Collision detection method based on motion uncertainty quantization and multi-modal intention

Abstract

The invention relates to a collision detection method based on motion uncertainty quantization and multi-modal intention. The method comprises the steps of U1, motion uncertainty ellipse modeling; u2, carrying out obstacle multi-mode behavior probability identification; u3, dynamic virtual lane generation and obstacle screening; and U4, space-time risk matrix judgment. According to the method, the omission ratio of a structured scene (high speed) is reduced, mistaken braking of temporary deviation (such as well lid avoidance) is effectively avoided, the time consumption of single-time detection is greatly reduced, high-frequency collision detection is supported, and decision delay of the high-speed scene is avoided.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a collision detection method based on motion uncertainty quantification and multimodal intent. Background Technology

[0002] As autonomous driving technology gradually becomes more widespread towards Level 2+ and Level 3, collision detection, as a core module ensuring vehicle driving safety, directly determines the safety, user acceptance, and scenario adaptability of the autonomous driving system. Currently, autonomous vehicles need to operate in a variety of complex scenarios, including structured roads such as highways and urban main roads, as well as unstructured roads such as parking lots, construction sections, and rural roads without lane markings. This poses three core requirements for collision detection technology: dynamic deviation adaptability (it needs to cover motion errors such as vehicle steering clearance and braking delay, as well as positional deviations caused by obstacle perception noise), intent recognition accuracy (it needs to accurately distinguish between obstacle behaviors such as "active lane changing," "temporary deviation," and "stationary occupation," avoiding misjudgments), and scenario universality (it needs to be compatible with diverse scenarios such as those with / without lane markings, and high / low perception accuracy, while ensuring real-time performance).

[0003] To meet the above requirements, current collision detection technology has formed a basic technical system: for dynamic changes in vehicle occupancy, some technologies adjust the bounding box size by linking vehicle speed with the detection time to cover the risk of lateral displacement at different vehicle speeds; for geometric collision accuracy, some technologies use directional bounding boxes (OBB) to accurately represent the outlines of vehicles and obstacles, and combine them with separation axle tests (SAT) to determine whether there is overlap; for detection efficiency, some technologies reduce the number of obstacles to be analyzed and reduce invalid calculations by using lane assignment filtering or tree structure clustering.

[0004] However, the existing technological system still suffers from three common defects that cannot be solved by simple combinations of technologies, resulting in security risks and efficiency bottlenecks in complex scenarios: 1. Unquantified Dynamic Uncertainties: Current technologies often assume "absolute certainty" in the position and attitude of vehicles and obstacles, failing to quantify and cover motion errors (such as 0.2m lateral deviation caused by steering system clearance during low-speed parking, and 3.3m longitudinal deviation corresponding to a 0.1s braking delay during high-speed driving) and perception noise (such as 0.1m positional error of LiDAR at a distance of 10m, and edge blurring offset when a monocular camera detects pedestrians). This results in a false alarm rate of 20%-30% in low-precision scenarios such as rain, obstructions, and long-distance perception, while in high-speed dynamic scenarios, there is a high risk of missed detection due to uncovered biases.

[0005] 2. Limited Dimensions in Obstacle Intent Recognition: Current technologies rely heavily on single features (such as the deviation angle of the movement direction from the lane and the distance from the obstacle to the lane line) when judging the behavior trend of obstacles, without integrating multi-dimensional information such as turn signal status, speed change rate, and historical trajectory curvature. This makes it difficult to effectively distinguish between "active lane change (turn signal on + continuous lateral deviation)" and "temporary deviation (short-term lateral displacement to avoid manhole covers or potholes)," resulting in a false positive rate that has remained at 15%-20% for a long time. This can easily lead to unnecessary deceleration or braking, affecting driving comfort and efficiency.

[0006] 3. Lack of adaptability to unstructured scenarios: Most collision detection technologies rely on fixed lane lines or structured coordinates as screening benchmarks, such as only retaining obstacles in the lane to be driven and adjacent lanes. However, in scenarios without lane lines, such as parking lots and construction sections, there is a lack of effective reference benchmarks. It is necessary to traverse all perceived obstacles for detection, resulting in a 2-3 times increase in computing power compared to structured scenarios. This not only makes it difficult to meet the 100ms-level real-time requirement (a common threshold for low- to mid-range automotive hardware), but also makes it prone to misjudgments due to interference from invalid obstacles.

[0007] The aforementioned shortcomings make it difficult for existing collision detection technologies to fully adapt to the safety and scenario requirements of L2+ and above autonomous driving, and breakthrough technical solutions are urgently needed to address these issues.

[0008] In the prior art, Chinese patent application (application number: 202210784892.3) discloses a collision detection method, device, vehicle, and storage medium. This method obtains the vehicle's current speed at the current moment and determines the collision detection time based on the vehicle's trajectory; it then determines the width of the rectangle corresponding to the collision detection time based on the current speed and the collision detection time; and finally performs collision detection on the vehicle based on the rectangle corresponding to the collision detection time. However, this solution only adjusts the width statically based on vehicle speed, without considering vehicle motion errors (such as steering system clearance causing a 0.5° steering angle error at high speeds, resulting in a lateral deviation of 0.087m within 10m) and braking delay errors (a 0.1s delay corresponds to a 3.3m longitudinal deviation at a speed of 120km / h). This leads to missed collision risks in high-speed scenarios due to insufficient coverage of these errors. Simultaneously, it does not consider obstacle perception errors (such as blurred edges and a 0.15m bounding box offset when a monocular camera detects pedestrians), resulting in a 35% overlap of blank areas in low-speed scenarios and a significantly increased false alarm rate. The solution assumes the obstacle bounding box position is fixed and does not incorporate obstacle speed adjustment logic. For example, when the vehicle and the obstacle are traveling at high speed in the same direction (relative speed <0.5m / s), if at some moment the bounding boxes slightly overlap (overlapping area 0.1m), the obstacle may be affected. 2If two objects are traveling at low speeds (relative speed 1m / s) and are not currently overlapping but will collide in 1 second, the collision risk will be misjudged. The width adjustment logic of the solution is based on "structured road speed characteristics" (such as 20km / h and 60km / h thresholds corresponding to urban areas and highways). In parking lot scenarios without lane lines, the vehicle speed is mostly below 5km / h, and the width is fixed at 0.1m, which cannot cover the turning deviation during parking (which can reach 0.2m), leading to misjudgment of collision risk with obstacles such as curbs and wheel chocks.

[0009] In the prior art, Chinese patent application (application number: 202210904044.1) discloses an autonomous driving collision detection method based on directional bounding boxes. First, relatively reliable obstacle bounding box (OBB) data is acquired in real-time from the vehicle perception system. Second, pre-planned candidate trajectories are obtained from the vehicle trajectory planning module and discretized; for each discrete point, a vehicle OBB bounding box is created. Third, the distance between the vehicle center and the obstacle center is initially calculated to determine whether they are absolutely safe (no collision will occur). Finally, a separation axis test is performed on obstacles and vehicles with potential collisions to determine whether a collision has occurred, and the collision risk is calculated based on the shortest distance between them, providing further decision-making information for the autonomous driving system. However, in this scheme, the vertex coordinates and attitude angles of the OBB rely on high-precision perception data (LiDAR point cloud density ≥ 200 points / m²). 2 (Attitude angle error <0.5°), but in real-world scenarios: at night or in foggy weather, the lidar point cloud density drops to 50 points / m. 2 The following issues arise: OBB attitude angle error increases to over 1.5°, leading to deviations in projection interval calculations and misclassifying non-collision scenes as collisions (a 40% increase in misclassification rate); when a monocular camera detects distant obstacles (>15m), vertex coordinate error reaches 0.2m, OBB contour distortion occurs, and the misclassification rate for overlapping blank areas reaches 30%. SAT testing requires traversing the normal vector axes of 8 edges, and each axis needs to complete three steps: vertex projection, interval calculation, and overlap judgment, with a single detection taking 180-220ms; while the real-time threshold of commonly used mid-to-low-end automotive MCUs for L2 autonomous driving (such as Texas Instruments TDA4VM) is 100ms, this solution cannot meet the requirements of high-frequency collision detection (such as a 10Hz detection frequency), resulting in decision delay. The solution does not differentiate between scenario types. It performs "pre-screening + SAT testing" on all obstacles (including distant cones and vehicles in idle areas) in unstructured scenarios (such as construction sections). The number of objects to be detected is 3-4 times that in structured scenarios, which further increases the computing power consumption and may even lead to overload of vehicle hardware. Summary of the Invention

[0010] In view of the shortcomings of the prior art, the present invention provides a collision detection method based on motion uncertainty quantification and multimodal intent, which not only reduces the false detection rate in structured scenes (high speed) and effectively avoids false braking of temporary deviations (such as avoiding manhole covers), but also significantly reduces the time consumed per detection, supports high-frequency collision detection, and avoids decision delay in high-speed scenes.

[0011] To achieve the above and other related objectives, the present invention provides the following technical solution: A collision detection method based on motion uncertainty quantification and multimodal intent, the method comprising: U1. Ellipse modeling for motion uncertainty; U2. Obstacle multimodal behavior probability recognition; U3. Dynamic virtual lane generation and obstacle selection; U4. Spatiotemporal Risk Matrix Judgment.

[0012] Furthermore, in step U1, the motion uncertainty ellipse modeling includes: U11. Vehicle Uncertainty Ellipse Generation: Based on the vehicle's current speed V, steering angle, and IMU calibration error, the relevant deviation amplitude is calculated. The longitudinal deviation 'a' is, , The lateral deviation b is, , Among them, t pred To predict the duration, σ a For longitudinal acceleration error, σ b The lateral acceleration error is given by L, which is the wheelbase of the vehicle, and the distance from the rear axle center (x) is taken as the reference point. c ,y c Using as the center of the ellipse, generate an ellipse representing the uncertainty of the vehicle. , Where x is the horizontal axis variable of the ellipse, and y is the vertical axis variable of the ellipse; U12. Obstacle Uncertainty Ellipse Generation: Input perception quality parameters, LiDAR point cloud density ρ, and camera's confidence level for obstacle recognition conf. If ρ > 100 and conf > 0.9, then high-precision perception bias correction parameters a0 and b0 are obtained; if ρ < 50 or conf < 0.7, then low-precision perception bias correction parameters a1 and b1 are obtained. An obstacle uncertainty ellipse is generated with the obstacle center (x0, y0) as the ellipse center. , ; Where g is the horizontal axis variable parameter of the ellipse, and h is the vertical axis variable parameter of the ellipse; U13. Perform a Minkowski sum between the uncertain ellipse and the vehicle's OBB / polygon shape to obtain the uncertain expanded vehicle body.

[0013] Furthermore, the motion uncertainty ellipse modeling also includes: U14. Generate a spacetime elliptic tube for each discrete sub-segment within Δt, and set it to a lower boundary type TTC: TTC lb =min k TTC (tube) k Replace single-moment TTC to explicitly override braking / steering hysteresis propagation.

[0014] Furthermore, in step U2, the obstacle multimodal behavior probability recognition includes: U21. Multi-dimensional feature extraction: Historical trajectory feature extraction: lateral offset Δy, the lateral displacement of the obstacle relative to the current driving direction in the past 1 second; trajectory curvature past C, the average curvature of the obstacle trajectory in the past 1 second. Real-time state feature extraction, turn signal state S light Rate of change of velocity ; Scene feature extraction: The structured scene is defined by the distance d between the obstacle and the nearest lane line. lane In an unstructured scene, the distance d between an obstacle and other static obstacles is... obs ; U22. Behavioral Probability Calculation and Normalization: The probability of going straight, p str , , Lane change probability p lc , , Parking probability PSTP , Probability normalization, , s=p str +p lc +p stp Take P norm The behavior corresponding to the maximum value in the middle is used as the obstacle "idea graph", and the output probability distribution is used for subsequent risk correction. U23. Time consistency constraint for “active lane change vs. temporary deviation”.

[0015] Furthermore, in step U23, the time consistency constraint of "active lane change vs. temporary deviation" is to introduce a dwell time threshold m before probability normalization. d Second-order consistency with heading change The joint discrimination, if the lateral drift is only in m d A single peak appears within and Then, the lane change probability is suppressed by λ<1 if the turn signal is on and Continuous same sign greater than ω min , where ε, λ and ω min If it is a constant parameter, then the probability of lane change is amplified.

[0016] Furthermore, in step U3, the dynamic virtual lane generation and obstacle selection include: U31. Scene type judgment: In structured scenes, if the perception module detects continuous lane lines, in unstructured scenes, if no valid lane lines are detected, the dynamic virtual lane generation process is executed. U32. Dynamic virtual lane generation; U33. Obstacle Filtering: Keep obstacles that are "within the virtual lane + the main image shows straight / lane change" and exclude obstacles that are "outside the virtual lane + the main image shows parking".

[0017] Furthermore, in step U32, the dynamic virtual lane generation includes: U321. Obstacle Edge Point Clustering: Then, the K-means clustering algorithm (K=2) is used to divide the edge points into "left boundary point set P". left The right boundary point set P of "and" right ”; U322. Perform lane boundary fitting: the left and right boundaries are represented by cubic B-splines, r=[c L , c R The sampling matrix S maps control points to equally spaced arc length parameters s∈[0,L]. A constrained optimization algorithm is used to solve this problem, and its objective function is: , in, As a support point, w0 is the prior lane width, w d w s w ω and w t Let L be the weights of each term, and L be a cubic spline function. and c L and c R The second-order difference matrix of the points, r prev For [c L ,c RThe midpoint of the []. U323. Determine the lane range: Virtual lane width W vir Calculate the distance between the left and right boundaries perpendicular to the vehicle's direction of travel. , Virtual lane centerline Lmid=(L left +L right ) / 2, where b2 is a point on the left boundary of the road, b1 is a point on the right boundary of the road, k2 is the slope of the road, and L left For the cubic spline function fitting curve of the left boundary of the road, L right Fit a curve to a cubic spline function on the right side of the road.

[0018] Furthermore, the width d of the lane i The constraint of (r), ; Lane curvature / turning angle f i (r) constraint, ; Apply a passable region constraint, ensuring that boundary points fall within the passable voxel mask, and impose a minimum spacing constraint on the obstacle set: ; Where r=[c L ,c R ],f max Let L(r) be the maximum curvature of the lane, L(r) be the fitted curve of the lane, O be the coordinates of the center point of the obstacle, and δ be the maximum curvature of the lane. obs This is the minimum spacing.

[0019] Furthermore, in step U4, the spatiotemporal risk matrix decision includes: U41. Basic Parameter Calculation: Calculate the collision time TTC. , Among them, (r o ,g o (a) represents the coordinates of the vehicle. c ,b c (v) represents the coordinates of the obstacle. or ,v og (v) represents the vehicle's speed. ca ,v cb () represents the velocity of the obstacle; if the denominator is zero, then the relative velocity is zero. Perform an uncertainty overlap calculation and calculate the intersection area S between the vehicle ellipse and the obstacle ellipse. inter With the elliptical area S of the vehicle car proportion, ; U42. Calculation of Three-Dimensional Risk Weights: First, calculate the basic risk weights (Risk) that integrate the temporal and spatial dimensions. base , , Where α is the time dimension weight, β is the spatial dimension weight, and TTC max The maximum collision time, Overlap max To determine the maximum degree of uncertainty overlap, we will next adjust the risk weights based on the obstacle intention graph to correct the intent. , Where γ is the correction parameter; U43. Risk Level and Decision Output.

[0020] Furthermore, the risk level and decision output are if Risk final A risk level greater than or equal to 0.8 indicates high risk, particularly during emergency braking of the vehicle; if the Risk... final A risk level greater than or equal to 0.4 and less than 0.8 indicates medium risk, requiring minor path adjustments or slowdowns; if the risk level is... final If it is less than 0.4, continuous monitoring will be conducted.

[0021] The present invention has the following positive effects: 1. This invention significantly reduces false alarms and missed detections by "quantifying motion uncertainty + multimodal intent recognition": in low-perception quality scenarios such as rainy nights, the percentage of blank overlap in obstacle bounding boxes decreases; the false positive rate in unstructured scenarios (parking lots) decreases; and the missed detection rate in structured scenarios (highways) decreases, effectively avoiding false braking of temporary deviations (such as avoiding manhole covers).

[0022] 2. This invention significantly improves real-time performance and reduces computing power by using "dynamic virtual lane filtering + lightweight ellipse intersection calculation": the number of objects to be detected is reduced compared to the traversal approach; the time consumed per detection is greatly reduced, high-frequency collision detection is supported, and decision delays in high-speed scenarios are avoided.

[0023] 3. By breaking through the dependence on structured roads, the scope of scene adaptation is significantly expanded: in addition to existing highway / city roads, unstructured scenes are added, increasing the overall coverage; under low-quality perception conditions such as low point cloud density or low recognition confidence, the accuracy of risk judgment is improved.

[0024] 4. This invention achieves differentiated decision-making by relying on the "spatiotemporal risk matrix", which significantly improves driving comfort: braking is triggered only for high risks, reducing unnecessary braking times; for medium risks, path fine-tuning and speed limiting are used to reduce violent control actions while ensuring safety. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the motion uncertainty ellipse of the present invention; Figure 3 This is a schematic diagram of the multimodal behavior probability recognition logic of the present invention; Figure 4 This is a schematic diagram of the dynamic virtual lane generation of the present invention; Figure 5 This is a schematic diagram of the spatiotemporal risk matrix of the present invention. Detailed Implementation

[0026] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0027] Example 1: As Figure 1 As shown, a collision detection method based on motion uncertainty quantification and multimodal intent is presented, the method comprising: U1. Ellipse modeling for motion uncertainty; U2. Obstacle multimodal behavior probability recognition; U3. Dynamic virtual lane generation and obstacle selection; U4. Spatiotemporal Risk Matrix Judgment.

[0028] In this embodiment, as Figure 2 As shown, in step U1, the motion uncertainty ellipse modeling includes: U11. Vehicle Uncertainty Ellipse Generation: Based on the vehicle's current speed V, steering angle, and IMU calibration error, the relevant deviation amplitude is calculated. The longitudinal deviation 'a' is, , The lateral deviation b is, , Among them, t pred To predict the duration, σ a For longitudinal acceleration error, σ b The lateral acceleration error is given by L, which is the wheelbase of the vehicle, and the distance from the rear axle center (x) is taken as the reference point. c ,y c Using as the center of the ellipse, generate an ellipse representing the uncertainty of the vehicle. , Where x is the horizontal axis variable of the ellipse, and y is the vertical axis variable of the ellipse; U12. Obstacle Uncertainty Ellipse Generation: Input perception quality parameters, LiDAR point cloud density ρ, and camera's confidence level for obstacle recognition conf. If ρ > 100 and conf > 0.9, then high-precision perception bias correction parameters a0 and b0 are obtained (a0 = 0.05m, b0 = 0.05m). If ρ < 50 or conf < 0.7, then low-precision perception bias correction parameters a1 and b1 are obtained (a1 = 0.2m, b1 = 0.2m). An obstacle uncertainty ellipse is generated with the obstacle center (x0, y0) as the ellipse center. , ; Where g is the horizontal axis variable parameter of the ellipse, and h is the vertical axis variable parameter of the ellipse; U13. Perform a Minkowski sum between the uncertain ellipse and the vehicle's OBB / polygon shape to obtain the uncertain expanded vehicle body.

[0029] In this embodiment, the motion uncertainty ellipse modeling further includes: U14. Generate a spacetime elliptic tube for each discrete sub-segment within Δt, and set it to a lower boundary type TTC: TTC lb =min k TTC (tube) k Replace single-moment TTC to explicitly override braking / steering hysteresis propagation.

[0030] Furthermore, in step U2, the obstacle multimodal behavior probability recognition includes: U21. Multi-dimensional feature extraction: Historical trajectory feature extraction: lateral offset Δy (Δy greater than 0.5m, indicating lane change trend); lateral displacement of the obstacle relative to the current driving direction in the past 1 second; trajectory curvature past C (C greater than 0.1rad / m, indicating steering behavior); average curvature of the obstacle trajectory in the past 1 second. Real-time state feature extraction, turn signal state S light 1 = On (lane change intention), 0 = Off, speed change rate ; Scene feature extraction: The structured scene is defined by the distance d between the obstacle and the nearest lane line. lane , (d lane (A distance less than 0.5m indicates proximity to the lane boundary) In unstructured scenarios, this refers to the distance d between an obstacle and other static obstacles. obs ; (d obs (Less than 2m prompts for avoidance behavior) U22. Behavioral Probability Calculation and Normalization: The probability of going straight, p str , , Lane change probability p lc , , Parking probability PSTP , Probability normalization, , s=p str +p lc +p stp Take P norm The behavior corresponding to the maximum value in the middle is used as the obstacle "idea graph", and the output probability distribution is used for subsequent risk correction. U23. Time consistency constraint for “active lane change vs. temporary deviation”.

[0031] In this embodiment, in step U23, the time consistency constraint of "active lane change vs. temporary deviation" is to introduce a dwell time threshold m before probability normalization. d Second-order consistency with heading change The joint discrimination, if the lateral drift is only in m d A single peak appears within and Then, the lane change probability is suppressed by λ<1 if the turn signal is on and Continuous same sign greater than ω min , where ε, λ and ω min If it is a constant parameter, then the probability of lane change is amplified.

[0032] Example 2: Based on the collision detection method based on motion uncertainty quantization and multimodal intent in Example 1, the present invention will be further explained and described below.

[0033] like Figure 1 As shown, a collision detection method based on motion uncertainty quantification and multimodal intent is presented, the method comprising: U1. Ellipse modeling for motion uncertainty; U2. Obstacle multimodal behavior probability recognition; U3. Dynamic virtual lane generation and obstacle selection; U4. Spatiotemporal Risk Matrix Judgment.

[0034] In this embodiment, as Figure 3 or Figure 4As shown, in step U3, the dynamic virtual lane generation and obstacle selection include: U31. Scene type judgment: In structured scenes, if the perception module detects continuous lane lines, in unstructured scenes, if no valid lane lines are detected, the dynamic virtual lane generation process is executed. U32. Dynamic virtual lane generation; U33. Obstacle Filtering: Keep obstacles that are "within the virtual lane + the main image shows straight / lane change" and exclude obstacles that are "outside the virtual lane + the main image shows parking".

[0035] In this embodiment, step U32, the dynamic virtual lane generation includes: U321. Obstacle Edge Point Clustering: Then, the K-means clustering algorithm (K=2) is used to divide the edge points into "left boundary point set P". left The right boundary point set P of "and" right ”; U322. Perform lane boundary fitting: the left and right boundaries are represented by cubic B-splines, r=[c L , c R The sampling matrix S maps control points to equally spaced arc length parameters s∈[0,L]. A constrained optimization algorithm is used to solve this problem, and its objective function is: , in, As a support point, w0 is the prior lane width, w d w s w ω and w t Let L be the weights of each term, and L be a cubic spline function. and c L and c R The second-order difference matrix of the points, r prev For [c L ,c R The midpoint of the []. U323. Determine the lane range: Virtual lane width W vir Calculate the distance between the left and right boundaries perpendicular to the vehicle's direction of travel. , Virtual lane centerline Lmid=(L left +L right ) / 2, where b2 is a point on the left boundary of the road, b1 is a point on the right boundary of the road, k2 is the slope of the road, and L left For the cubic spline function fitting curve of the left boundary of the road, L rightFit a curve to a cubic spline function on the right side of the road.

[0036] In this embodiment, the width d of the lane i The constraint of (r), ; Lane curvature / turning angle f i (r) constraint, ; Apply a passable region constraint, ensuring that boundary points fall within the passable voxel mask, and impose a minimum spacing constraint on the obstacle set: ; Where r=[c L ,c R ],f max Let L(r) be the maximum curvature of the lane, L(r) be the fitted curve of the lane, O be the coordinates of the center point of the obstacle, and δ be the maximum curvature of the lane. obs This is the minimum spacing.

[0037] In this embodiment, as Figure 5 As shown, in step U4, the spatiotemporal risk matrix decision includes: U41. Basic Parameter Calculation: Calculate the collision time TTC. , Among them, (r o ,g o (a) represents the coordinates of the vehicle. c ,b c (v) represents the coordinates of the obstacle. or ,v og (v) represents the vehicle's speed. ca ,v cb () represents the velocity of the obstacle; if the denominator is zero, then the relative velocity is zero. Perform an uncertainty overlap calculation and calculate the intersection area S between the vehicle ellipse and the obstacle ellipse. inter With the elliptical area S of the vehicle car proportion, ; U42. Calculation of Three-Dimensional Risk Weights: First, calculate the basic risk weights (Risk) that integrate the temporal and spatial dimensions. base , , Where α is the time dimension weight, β is the spatial dimension weight, and TTC max The maximum collision time, Overlap max To determine the maximum degree of uncertainty overlap, we will next adjust the risk weights based on the obstacle intention graph to correct the intent. , Where γ is the correction parameter; U43. Risk Level and Decision Output.

[0038] In this embodiment, the risk level and decision output are if Risk final A risk level greater than or equal to 0.8 indicates high risk, particularly during emergency braking of the vehicle; if the Risk... final A risk level greater than or equal to 0.4 and less than 0.8 indicates medium risk, requiring minor path adjustments or slowdowns; if the risk level is... final If it is less than 0.4, continuous monitoring will be conducted.

[0039] The present invention provides a collision detection system based on motion uncertainty quantization and multimodal intent, including a computer device programmed or configured to perform the steps of the collision detection method based on motion uncertainty quantization and multimodal intent.

[0040] The present invention provides a computer-readable storage medium storing a computer program programmed or configured to perform the collision detection method based on motion uncertainty quantification and multimodal intent.

[0041] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0042] In summary, this invention not only reduces the false detection rate in structured scenarios (high speed) and effectively avoids false braking of temporary deviations (such as avoiding manhole covers), but also significantly reduces the time consumed per detection, supports high-frequency collision detection, and avoids decision delays in high-speed scenarios.

[0043] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A method for collision detection based on motion uncertainty quantification with multi-modal intent, the method comprising: The method comprises: U1. Motion uncertainty ellipse modeling; U2. Obstacle multi-modal behavior probability identification; U3. Dynamic virtual lane generation and obstacle screening; U4. Spatio-temporal risk matrix decision.

2. The method for collision detection based on quantifying motion uncertainty and multi-modal intent of claim 1, wherein, In step U1, the motion uncertainty ellipse modeling comprises: U11. Vehicle uncertainty ellipse generation: from the current vehicle speed V, steering angle, IMU calibration error, the relevant deviation amplitude is calculated according to the above input parameters, The longitudinal deviation a is, ， The lateral deviation b is, ， where t pred is the prediction horizon, σ a is the longitudinal acceleration error, σ b is the lateral acceleration error, L is the vehicle wheelbase, and the rear axle center of the ego vehicle (x c ,y c ) is the center of the ellipse, the vehicle uncertainty ellipse is generated, , Where x is the horizontal axis variable of the ellipse, and y is the vertical axis variable of the ellipse; U12. Obstacle uncertainty ellipse generation: input the perception quality parameters, laser radar point cloud density ρ, and camera obstacle recognition confidence conf, If ρ>100 and conf>0.9, high-precision perception deviation correction parameters a0 and b0 are obtained, and if ρ<50 or conf<0.7, low-precision perception deviation correction parameters a1 and b1 are obtained, and the obstacle center (x0, y0) is taken as the ellipse center, and the generated obstacle uncertainty ellipse is, ， ； Where g is the horizontal axis variable parameter of the ellipse, and h is the vertical axis variable parameter of the ellipse; U13. Minkowski sum of uncertainty ellipse and vehicle shape OBB / polygon to obtain uncertainty expansion vehicle body.

3. The method for collision detection based on quantifying motion uncertainty and multi-modal intent of claim 2, wherein, The motion uncertainty ellipse modeling further comprises: U14. Generate a space-time ellipsoidal tube for each discrete sub-segment within Δt, and replace the single time TTC: TTC lb = min k TTC(tube k ) replaces the single time TTC, explicitly covering braking / steering hysteresis propagation.

4. The motion uncertainty quantification based multi-modal intent collision detection method of claim 1, wherein, In step U2, the obstacle multi-modal behavior probability identification comprises: U21. Multi-dimensional feature extraction: Historical trajectory feature extraction, lateral displacement Δy, lateral displacement of the obstacle relative to the current driving direction in the past 1s, trajectory curvature C in the past 1s, average curvature of the obstacle trajectory; Real-time status feature extraction, turn signal status S light , rate of speed change ; Scene feature extraction, structured scene is the distance d of the obstacle and the nearest lane line lane , unstructured scene is the distance d of the obstacle and other static obstacles obs ; U22. Behavior probability calculation and normalization: Straight ahead probability p str , ， Lane change probability p lc , ， Parking probability pstp, ， Probability normalization, ， s=p str +p lc +p stp Take P norm The behavior corresponding to the maximum value in the middle is used as the obstacle "idea graph", and the output probability distribution is used for subsequent risk correction. U23. Time consistency constraint of "active lane change vs. temporary deviation".

5. The motion uncertainty quantification based multi-modal intent collision detection method of claim 4, wherein, In step U23, the time consistency constraint of "active lane change vs. temporary deviation" is to introduce a threshold m of residence duration before probability normalization d , combined with the second-order consistency of heading change , if the lateral drift only appears as a single peak within m d and , then the lane change probability is suppressed by λ<1, if the turn signal is on and the duration of the same sign is greater than ω min , where ε, λ and ω min are constant parameters, then the lane change probability is amplified.

6. The motion uncertainty quantification based multi-modal intent collision detection method of claim 1, wherein, In step U3, the dynamic virtual lane generation and obstacle screening comprises: U31. Scene type judgment: structured scene, if the perception module detects continuous lane lines, unstructured scene: if no valid lane line is detected, the dynamic virtual lane generation process is executed; U32. Dynamic virtual lane generation; U33. Obstacle screening: keep the obstacles "in the virtual lane + main intention is straight / laning", and exclude the obstacles "outside the virtual lane + main intention is parking".

7. The motion uncertainty quantification based and multi-modal intent collision detection method of claim 6, wherein, In step U32, the dynamic virtual lane generation comprises: U321. Obstacle edge point clustering: Then the K-means clustering algorithm (K=2) is adopted to divide the edge points into "left boundary point set P left " and "right boundary point set P right ". U322. Obstacle edge point clustering: Then the K-means clustering algorithm (K=2) is adopted to divide the edge points into "left boundary point set P left " and "right boundary point set P right ". U322. Perform lane boundary fitting: left and right boundaries are represented by cubic B-splines, r=[c L , c R ], the sampling matrix S maps the control points to equi-spaced arc-length parameters s ∈ [0, L], which is solved using a constrained optimization algorithm with the objective function ， wherein, w0is a prior lane width, w d w s w ω and w t are respective weights, L is a cubic spline function, and are second-order difference matrices of c L and c R points, r prev is a middle point in [c L , c R ]. U323. Determining lane extent: virtual lane width W vir , calculating the distance of the left / right boundary in the direction perpendicular to the driving direction of the ego vehicle, ， Virtual lane center line Lmid = (L left + L right ) / 2, where b2 is a point on the left road boundary, bl is a point on the right road boundary, k2 is the slope of the road, L left is the cubic spline function fitted curve of the left road boundary, and L right is the cubic spline function fitted curve of the right road boundary.

8. The motion uncertainty quantification based multi-modal intent collision detection method of claim 7, wherein: Width of the lane d i (r) constraints, ； Lane curvature / turning angle f i (r) constraints, ； The passable domain constraint is performed, the boundary points fall within the passable voxel mask, and the minimum distance constraint is applied to the obstacle set: ； where r = [c L ,c R ], f max is the maximum curvature of the lane, L(r) is the fitted curve of the lane, O is the center point coordinate of the obstacle, and δ obs is the minimum distance.

9. The motion uncertainty quantification based multi-modal intent collision detection method of claim 1, wherein, In step U4, the spatio-temporal risk matrix decision comprises: U41. Basic parameter calculation: collision time TTC calculation is performed, ， wherein (r o ,g o ) is the coordinate point of the vehicle, (a c ,b c ) is the coordinate point of the obstacle, (v or ,v og ) is the speed of the ego vehicle, (v ca ,v cb ) is the speed of the obstacle, and the relative speed is zero if the denominator is zero. An uncertainty overlap is performed to calculate the intersection area S of the vehicle ellipse and the obstacle ellipse inter with the area of the vehicle ellipse S car , ； U42. Three-dimensional risk weight calculation: First, calculate the base risk weight Risk that fuses the time and spatial dimensions base , ， wherein, ɑ is the time dimension weight, β is the space dimension weight, TTC max is the collision time maximum, Overlap max is the uncertainty overlap maximum, next, the risk weight is adjusted combined with the main idea of the obstacle to modify the idea, ， Where γ is a correction parameter; U43. Risk level and decision output.

10. The motion uncertainty quantification based multi-modal intent collision detection method of claim 9, wherein: The risk level and decision output are high risk, vehicle emergency braking if Risk final is greater than or equal to 0.8; medium risk, path fine tuning or deceleration if Risk final is greater than or equal to 0.4 and less than 0.8; and continue monitoring if Risk final is less than 0.4.

Images