An automatic driving behavior regulation method and system based on perception reliability prior

Abstract

This invention discloses an autonomous driving behavior control method and system based on perception reliability priors, relating to the fields of autonomous driving and intelligent vehicle safety control technology. Addressing the soft anomaly problem of "appearing usable but actually distorted" perception results in autonomous driving systems, this method extracts at least three types of runtime features, including sensor health, multimodal consistency, and temporal stability, and jointly generates local and global perception reliability priors. Independent of the feature fusion process within the perception network, this system directly performs hierarchical behavior gating at the kinematic envelope level: restricting local driving actions for specific targets or regions based on local priors; and shrinking or performing degraded control on the candidate trajectory set based on global priors. This invention can suppress the risk of high-confidence perception false detections propagating to the planning and control link, achieving local action permission constraints and global system defense, thereby improving the safety and robustness of autonomous driving systems.

Description

Technical Field

[0001] The invention relates to the field of autonomous driving and intelligent vehicle safety control technology, and in particular to an autonomous driving behavior regulation method and system based on prior perception reliability. It is applicable to the reliability assessment of the perception results of the vehicle's surrounding environment, and to the gating, restriction, degradation or minimum risk control of the autonomous vehicle behavior based on the reliability assessment results. Background Technology

[0002] Existing autonomous driving systems typically rely on a variety of sensors, such as cameras, lidar, millimeter-wave radar, inertial measurement units, global navigation and positioning units, and vehicle chassis status acquisition units, to perceive the environment around the vehicle and perform target detection, lane recognition, drivable area estimation, target tracking, trajectory prediction, path planning, and vehicle control.

[0003] However, in actual road operation, autonomous driving perception systems often exhibit an abnormal state with high safety risks: the relevant sensors and perception algorithms continue to output results, and on the surface, the output appears continuous, stable, and usable, but its representation of the external environment has become distorted. This type of anomaly is not due to complete sensor failure, nor is it due to the system ceasing to output; rather, it is a soft anomaly situation of "appearing usable but actually distorted."

[0004] For example, under conditions such as backlight, nighttime glare, rain and fog, lens smudges, reflective material interference, sparse point clouds, multipath reflection, time synchronization deviation, slight drift of extrinsic parameters, sensor installation offset, and distributed external scenes, the autonomous driving system may still output target boxes, lane lines, depth estimation results, occupancy information, or drivable area information, but these results are no longer reliable in terms of semantic category, spatial location, boundary shape, motion trend, and temporal continuity.

[0005] Existing technologies typically focus on improving the recognition accuracy of perception models, increasing the confidence of single models, performing multi-sensor fusion, or triggering degraded control when explicit faults are detected. However, for soft anomalies where the perception results appear usable but are actually distorted, existing solutions often lack a hierarchical reliability characterization mechanism for the real-time decision chain of autonomous driving, and also lack a systematic solution to directly transform such reliability characterizations into local action permission restrictions and global candidate behavior or trajectory set shrinkage.

[0006] Furthermore, although some perception technologies based on "multimodal feature gating" or "uncertainty fusion" have emerged in recent years, most of these technologies are limited to the network weight allocation layer within the environmental perception module. Their core assumption is that as long as the front-end fusion is done well enough, the data output to the downstream decision-making layer will necessarily be secure.

[0007] However, in the actual engineering operation of high-level autonomous driving (NOA / FSD) systems, due to the inexhaustibility of long-tailed out-of-scene (OOD) distributions, perception models are highly susceptible to "high-confidence misclassification." When the neural network experiences feature extraction distortion due to factors such as lighting reflections or water mist, it may still output false targets (such as ghost obstacles) or incorrect geometric boundaries with extremely high confidence scores. When existing predictive planning modules (PNC) receive such "seemingly usable, actually distorted, and accompanied by high confidence" data, they directly treat it as a hard constraint for trajectory planning, thereby triggering abnormal control responses or unexpected braking behaviors (such as ghost braking) or frequently triggered unexpected degradation behaviors (such as frequent requests for manual intervention).

[0008] In particular, in existing technologies, the confidence level output by the perception model is usually only used as an auxiliary parameter of the recognition result, and it is generally not used as the core action gating basis for prediction, planning, and control modules. As a result, erroneous perception results may still be used by subsequent decision-making modules, leading to the continuous transmission and amplification of risks.

[0009] Therefore, it is necessary to propose a new technical solution to systematically assess the credibility of environmental perception results before they enter the autonomous driving decision chain, and to impose hierarchical constraints, downgrades, or blocks on autonomous driving behavior based on the assessment results, thereby improving the safety and robustness of the autonomous driving system under complex working conditions. Summary of the Invention

[0010] The primary objective of this invention is to provide a method and system for regulating autonomous driving behavior based on prior perception reliability. This addresses the objective technical problem in existing high-order autonomous driving systems where the Predictive and Planning Control (PNC) layer is easily misled by "high-confidence soft-distortion data" output by the perception module, leading to control oscillations and dangerous misoperations. This invention overcomes the limitations of existing technologies that over-rely on a single perception confidence level, proposing a system-level behavior regulation mechanism independent of the weight allocation within the perception network.

[0011] Another objective of this invention is to address the technical problem of excessively coarse granularity in existing degradation strategies. By establishing a hierarchical gating mechanism of "local action permission restriction" and "global candidate trajectory set contraction," this invention enables the vehicle to precisely block high-risk local actions such as specific lane changes and obstacle avoidance when perception credibility decreases, while simultaneously ensuring overall driving safety through the contraction of the physical kinematic envelope. This, in turn, minimizes frequent and abrupt system takeovers while ensuring functional safety, thereby improving the availability and ride comfort of intelligent vehicles in complex environments.

[0012] To achieve the above objectives, the present invention provides an autonomous driving behavior control method based on prior perception reliability, comprising:

[0013] Acquire multi-source perception data of the vehicle's surrounding environment and generate environmental perception results based on the multi-source perception data;

[0014] Extract at least three types of features corresponding to the multi-source sensing data and the environmental sensing results. The at least three types of features include: sensor health status features, multimodal consistency features, temporal stability features, anomaly features, and uncertainty features.

[0015] The perception reliability prior is generated jointly based on the at least three types of features. The perception reliability prior includes at least: a global perception reliability prior that characterizes the overall perception reliability of the environment, and local perception reliability priors that correspond to specific targets, specific areas, specific lane lines and / or specific occupants, respectively.

[0016] Based on the comparison between the local perception reliability prior and the corresponding local threshold, local driving actions related to the specific target, specific area, specific lane line and / or specific occupancy unit are restricted;

[0017] Based on the comparison between the global perception reliability prior and the corresponding global threshold, the candidate driving behavior and / or candidate trajectory set of the vehicle is narrowed, and degraded control or minimum risk maneuver is performed when the global perception reliability prior is lower than the preset safety threshold.

[0018] Perform autonomous driving prediction, planning, and control within the narrowed set of candidate driving behaviors and / or candidate trajectories.

[0019] In some implementations, the multi-source sensing data includes at least one of the following: camera image data, lidar point cloud data, millimeter-wave radar data, inertial measurement unit data, global navigation and positioning unit data, wheel speed data, and vehicle chassis status data.

[0020] In some implementations, before generating environmental perception results based on the multi-source sensing data, the method further includes performing time synchronization, spatial registration, noise reduction, distortion correction, and / or calibration correction on the multi-source sensing data.

[0021] In some implementations, the environmental perception results include at least one of the following: target detection results, target tracking results, semantic segmentation results, depth estimation results, lane recognition results, occupancy information, and drivable area recognition results.

[0022] In some implementations, the at least three types of features include at least: sensor health status features, multimodal consistency features, and timing stability features.

[0023] In some embodiments, the sensor health status characteristics include at least one of the following: exposure anomaly characteristics, image blurring characteristics, lens smudge occlusion characteristics, raindrop adhesion characteristics, point cloud sparsity characteristics, radar echo anomaly characteristics, frame rate jitter characteristics, time synchronization deviation characteristics, sensor extrinsic parameter drift characteristics, and installation offset characteristics.

[0024] In some implementations, the multimodal consistency features include at least one of the following: spatial alignment consistency between visual targets and point cloud targets, consistency between visual tracking speed and radar speed measurement results, consistency between the boundaries of multi-source drivable areas, consistency between lane recognition results and map priors, and consistency between target depth estimation and spatial reconstruction results.

[0025] In some implementations, the timing stability features include at least one of the following: continuous frame jump features of target position, abnormal change features of target velocity, discontinuity features of target boundary contour, sudden generation or sudden disappearance of target, lane line continuity disruption features, and non-physical jitter features of occupied units.

[0026] In some implementations, the perceived reliability prior also includes a structured perceived reliability prior, which includes at least one of the following: semantic reliability, geometric reliability, and temporal reliability.

[0027] In some implementations, the local driving action includes at least one of the following: performing a lane change action to the corresponding area, performing a detour action around the corresponding target, performing a lane keeping correction action based on the corresponding lane line, and performing a local obstacle avoidance path selection action based on the corresponding occupancy unit; when the prior reliability of the corresponding local perception is lower than the corresponding local threshold, at least one of the above local driving actions is restricted.

[0028] In some implementations, shrinking the candidate driving behavior and / or candidate trajectory set includes at least one of the following: limiting the maximum vehicle speed, increasing the following distance, expanding the collision prediction safety boundary, reducing the candidate trajectory set, and switching to conservative trajectory planning; and when the global perception reliability prior is lower than a lower level of safety threshold, performing a minimum risk maneuver or safe stopping.

[0029] This invention also provides an autonomous driving behavior control system based on perception reliability priors, comprising: a multi-source perception data acquisition module, an environment perception module, a feature extraction module, a perception reliability prior generation module, a behavior gating module, a prediction planning module, and a control execution module; wherein, the perception reliability prior generation module is used to jointly generate a global perception reliability prior and a local perception reliability prior based on at least three types of features extracted during current runtime, and the behavior gating module is used to restrict local driving actions according to the local perception reliability prior and to shrink the candidate driving behavior and / or candidate trajectory set according to the global perception reliability prior.

[0030] Beneficial effects:

[0031] Compared with the prior art, the present invention has at least the following beneficial effects:

[0032] First, this invention does not rely solely on the confidence level or a single anomaly index output by a single perception model. Instead, it comprehensively characterizes the credibility of current environmental perception by jointly utilizing at least three types of runtime evidence, including sensor health status, multimodal consistency, temporal stability, and anomaly and / or uncertainty features. Therefore, it can more effectively identify soft anomaly states in the autonomous driving perception chain that are "appearing usable but actually distorted".

[0033] Second, this invention does not only generate a single overall risk quantity, but simultaneously generates global perception reliability priors and local perception reliability priors, thereby reflecting the overall environmental perception credibility and the local credibility of specific targets, areas, lane lines or occupied units, providing a hierarchical basis for subsequent autonomous driving behavior control.

[0034] Third, this invention constructs an independent and physically-based two-layer behavior gating mechanism of "local action permission restriction + global trajectory envelope contraction". Compared with existing technologies that only adjust network weights within the algorithm, the gating mechanism of this invention directly acts on the kinematic envelope or geometric space generated by the prediction planning module. For example, by introducing a dynamic safety redundancy coefficient in the global contraction. Increase following distance and collision prediction time This invention directly transforms abstract perceived risk into concrete vehicle chassis kinematic constraints. This hierarchical mechanism avoids the nonlinear runaway problem caused by coarsely adjusting vehicle behavior with a single global risk parameter, achieving fine-grained restrictions on driving action permissions. Unlike existing technologies that only perform feature fusion or uncertainty weighting within the perception model, the behavior gating of this invention directly acts on the kinematic envelope or geometric space generated by the prediction planning module, achieving physical constraints on autonomous driving behavior at the system structure level.

[0035] Fourth, when perception confidence decreases, the present invention can specifically restrict lane changes to specific areas, detours around specific targets, lane keeping corrections based on specific lane lines, and local obstacle avoidance path selection. It can also simultaneously limit the maximum vehicle speed, expand the safety boundary, reduce the candidate trajectory set, switch to conservative trajectory planning, and perform degraded control or minimum risk maneuvers when necessary, thereby improving the safety and robustness of the autonomous driving system in complex weather, complex lighting, sensor soft distortion, and distributed scenarios.

[0036] Fifth, by establishing the generation of perceptual reliability priors on the joint basis of multiple types of evidence during the current operation of the autonomous vehicle, this invention enables the scheme to directly reflect the real-time credibility of the current vehicle, the current perception chain, and the current scenario, and is therefore more suitable as an online behavioral gating basis before autonomous driving planning and control. Attached Figure Description

[0037] Figure 1 This is a block diagram of the overall structure of the autonomous driving behavior control system based on perception reliability prior of the present invention.

[0038] Figure 2 This is a flowchart of the autonomous driving behavior control method based on perceived reliability prior of the present invention;

[0039] Figure 3 This is a schematic diagram of the input-output relationship of the perceived reliability prior generation module in this invention;

[0040] Figure 4 is a schematic diagram of the behavior gating logic based on perceived reliability prior in this invention.

[0041] Figure 5 This is a schematic diagram of the hierarchical behavior gating of the present invention for the "ghost braking" problem under heavy rain conditions. It shows the abnormal target recognition caused by soft perceptual distortion, and the process of regulating emergency braking requests and candidate trajectory sets based on local perception reliability priors and global perception reliability priors.

[0042] [Explanation of Labels in the Attached Image]

[0043] Among them, 10 represents the external traffic environment, 20 represents the vehicle, 100 represents the autonomous driving behavior control system, 110 represents the multi-source perception data acquisition module, 111 represents the camera unit, 112 represents the LiDAR unit, 113 represents the millimeter-wave radar unit, 114 represents the positioning and attitude acquisition unit, 120 represents the data preprocessing module, 130 represents the sensor health monitoring module, 140 represents the environmental perception module, 150 represents the perception reliability prior generation module, 151 represents the health status feature extraction unit, 152 represents the multimodal consistency analysis unit, and 153 represents the temporal stability analysis unit. 154 is the anomaly and uncertainty analysis unit; 155 is the PRP fusion output unit; 160 is the behavior gating module; 170 is the prediction planning module; 180 is the control execution module; 190 is the closed-loop feedback and update module; 200 is the "ghost obstacle" identification result; 210 is the emergency braking trigger request (AEB trigger request); 220 is the braking request after deweighting or delayed confirmation; 230 is the local driving action restriction output; 240 is the dynamic safety redundancy coefficient; 250 is the contracted candidate trajectory set; 260 is the expanded safety boundary or kinematic envelope. Detailed Implementation

[0044] To more clearly illustrate the present invention, the following description, in conjunction with preferred embodiments and accompanying drawings, further explains the invention. Similar components in the drawings are indicated by the same reference numerals. Those skilled in the art should understand that the specific description below is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0045] To facilitate understanding of some of the terms used in this application, the relevant terms are explained below. These explanations are intended to clarify the technical meanings within this application and are not intended to unnecessarily limit the scope of protection of this application.

[0046] The "Perception Reliability Prior" (PRP) described in this application refers to reliability information in an autonomous driving system that characterizes the credibility of current environmental perception based on multi-source perception data acquired during vehicle operation and / or environmental perception results generated based on the multi-source perception data. This information serves as a preliminary constraint, reference, or gating basis for the autonomous driving decision chain before subsequent prediction, planning, and control phases. The PRP can be represented by scores, levels, labels, vectors, or combinations thereof, and is not limited to a single probability value.

[0047] In this application, the term "prior" refers to pre-existing constraint information relative to the subsequent prediction, planning, and control stages. It means information provided to the decision chain as constraints or references before the execution of autonomous driving behavior, and is not limited to static prior knowledge given offline in advance, detached from current observation data, nor is it limited to statistical priors in a strictly Bayesian sense. In some implementations, the PRP in this application is jointly generated from multiple types of evidence during current runtime, and its generation does not require activation by matching historical abnormal event entries as a necessary prerequisite.

[0048] The "global perception reliability prior" mentioned in this application refers to the representation of the overall environmental perception reliability at the current moment; the "local perception reliability prior" refers to the representation of the reliability of a specific target, a specific region, a specific lane line, a specific occupant unit, and / or a specific local perception task. The term "local" is not limited to a local area of ​​an image, but may also correspond to a spatial region, a target instance, a functional region, or a local perception task.

[0049] The "structured perceived reliability prior" mentioned in this application refers to the representation form after decomposing the perceived credibility according to different reliability dimensions. The structured perceived reliability prior may include at least one or more of semantic reliability, geometric reliability, and time-series reliability. The term "structured" is not limited to a specific data structure, tensor form, or network output form, but refers to a multi-dimensional and distinguishable way of representing reliability.

[0050] The “semantic reliability” mentioned in this application refers to the degree of credibility of the environmental perception results in terms of target category, scene semantics or functional attributes; the “geometric reliability” refers to the degree of credibility of the environmental perception results in terms of geometric or motion quantities such as position, size, depth, speed, orientation, and boundary shape; and the “temporal reliability” refers to the consistency, continuity and stability of the environmental perception results over consecutive time periods.

[0051] The "kinematic envelope" mentioned in this application refers to the feasible range of motion defined by vehicle dynamics capabilities and safety constraints. This range is determined by the following parameters: vehicle speed, acceleration, braking capability, steering capability, following distance, collision prediction time, and candidate trajectory boundaries.

[0052] The "dynamic safety redundancy coefficient" mentioned in this application refers to a safety amplification coefficient that is dynamically adjusted based on the prior knowledge of global perception reliability. It is used to proportionally expand the following distance, collision prediction time, and trajectory boundary to enhance the safety margin of the system under unreliable perception conditions.

[0053] The "behavioral gating" described in this application refers to a control process that allows, restricts, downgrades, blocks, prioritizes, and / or reduces the action set of candidate driving behaviors of a vehicle based on the PRP (Predictive Planning Principle). This process directly affects the kinematic envelope or geometric space generated by the predictive planning module, rather than merely affecting the feature fusion or weight allocation process within the perception model. The behavioral gating is not limited to binary switch control; it can also be multi-level constraint control or continuous adjustment control.

[0054] The “partial driving action” described in this application refers to a driving action or partial behavior selection corresponding to a specific target, a specific area, a specific lane line and / or a specific occupancy unit; the partial driving action may include one or more of the following: performing a lane change action to the corresponding area, performing a detour action around the corresponding target, performing a lane keeping correction action based on the corresponding lane line, and performing a partial obstacle avoidance path selection action based on the corresponding occupancy unit.

[0055] The “candidate driving behavior” mentioned in this application refers to the driving actions, behavior strategies or trajectory schemes that an autonomous driving system can choose from given environmental perception results and vehicle operating status; the “candidate trajectory set” refers to a set of multiple selectable driving trajectories; and the “conservative trajectory planning” refers to a planning method that generates driving trajectories with lower speed, larger safety boundaries, higher constraint strength or fewer degrees of freedom of action compared to conventional autonomous driving planning.

[0056] The "degraded control" described in this application refers to a control method that reduces or conservatively adjusts vehicle control authority, behavioral freedom, operating speed, trajectory selection range, or action set when the perceived trust level decreases, system risk increases, or preset safety conditions are not met. The "minimum risk maneuver" refers to the control process that causes the vehicle to switch to a relatively low-risk state when the system determines that continuing the current autonomous driving behavior poses a high risk. This may include deceleration, lane keeping, pulling over, safe stopping, or other risk-reducing actions.

[0057] The term "appears usable but is actually distorted" as used in this application refers to a state where the relevant sensors and / or perception algorithms continue to output results, and the output appears continuous, stable, or available for subsequent modules to use. However, the output exhibits significant deviations in semantics, geometric position, boundary morphology, motion trends, temporal continuity, or accuracy of environmental representation, thus failing to support safe autonomous driving decisions. This state falls under the category of soft perceptual anomalies in a non-complete failure state.

[0058] The "sensor health status characteristics" mentioned in this application refer to the characteristic information used to characterize whether the current working state of the sensor is normal, stable, or reliable; the "multimodal consistency characteristics" refer to the consistency characterization of different sensors, different sensing channels, and / or different sensing results in terms of space, time, semantics, geometry, or motion; the "temporal stability characteristics" refer to the characteristic characterization of whether the environmental sensing results remain continuous, smooth, reasonable, and in accordance with physical laws at consecutive moments; the "abnormal characteristics" refer to the characteristic information used to characterize the deviation of the current input, current scene, or current sensing output from the normal state; and the "uncertainty characteristics" refer to the characteristic information used to characterize the insufficient reliability of the current sensing output or the increased dispersion of the results. Detailed Implementation

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. All equivalent substitutions, improvements, and modifications made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0060] Implementation Method 1: Method Implementation

[0061] like Figures 1 to 4 As shown, this embodiment provides an autonomous driving behavior control method based on prior perception reliability, which can be applied to autonomous vehicles, intelligent driving controllers, or unmanned driving platforms.

[0062] First, multi-source perception data of the vehicle's surrounding environment is acquired. This multi-source perception data may come from one or more of the following: cameras, lidar, millimeter-wave radar, inertial measurement units, global navigation and positioning units, wheel speed sensors, and vehicle chassis status acquisition units.

[0063] In some implementations, the multi-source sensing data is preprocessed, including at least one of time synchronization, spatial registration, noise reduction, distortion correction, and calibration correction. This preprocessing reduces the impact of sensor time and spatial biases and signal noise on subsequent environmental sensing and sensing reliability assessment.

[0064] Subsequently, environmental perception results are generated based on the preprocessed multi-source perception data. The environmental perception results may include one or more of the following: target detection results, target tracking results, semantic segmentation results, depth estimation results, lane recognition results, occupancy information, and drivable area recognition results.

[0065] After obtaining the environmental perception results, at least three types of runtime features corresponding to the multi-source perception data and the environmental perception results are extracted to generate a priori perception reliability. The at least three types of runtime features include: sensor health status features, multimodal consistency features, temporal stability features, anomaly features, and uncertainty features.

[0066] The sensor health status features are used to characterize whether the working status of each sensor in the current vehicle perception chain is normal, stable, or reliable. For example, for cameras, features such as exposure anomaly, image blur, lens smudge occlusion, and raindrop adhesion can be extracted; for lidar, features such as point cloud sparsity and echo anomaly can be extracted; for millimeter-wave radar, features such as echo anomaly can be extracted; in addition, features such as frame rate jitter, time synchronization deviation, sensor extrinsic parameter drift, and installation offset can also be extracted.

[0067] The multimodal consistency feature is used to characterize whether there is consistency between different sensors, different perception channels, and different perception results. For example, it can analyze the spatial alignment consistency between visual targets and point cloud targets, the consistency between visual tracking speed and radar speed measurement results, the consistency of multi-source drivable area boundaries, the consistency between lane recognition results and map priors, and the consistency between target depth estimation and spatial reconstruction results.

[0068] The temporal stability feature is used to characterize the stability of environmental perception results over continuous time. For example, it can detect whether the target position undergoes abrupt changes that do not conform to physical laws in consecutive frames, whether the target speed and boundary contours experience abnormal jitter, whether the target appears or disappears abnormally, whether the continuity of lane lines is disrupted, and whether the occupied unit experiences non-physical jitter.

[0069] The aforementioned anomaly features and uncertainty features are used to characterize whether the current scene or the current perception output deviates from the normal state or is in a low-confidence state. For example, corresponding features can be constructed using anomaly detection results, out-of-distribution detection results, or perception uncertainty results.

[0070] Based on the above three types of runtime features, a perception reliability prior is generated. The perception reliability prior includes at least: a global perception reliability prior characterizing the overall environmental perception reliability at the current moment, and local perception reliability priors corresponding to specific targets, specific regions, specific lane lines, and / or specific occupied units, respectively.

[0071] In some implementations, the perception reliability prior also includes a structured perception reliability prior. The structured perception reliability prior may include at least one or more of semantic reliability, geometric reliability, and temporal reliability, used to further describe the state of the current environment perception result across different reliability dimensions.

[0072] After obtaining the perceived reliability prior, it is input into the behavior gating module. The behavior gating module does not adjust vehicle behavior based solely on a single overall risk indicator, but rather performs hierarchical behavior gating based on both local and global perceived reliability priors.

[0073] Specifically, when the prior reliability of local perception is lower than the corresponding local threshold, local driving actions related to the specific target, specific area, specific lane line, and / or specific occupancy unit are restricted. The local driving actions may include at least one of the following: performing a lane change action to the corresponding area, performing a detour action around the corresponding target, performing a lane keeping correction action based on the corresponding lane line, and performing a local obstacle avoidance path selection action based on the corresponding occupancy unit.

[0074] Simultaneously, when the global perception reliability prior is lower than the corresponding global threshold, the candidate driving behavior and / or candidate trajectory set of the vehicle is reduced. The reduction may include at least one of the following: limiting the maximum vehicle speed, increasing the following distance, expanding the collision prediction safety boundary, reducing the candidate trajectory set, and switching to conservative trajectory planning; when the global perception reliability prior is further lower than the lower level safety threshold, degraded control, minimum risk maneuver, or safe stopping is performed.

[0075] Under the dual constraints of the aforementioned local driving action restrictions and the shrinking of the global candidate behavior and / or candidate trajectory set, the prediction and planning module performs autonomous driving prediction and planning within the shrinking action permission range, and the control execution module controls the vehicle to perform the corresponding driving actions according to the planning results.

[0076] For example, when the local perception reliability prior in the adjacent lane area to the left of the vehicle decreases, lane-changing actions to the left can be restricted; when the global perception reliability prior decreases at the same time, the maximum vehicle speed can be further restricted, the following distance can be increased, and the candidate trajectory set can be reduced; when the global perception reliability prior is further lower than the minimum safety threshold, minimum risk maneuvering or safe stopping can be triggered.

[0077] In some implementations, the system may also record low-perception reliability events, perception conflict events, and abnormal scenario data for subsequent parameter optimization, threshold tuning, or model updates; however, the recording and updating process is not the primary triggering basis for perception reliability prior generation and behavior gating in this implementation.

[0078] Implementation Method Two: System Implementation Method

[0079] like Figure 1As shown, this embodiment provides an autonomous driving behavior control system 100 based on perception reliability priors, including a multi-source perception data acquisition module 110, a data preprocessing module 120, a sensor health monitoring module 130, an environmental perception module 140, a perception reliability prior generation module 150, a behavior gating module 160, a prediction planning module 170, a control execution module 180, and a closed-loop feedback and update module 190.

[0080] The multi-source perception data acquisition module 110 is used to acquire multi-source perception data of the environment surrounding the vehicle 20. The multi-source perception data acquisition module 110 may include a camera unit 111, a lidar unit 112, a millimeter-wave radar unit 113, and a positioning and attitude acquisition unit 114.

[0081] The data preprocessing module 120 is used to perform at least one of the following on the multi-source sensing data: time synchronization, spatial registration, noise reduction, distortion correction, and calibration correction.

[0082] The sensor health monitoring module 130 is used to extract and output sensor health status characteristics.

[0083] The environment perception module 140 is used to generate environment perception results based on preprocessed multi-source perception data.

[0084] The perception reliability prior generation module 150 is used to extract at least three types of runtime features based on the multi-source perception data and the environmental perception results, and jointly generate global perception reliability prior and local perception reliability prior based on the at least three types of runtime features.

[0085] In some implementations, such as Figure 3 As shown, the perceived reliability prior generation module 150 may include a health status feature extraction unit 151, a multimodal consistency analysis unit 152, a time series stability analysis unit 153, an anomaly and uncertainty analysis unit 154, and a PRP fusion output unit 155.

[0086] The behavior gating module 160 is used to perform hierarchical behavior gating based on the global perception reliability prior and the local perception reliability prior, respectively. Specifically, the behavior gating module 160 restricts corresponding local driving actions based on the local perception reliability prior, and restricts candidate driving behaviors and / or candidate trajectory sets based on the global perception reliability prior.

[0087] The prediction and planning module 170 is used to perform autonomous driving prediction and planning under the constraints of the behavior gating module 160.

[0088] The control execution module 180 is used to control the vehicle to perform corresponding driving actions based on the planning results.

[0089] The closed-loop feedback and update module 190 is used to record low-perception reliability events, perception conflict events, and abnormal scenario data, and to update the relevant parameters, thresholds, or strategies of the perception reliability prior generation module 150 and the behavior gating module 160.

[0090] Implementation Method 3: Typical Application Scenarios

[0091] This embodiment uses nighttime urban road conditions as an example to illustrate the application process of the present invention. When a vehicle is driving on a road at night, there are backlighting headlights ahead, a small amount of dirt adheres to the surface of the camera lens, and the road guardrail has strong reflective properties.

[0092] Under this condition, the camera continues to output images, the target detection module can still output the vehicle outline and lane line results, the lidar continues to output point cloud data, and the millimeter-wave radar continues to output distance and relative speed information, meaning that the system is still in a "usable" state on the surface.

[0093] However, due to backlighting, dirt, and reflection interference, the local overexposed areas in the image increase, the target boundary fluctuations increase, the lane line continuity decreases, the spatial overlap between the visual target and the point cloud target decreases, the consistency between the visual tracking speed and the radar speed measurement results decreases, and the target outline in consecutive frames shows non-physical jumps.

[0094] The system extracts exposure anomaly features and blur features through the sensor health monitoring module 130, extracts visual and point cloud consistency decline features through the multimodal consistency analysis unit 152, extracts boundary jump and timing jitter features through the temporal stability analysis unit 153, and extracts anomaly features or uncertainty features through the anomaly and uncertainty analysis unit 154.

[0095] Subsequently, the PRP fusion output unit 155 generates global perception reliability priors and local perception reliability priors based on the above at least three types of runtime features, wherein the geometric reliability of the forward target decreases, the local perception reliability prior of the left front region decreases, and the lane line temporal reliability decreases.

[0096] The behavior gating module 160 restricts lane-changing actions to the left front area based on the aforementioned local perception reliability prior results, and restricts the vehicle's maximum speed, increases the following distance, and reduces the candidate trajectory set based on the aforementioned global perception reliability prior results. At the same time, the prediction planning module 170 is switched to conservative trajectory planning mode. When the global perception reliability prior further decreases to below the minimum safety threshold, the control execution module 180 performs minimum risk maneuvers or safe stops.

[0097] Therefore, even if the relevant sensors and perception algorithms in the system continue to output results, the present invention can still identify risky states in the autonomous driving perception chain that appear usable but are actually distorted, and prevent the execution of high-risk driving behaviors through hierarchical behavior gating that restricts local driving actions and shrinks global candidate behaviors and / or candidate trajectory sets.

[0098] Example 4: "Ghost Braking" Defense and Layered Gating Application under Heavy Rain Conditions

[0099] This embodiment uses a typical "phantom braking" defense scenario under heavy rain conditions on a highway as an example to further illustrate the system-level control process of the present invention under the condition of perceived soft distortion, such as... Figure 5 As shown.

[0100] Scenario Description: A vehicle is cruising at 100 km / h on a highway when it encounters a sudden downpour. A large truck is driving ahead, kicking up a large amount of water mist. At this moment, the reflection of the water on the road surface and the water mist are incorrectly identified as "stationary white box-shaped obstacles" by the visual neural network of the forward-looking camera in some frames of the image. Furthermore, due to the limitations of neural network feature extraction, the confidence level of its output classification of this target is as high as 95%.

[0101] The dilemma of existing technology: In traditional autonomous driving systems, because a single vision model gives a very high target confidence level, and the "ghost obstacle" happens to be located directly in front of the vehicle's driving trajectory, the planning and control layer will immediately respond and trigger automatic emergency braking (AEB), causing the vehicle to experience dangerous "ghost braking" at high speeds, which can easily lead to rear-end collisions.

[0102] The processing procedure of this invention is as follows:

[0103] Step A: Feature Extraction and Soft Anomaly Detection. Although the visual perception module reported a high-confidence target, the sensor health monitoring module 130 extracted "lens raindrop attachment features" and "local exposure anomaly features"; the multimodal consistency analysis unit 152 found extreme conflict, that is, the "spatial alignment consistency between the visual target and the point cloud target" is close to zero (the millimeter-wave radar did not return a solid echo, and although the lidar point cloud was sparse, it did not present any physical obstacles); at the same time, the temporal stability analysis unit 153 detected that the "obstacle" was "suddenly generated and the boundary was not physically abrupt" (without a reasonable temporal motion trajectory).

[0104] Step B: PRP Joint Generation. Based on the above at least three types of runtime characteristics, the perception reliability prior generation module 150 performs online diagnosis: the local perception reliability prior score for the "obstacle" in front drops sharply (far below the local safety threshold); at the same time, affected by the rainstorm environment, the global perception reliability prior also drops to the first level of safety range.

[0105] Step C: Layered behavior gating execution. (1) Local restriction (defense against ghost braking): The behavior gating module 160 determines that the local perception reliability of the specific target ahead is extremely low, so it directly restricts the local driving actions related to the target, that is, it downweights the braking request and enters the delayed confirmation or conservative deceleration verification process. The vehicle does not brake suddenly. (2) Global contraction (improving system robustness): At the same time, in response to the decrease in the prior global perception reliability, the behavior gating module 160 contracts the candidate trajectory set of the prediction planning module 170. The system introduces a dynamic safety redundancy coefficient. (In this embodiment, we take) ), will calibrate the following distance and collision prediction time The speed limit has been increased by 20%, and the global maximum speed limit has been reduced from 100 km / h to 80 km / h.

[0106] Implementation Results: Ultimately, the vehicle smoothly reduced its cruising speed and maintained a safe distance, thereby suppressing the risk of visual misdetections propagating to the planning and control link and triggering abnormal braking. This invention, independent of the internal feature weights of the perception model, implements gating directly at the physical kinematic envelope level, effectively solving the technical problem of autonomous driving systems being misled by soft anomalies that are "superficially high in confidence but actually distorted," leading to system oscillations.

Claims

1. A method for regulating autonomous driving behavior based on prior perception reliability, characterized in that, include: Acquire multi-source perception data of the vehicle's surrounding environment and generate environmental perception results based on the multi-source perception data; Extract at least three types of features corresponding to the multi-source sensing data and the environmental sensing results. The at least three types of features include: sensor health status features, multimodal consistency features, temporal stability features, anomaly features, and uncertainty features. Based on the joint generation of the at least three types of features, a perceptual reliability prior is generated, which includes at least the following: (1) Global perception reliability priors characterizing the overall environmental perception credibility; (2) Local perception reliability priors corresponding to specific targets, specific areas, specific lane lines and / or specific occupancy units respectively; Based on the comparison between the local perception reliability prior and the corresponding local threshold, local driving actions related to the specific target, specific area, specific lane line and / or specific occupancy unit are restricted; Based on the comparison between the global perception reliability prior and the corresponding global threshold, the candidate driving behavior and / or candidate trajectory set of the vehicle is shrunk, and degraded control or minimum risk maneuver is performed when the global perception reliability prior is lower than a preset safety threshold; wherein, the restriction on local driving actions and the shrinkage of the candidate driving behavior and / or candidate trajectory set of the vehicle are independent of the feature fusion layer or weight allocation network inside the environment perception module, and directly act on the kinematic envelope or geometric space generated by the prediction planning module; autonomous driving prediction, planning and control are performed within the shrinkage candidate driving behavior and / or candidate trajectory set.

2. The method according to claim 1, characterized in that, The multi-source sensing data includes at least one of the following: camera image data, lidar point cloud data, millimeter-wave radar data, inertial measurement unit data, global navigation and positioning unit data, wheel speed data, and vehicle chassis status data.

3. The method according to claim 1 or 2, characterized in that, Before generating environmental perception results based on the multi-source sensing data, the method further includes: performing time synchronization, spatial registration, noise reduction, distortion correction, and / or calibration correction on the multi-source sensing data.

4. The method according to any one of claims 1 to 3, characterized in that, The environmental perception results include at least one of the following: target detection results, target tracking results, semantic segmentation results, depth estimation results, lane recognition results, occupancy information, and drivable area recognition results.

5. The method according to any one of claims 1 to 4, characterized in that, The at least three types of features include at least: sensor health status features, multimodal consistency features, and time series stability features.

6. The method according to claim 5, characterized in that, The sensor health status characteristics include at least one of the following: exposure anomaly characteristics, image blurring characteristics, lens smudge occlusion characteristics, raindrop adhesion characteristics, point cloud sparsity characteristics, radar echo anomaly characteristics, frame rate jitter characteristics, time synchronization deviation characteristics, sensor extrinsic parameter drift characteristics, and installation offset characteristics.

7. The method according to claim 5, characterized in that, The multimodal consistency features include at least one of the following: spatial alignment consistency between visual targets and point cloud targets, consistency between visual tracking speed and radar speed measurement results, consistency between the boundaries of multi-source drivable areas, consistency between lane recognition results and map priors, and consistency between target depth estimation and spatial reconstruction results.

8. The method according to claim 5, characterized in that, The temporal stability features include at least one of the following: continuous frame jump features of target position, abnormal change features of target velocity, discontinuity features of target boundary contour, sudden generation or sudden disappearance of target, lane line continuity disruption features, and non-physical jitter features of occupied units.

9. The method according to any one of claims 1 to 8, characterized in that, The perceived reliability prior also includes a structured perceived reliability prior, which includes at least one of the following: semantic reliability, geometric reliability, and temporal reliability.

10. The method according to any one of claims 1 to 9, characterized in that, The local driving action includes at least one of the following: performing a lane change action to the corresponding area, performing a detour action around the corresponding target, performing a lane keeping correction action based on the corresponding lane line, and performing a local obstacle avoidance path selection action based on the corresponding occupancy unit; when the prior reliability of the corresponding local perception is lower than the corresponding local threshold, at least one of the above local driving actions is restricted.

11. The method according to any one of claims 1 to 10, characterized in that, Based on the comparison between the global perception reliability prior and the corresponding global threshold, the vehicle candidate driving behavior and / or candidate trajectory set is shrunk by at least one of the following: limiting the maximum vehicle speed, increasing the following distance, expanding the collision prediction safety boundary, reducing the candidate trajectory set, and switching to conservative trajectory planning; and when the global perception reliability prior is lower than the lower level safety threshold, the minimum risk maneuver or safe stop is performed.

12. The method according to claim 11, characterized in that, The shrinking of the candidate driving behavior and / or candidate trajectory set includes introducing a dynamic safety redundancy coefficient. When the global perception reliability prior drops to the first level of safety range, the system multiplies the preset calibrated following distance and collision prediction time by the dynamic safety redundancy coefficient to amplify the envelope, thereby forcibly widening the vehicle's safety boundary; wherein, the dynamic safety redundancy coefficient... satisfy ,and The degree of decrease in the global perceived reliability prior monotonically increases; in some embodiments, The range of values ​​for is: ,in This is a preset upper limit value based on vehicle dynamics performance and safety constraints.

13. An autonomous driving behavior control system based on prior perception reliability, characterized in that, include: The multi-source perception data acquisition module is used to acquire multi-source perception data of the vehicle's surrounding environment. The environment perception module is used to generate environment perception results based on the multi-source perception data; The feature extraction module is used to extract at least three types of features corresponding to the multi-source perception data and the environmental perception results. The at least three types of features include at least three types of sensor health status features, multimodal consistency features, temporal stability features, anomaly features, and uncertainty features. The perception reliability prior generation module is used to jointly generate a perception reliability prior based on the at least three types of features. The perception reliability prior includes at least a global perception reliability prior and a local perception reliability prior. The behavior gating module is used to: restrict local driving actions related to the corresponding target, area, lane line, and / or occupied unit according to the comparison result of the local perception reliability prior and the corresponding local threshold; and shrink the vehicle candidate driving behavior and / or candidate trajectory set according to the comparison result of the global perception reliability prior and the corresponding global threshold, and trigger degraded control or minimum risk maneuver when the global perception reliability prior is lower than a preset safety threshold. The prediction and planning module is used to perform autonomous driving prediction and planning within the narrowed set of candidate driving behaviors and / or candidate trajectories. The control execution module is used to control the vehicle to perform corresponding driving actions based on the planning results.

14. The system according to claim 13, characterized in that, The multi-source sensing data acquisition module includes at least one of the following: a camera unit, a lidar unit, a millimeter-wave radar unit, and a positioning and attitude acquisition unit; the system also includes a data preprocessing module for performing time synchronization, spatial registration, noise reduction, distortion correction, and / or calibration correction on the multi-source sensing data.

15. The system according to claim 13 or 14, characterized in that, The perceptual reliability prior generation module is configured to generate a structured perceptual reliability prior based on the generation of global perceptual reliability prior and local perceptual reliability prior; wherein the structured perceptual reliability prior includes at least one of the following: semantic reliability, geometric reliability, and temporal reliability.

16. The system according to any one of claims 13 to 15, characterized in that, The behavior gating module is configured to: restrict at least one of the following local driving actions based on the comparison result between the local perception reliability prior and the corresponding local threshold: performing a lane change action to the corresponding area, performing a detour action around the corresponding target, performing a lane keeping correction action based on the corresponding lane line, and performing a local obstacle avoidance path selection action based on the corresponding occupancy unit; and, based on the comparison result between the global perception reliability prior and the corresponding global threshold, restrict at least one of the following: maximum vehicle speed, following distance, collision prediction safety boundary and candidate trajectory set, and trigger degraded control, minimum risk maneuver or safe stopping when the global perception reliability prior is lower than the lower level safety threshold.

17. A vehicle, characterized in that, Includes the autonomous driving behavior control system based on perceived reliability prior as described in any one of claims 13 to 16.

18. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 12.

Images