close-range obstacles beside the vehicle; the middle grid (N4cm×N4cm, covering a range of N5m×N5m) is used for judging the lane-level drivable area; the low-resolution top grid (N6cm×N6cm, covering a range of N7m×N7m) is used for controlling the macro road conditions in the far field. This method compresses the high-density point cloud from three dimensions to two dimensions, which reduces the data scale for subsequent risk field calculation.

[0062] Double-threshold three-state occupancy determination: each grid establishes a confidence score based on multi-sensor cross-validation :

[0063]

[0064]

[0065]

[0066] wherein the states are divided into three categories: drivable (confidence ≥ K1), unknown (K2 < confidence < K1), and undrivable (confidence ≤ K2), is the independent occupancy confidence of the grid by the m-th sensor is (0~1), is the confidence matrix output by the built-in module of the sensor (e.g., LiDAR point cloud density, camera segmentation ratio multiplied by signal-to-noise ratio, etc.), is the number of valid sensors for cross-validation, M is the total number of sensors, is the high-confidence detection indicator of the m-th sensor, K0 is the high-confidence determination threshold for a single sensor, is a high-confidence coefficient (taken as 1), is a low-confidence coefficient (K2 < < K1), is the confidence coefficient when there is no clear detection ( < K2).

[0067] Unknown state grids are treated as medium risk by default; occupancy detected by a single sensor is marked as low-confidence occupancy, and it needs to be cross-validated by at least two types of sensors to be determined as high-confidence occupancy. In the event of sensor failure, the system calls historical grid data for recursive completion, the recursive duration does not exceed t1 seconds, and if it times out, the corresponding area is marked as high risk.

[0068] 2. Risk field: three-dimensional coupled dynamic risk field model.

[0069] The risk field is the direct basis for planning path decisions. The total risk value of each grid is composed of three dimensions: static intrinsic risk, dynamic interactive risk, and uncertainty risk, coupled through dynamic weights. Static intrinsic risk This expresses the risks posed by road structural properties and fixed obstacles. It is based on the distance from the grid to the road boundary. and distance to the center line of the lane calculate:

[0070] in: (For example, in an urban setting, W1=0.6, W2=0.4). Distance threshold for boundary influence (e.g.) =0.8m (Risk halved for 0.55m) Lane impact distance threshold (e.g.) =1.5m (Risk halved for 1.04m).

[0071] The closer to the road boundary and the farther away from the center line of the lane, the higher the static risk.

[0072] Dynamic interaction risk This characterizes the collision risk posed by moving traffic participants. A multi-source Gaussian kernel stacking method is used for spatial distribution modeling.

[0073]

[0074]

[0075]

[0076]

[0077] In the formula: The influence weight of the kth dynamic point represents the potential hazard and urgency of the collision, reflecting the collision energy and motion mutation capability of the dynamic obstacle; Its standard deviation of influence; , Let be the predicted position of the k-th dynamic point at time t. , It is obtained by the difference of the center displacement of dynamic points in two consecutive frames. , Obtained by velocity difference of three consecutive frames; The effective mass density is estimated by the number of grid cells occupied by dynamic points and the average value of lidar reflection intensity; The area is the area of ​​the connected region of the dynamic point in the top view. The larger the area, the more severe the collision consequences. The absolute speed increases through inter-frame connected component differences; the faster the speed, the greater the kinetic energy, and the more severe the consequences of a collision. The absolute value of the yaw rate is obtained through inter-frame connected component difference. The trajectory of an object making a sharp turn is more difficult to predict than that of an object traveling in a straight line, thus increasing its hazard. This is the turning penalty (values ​​range from 0.5 to 1.0, default is 0.6). Let be the length of the connected domain of the kth dynamic point, and let be the maximum span of the space occupied by the obstacle along the driving direction. The width of the connected domain of the k-th dynamic point is the maximum span of the obstacle in the space perpendicular to the driving direction. The safety time interval constant is used to... Convert to forward risk extension distance (default 2s); The angular velocity coefficient is used to... Convert to lateral risk (default 0.45); There is no need to classify dynamic points; simply sensing their position and motion state is sufficient to overlay the spatial risk distribution of all dynamic objects.

[0078] Uncertainty risk The implicit risks arising from insufficient confidence in representation. Grid-based multi-source evidence fusion confidence. calculate:

[0079] The lower the confidence level (the more severe the sensor occlusion and the greater the degree of data loss), the higher the uncertainty risk. This method achieves continuous quantitative expression of implicit risks in densely populated areas with occlusion and in low-light blind spots of cameras.

[0080] Dynamic coupling weights: The three risk dimensions mentioned above are coupled with time-varying weights, resulting in a total risk value:

[0081]

[0082]

[0083] in ,but .

[0084] For the collision time, Static risk weights ( This is the basic weight for static risk, defaulting to K3, with a value range of [0,1]. Its function is to balance the proportion of static risk in the total risk. For dynamic risk weights ( The dynamic risk base weight, defaults to K4, and takes values ​​in the range [0,1]. Uncertainty risk weights ( λ is the basic weight for uncertainty risk (default is K5, value range [0,1], its function is to balance the risk weight in the uncertainty region), and λ is the collision time coefficient (default is K6, value range >0, its function is to control the influence of collision time on dynamic weight).

[0085] For dynamic grids, an occupancy state function Occ_k(t) is defined, which is a binary function (1 for occupied, 0 for empty) that shows the k-th grid being occupied by a dynamic object at time t. Grids that remain unchanged for multiple consecutive frames and have a constant position are considered static obstacles. Using the historical difference of Occ_k(t), the velocity and acceleration can be obtained. d(i,j) is the distance from the grid to the edge of the vehicle. V_radial is the radial velocity of the grid obtained through the difference of consecutive frames (the projected velocity along the direction from the vehicle to the center of the grid). When V_radial < 0, the object is close to the vehicle.

[0086] The smaller the TTC (the more imminent the potential collision risk), the larger β(t), the higher the weight of dynamic interaction risk in the total risk, and the more sensitive the planner is to immediate threats. The weighted coupling structure simulates the intuitive decision-making logic of human drivers shifting their focus from road geometric constraints (static risk) to dynamic avoidance (dynamic interaction risk) in emergency situations.

[0087] 3. Trajectory: Risk gradient-guided trajectory search.

[0088] Risk gradient descent and cost function: An initial trajectory path is generated from the global navigation path, with the search direction following the gradient descent direction of the risk field. The trajectory evaluation function comprehensively considers safety and comfort.

[0089] in Let the arc length of the trajectory be . For acceleration, Impact level, weighting coefficient and It is responsible for balancing the principle of prioritizing safety with the comfort experience of passengers.

[0090] Constraint relaxation search and collision boundary verification: When the lowest-risk path violates vehicle dynamics constraints, a constraint relaxation mechanism is activated: the trajectory is allowed to enter an area with a risk value not exceeding K7 to search for alternative paths. Each time the risk value exceeds the threshold K8, the comfort constraint weight is reduced by K9, gradually expanding the search space until an executable, dynamically feasible path is found. All candidate trajectories undergo collision boundary verification; any grid cell within the trajectory coverage area with a maximum risk value exceeding K10 is directly discarded.

[0091] Temporal risk field prediction: The temporal convolutional network takes M1 consecutive frames of historical risk fields as input and predicts the risk field in M2 frames within the next t2 seconds. Each frame is spaced t3 seconds apart, forming a dynamic spatiotemporal risk map containing prediction information. The trajectory planning module simultaneously searches in the current risk field and the predicted risk field to achieve forward-looking trajectory selection.

[0092] Error prevention mechanism: three-layer cross-validation.

[0093] Perception layer verification: A grid coding strategy with multi-source evidence cross-verification is adopted to ensure the reliability of grid occupancy determination. When a single sensor's data is abnormal, the evidence weight of that sensor is automatically reduced (weight lower limit K11); if more than two types of sensors fail simultaneously, a degradation mode is triggered, reducing the risk field resolution to 1 / 4 of the original, and the maximum vehicle speed is limited to V1km / h.

[0094] Risk calculation layer verification: When the similarity between two adjacent risk fields is less than K12, an abnormal risk field jump is determined. The historical risk field data of the previous frame is automatically reused and sensor health diagnosis is triggered simultaneously. When the area of ​​a high-risk region suddenly increases by more than K13%, multi-sensor secondary cross-validation is initiated to confirm the authenticity of the risk before updating the risk field.

[0095] Planning layer verification: When the lateral offset of two adjacent planned trajectories exceeds D1m or the longitudinal acceleration difference exceeds a1m / s², a trajectory jump is determined, the previous frame trajectory is automatically reused, and the risk field is recalculated. When the total risk score of the trajectory coverage area exceeds the threshold K14, deceleration control is automatically triggered, and the speed reduction ratio is positively correlated with the degree to which the risk score exceeds the threshold.

[0096] Computing power optimization.

[0097] Sparse computation strategy: Only graticules with risk values ​​≥ K15 are subjected to fine-grained computation and time-series updates, while low-risk graticles directly reuse historical data: computational load is reduced by no less than E2%. FPGA parallel acceleration achieves a single-graticule computation latency of ≤ t4ms and a full-frame risk field computation latency of ≤ t5ms.

[0098] Dynamic resolution adjustment: When the vehicle speed is ≥ V2km / h, the resolution of the middle layer far-field grid is increased to maintain the field of vision coverage during high-speed cruising; when the vehicle speed is below V2km / h, the resolution of the bottom layer near-field grid is increased to enhance the accurate perception of low-speed maneuvering, so as to realize the accurate allocation of computing resources under different vehicle speed conditions.

[0099] This invention abandons the target detection stage, avoiding subsequent cascading failures caused by target omissions, misclassifications, and tracking losses. It maintains stable and reliable risk perception and path planning capabilities in long-tail scenarios such as dense, irregular traffic flow and obstacle deformation / occlusion. Perception uncertainty is incorporated as an independent risk dimension into the core layer of risk field calculation; areas with higher uncertainty have higher risk values, and vehicles automatically move away from blind spots. Dynamic risk weights decay exponentially with collision time; the smaller the TTC (Time To Collision) is, the higher the dynamic risk weight, and the planner automatically shifts its attention from road geometric constraints to immediate collision avoidance. Adaptive adjustment under different vehicle speeds, scenarios, and urgency levels enhances accident prevention capabilities. The collaborative design of sparse voxel hash compression, dynamic resolution control, sparse computing strategies, and FPGA parallel acceleration promises to achieve real-time risk field calculation on low-computing-power platforms.

[0100] To better implement the trajectory planning method for intelligent driving in this embodiment of the invention, based on the trajectory planning method for intelligent driving, correspondingly, as follows: Figure 3 As shown, this embodiment of the invention also provides a trajectory planning device for intelligent driving. The trajectory planning device 300 for intelligent driving includes: The acquisition module 301 is used to acquire multi-sensor data of the target vehicle. Centered on the target vehicle, based on a preset voxel size, the multi-sensor data is mapped to the target hash table through sparse voxel hash encoding. The first determining module 302 is used to project the target hash table onto the two-dimensional BEV space to generate multiple grids, and determine the confidence score of each grid based on the independent occupancy confidence of each sensor for each grid. The second determining module 303 is used to determine the static intrinsic risk of each grid cell based on the distance from each grid cell to the road boundary and the distance from each grid cell to the lane centerline, determine the dynamic interaction risk of each grid cell based on the collision risk of dynamic points within each grid cell, and determine the uncertainty risk of each grid cell based on the confidence score of each grid cell. The planning module 304 is used to determine the total risk value of each grid based on the static intrinsic risk, dynamic interaction risk and uncertainty risk of each grid, and to construct a trajectory risk evaluation function based on the total risk value of each grid. The trajectory planning of the target vehicle is carried out with the goal of minimizing the trajectory risk evaluation function.

[0101] The intelligent driving trajectory planning device 300 provided in the above embodiments can realize the technical solutions described in the above intelligent driving trajectory planning method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above intelligent driving trajectory planning method embodiments, and will not be repeated here.

[0102] like Figure 4 As shown, the present invention also provides a trajectory planning device 400. The trajectory planning device 400 includes a processor 401, a memory 402, a display 403, and a data acquisition device 404. Figure 4 Only some components of the trajectory planning device 400 are shown; however, it should be understood that implementation of all shown components is not required, and more or fewer components may be implemented instead.

[0103] In some embodiments, processor 401 may be a central processing unit (CPU), microprocessor, or other data processing chip, used to run program code stored in memory 402 or process data, such as the trajectory planning method for intelligent driving in this invention.

[0104] In some embodiments, processor 401 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 401 may be local or remote. In some embodiments, processor 401 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-cloud, etc., or any combination thereof.

[0105] In some embodiments, memory 402 may be an internal storage unit of trajectory planning device 400, such as a hard disk or memory of trajectory planning device 400. In other embodiments, memory 402 may also be an external storage device of trajectory planning device 400, such as a pluggable hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. equipped on trajectory planning device 400.

[0106] Furthermore, the memory 402 may include both internal storage units of the trajectory planning device 400 and external storage devices. The memory 402 is used to store the application software and various types of data installed on the trajectory planning device 400.

[0107] In some embodiments, display 403 may be an LED display, a liquid crystal display, a touch-screen liquid crystal display, or an organic light-emitting diode (OLED) touchscreen. Display 403 is used to display information from the trajectory planning device 400 and to display a visual user interface. Components 401-403 of the trajectory planning device 400 communicate with each other via a system bus.

[0108] In some embodiments, the data acquisition device 404 may be a chip or integrated circuit with data transmission and reception functions.

[0109] In one embodiment, when the processor 401 executes the intelligent driving trajectory planning program in the memory 402, the following steps can be implemented: Acquire multi-sensor data of the target vehicle, and map the multi-sensor data to the target hash table using sparse voxel hash encoding, with the target vehicle as the center and based on the preset voxel size. The target hash table is projected onto the two-dimensional BEV space to generate multiple grids. Based on the independent occupancy confidence of each sensor for each grid, the confidence score of each grid is determined. The static intrinsic risk of each grid is determined based on the distance from each grid to the road boundary and the distance from each grid to the lane centerline. The dynamic interaction risk of each grid is determined based on the collision risk of dynamic points within each grid. The uncertainty risk of each grid is determined based on the confidence score of each grid. Based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, the total risk value of each grid is determined, and a trajectory risk evaluation function is constructed based on the total risk value of each grid. The trajectory planning of the target vehicle is carried out with the goal of minimizing the trajectory risk evaluation function.

[0110] It should be understood that when the processor 401 executes the intelligent driving trajectory planning program in the memory 402, in addition to the functions mentioned above, it can also perform other functions, as can be found in the description of the corresponding method embodiments above.

[0111] Furthermore, this embodiment of the invention does not specifically limit the type of trajectory planning device 400 mentioned. The trajectory planning device 400 can be a portable electronic device such as a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, or laptop computer. Exemplary embodiments of portable electronic devices include, but are not limited to, portable electronic devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable electronic devices can also be other portable electronic devices, such as laptop computers with touch-sensitive surfaces (e.g., touch panels). It should also be understood that in some other embodiments of the invention, the trajectory planning device 400 may not be a portable electronic device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0112] The trajectory planning method, device, and equipment for intelligent driving provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A trajectory planning method for intelligent driving, characterized in that, include: Acquire multi-sensor data of the target vehicle, and map the multi-sensor data to the target hash table using sparse voxel hash encoding, with the target vehicle as the center and based on the preset voxel size. The target hash table is projected onto the two-dimensional BEV space to generate multiple grids. Based on the independent occupancy confidence of each sensor for each grid, the confidence score of each grid is determined. The static intrinsic risk of each grid is determined based on the distance from each grid to the road boundary and the distance from each grid to the lane centerline. The dynamic interaction risk of each grid is determined based on the collision risk of dynamic points within each grid. The uncertainty risk of each grid is determined based on the confidence score of each grid. Based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, the total risk value of each grid is determined, and a trajectory risk evaluation function is constructed based on the total risk value of each grid. The trajectory planning of the target vehicle is carried out with the goal of minimizing the trajectory risk evaluation function.

2. The trajectory planning method for intelligent driving according to claim 1, characterized in that, The process of determining the confidence score for each grid cell based on the independent occupancy confidence score of each sensor for each grid cell includes: Based on the independent occupancy confidence of each sensor for each grid, the confidence detection result of each sensor in each grid is determined; Based on the confidence detection results of each sensor in each grid, determine the cross-validation effective sensor count for each grid. The confidence score for each grid is determined based on the cross-validation effective sensor count for each grid and the independent occupancy confidence of each sensor for each grid.

3. The trajectory planning method for intelligent driving according to claim 2, characterized in that, The independent occupancy confidence of each sensor for each grid is determined based on the output confidence of each sensor within each grid.

4. The trajectory planning method for intelligent driving according to claim 2, characterized in that, The confidence score for each grid cell is determined based on the following formula: in, Represents grid Confidence score, Indicates sensor For grid Independent occupancy confidence level, Represents grid Cross-validation effective sensor count, , , This is the preset confidence level coefficient.

5. The trajectory planning method for intelligent driving according to claim 1, characterized in that, The determination of the total risk value for each grid cell based on its static intrinsic risk, dynamic interaction risk, and uncertainty risk includes: The first weight, second weight, and third weight are determined based on the collision time of each grid cell; The static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid are weighted and fused based on the first weight, the second weight, and the third weight to determine the total risk value of each grid.

6. The trajectory planning method for intelligent driving according to claim 5, characterized in that, The collision time for each grid is determined based on the distance from each grid to the edge of the target vehicle and the projected velocity of the target vehicle in the direction from the center of each grid.

7. The trajectory planning method for intelligent driving according to claim 6, characterized in that, The first weight, the second weight, and the third weight are determined based on the following formula: in, Indicates the first weight. Indicates the second weight. Indicates the third weight. Indicates the first basic weight. Indicates the second basic weight. Indicates the third basic weight. This represents the preset collision time coefficient. Represents grid The collision time, Represents grid Distance to the edge of the target vehicle Indicates the target vehicle to the grid. Projection velocity along the central direction.

8. The trajectory planning method for intelligent driving according to claim 1, characterized in that, The trajectory risk assessment function is determined based on the following formula: in, This represents the trajectory risk assessment function. This represents the total risk value of the raster. Indicates acceleration. Indicates the degree of impact. Represents the trajectory. The preset security weighting coefficient, This is the preset comfort level.

9. A trajectory planning device for intelligent driving, characterized in that, include: The acquisition module is used to acquire multi-sensor data of the target vehicle. Centered on the target vehicle, based on a preset voxel size, the multi-sensor data is mapped to the target hash table through sparse voxel hash encoding. The first determining module is used to project the target hash table onto the two-dimensional BEV space to generate multiple grids, and determine the confidence score of each grid based on the independent occupancy confidence of each sensor for each grid. The second determination module is used to determine the static intrinsic risk of each grid cell based on the distance from each grid cell to the road boundary and the distance from each grid cell to the lane centerline, determine the dynamic interaction risk of each grid cell based on the collision risk of dynamic points within each grid cell, and determine the uncertainty risk of each grid cell based on the confidence score of each grid cell. The planning module is used to determine the total risk value of each grid based on the static intrinsic risk, dynamic interaction risk, and uncertainty risk of each grid, and to construct a trajectory risk evaluation function based on the total risk value of each grid. The trajectory planning of the target vehicle is carried out with the goal of minimizing the trajectory risk evaluation function.

10. A trajectory planning device, characterized in that, Includes a data acquisition unit, memory, and processor, among which, The data acquisition device is communicatively connected to the processor and is used to acquire multi-sensor data of the target vehicle; The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the trajectory planning method for intelligent driving as described in any one of claims 1 to 7.

Images