Intelligent driving perception safety assessment method and device and computer program product

Abstract

The invention relates to the technical field of intelligent driving safety, and provides an intelligent driving perception safety assessment method and device and a computer program product. Comprising the steps that perception quality evaluation indexes are calculated based on a target detection result of a perception module, and the perception quality evaluation indexes comprise an alarm coverage rate, a false alarm rate, a classification quality score and an uncertainty quality score; according to a pre-constructed threshold function and the perception quality evaluation index, dynamically adjusting a safety threshold of collision risk evaluation; based on a risk weight distribution matrix, performing fusion calculation on the perception quality evaluation index and the collision time to obtain a fusion risk value; and outputting a safety assessment result according to a comparison result of the fusion risk value and the safety threshold. The bottleneck of information separation of a sensing layer and a decision-making layer in a traditional method is broken through, and the safety evaluation capability of an intelligent driving system in a long-tail scene is improved.

Description

Intelligent driving perception safety assessment methods, devices and computer program products Technical Field

[0001] This application relates to the field of intelligent driving safety technology, specifically to an intelligent driving perception safety assessment method, device, and computer program product. Background Technology

[0002] With the rapid development of autonomous driving technology, the safety assessment of intelligent driving systems has become a key issue restricting their large-scale commercialization. Currently, the safety assessment of intelligent driving systems mainly relies on traditional vehicle dynamics indicators, such as time-to-collision (TTC), vehicle speed, acceleration, and other macroscopic behavioral parameters. These indicators can reflect the overall motion state of the vehicle and potential collision risks to a certain extent and have been widely used in traditional automotive active safety systems. However, these indicators only quantify risks from the perspective of vehicle kinematics and fail to cover the functional performance of the perception module itself. When the perception module experiences performance fluctuations or functional deficiencies, traditional indicators cannot reflect this change in a timely manner.

[0003] ISO 21448:2022 clearly proposes the conceptual framework of Safety of the Intended Functionality (SOTIF), aiming to address safety risks caused by insufficient system functionality or performance limitations. The standard states that autonomous driving systems need to focus on the functional performance of perception modules in various scenarios, especially robustness in long-tail scenarios such as rain, snow, strong light, and occlusion. However, existing standards only remain at the conceptual guiding principle level, lacking directly applicable quantitative assessment methods in engineering practice. In real-world applications, when vehicles enter complex scenarios, the performance of perception modules may significantly degrade, but existing safety assessment systems often fail to detect this change, still employing fixed risk judgment criteria, leading to a mismatch between safety margins and actual perception capabilities.

[0004] Therefore, the existing technology has the following technical defects: traditional safety assessment indicators are independent of the performance of the perception module, and changes in the function of the perception layer cannot be reflected in the safety assessment results; when the perception performance declines in long-tail scenarios, the system still operates in the conventional mode, which exposes the safety risks; at the same time, the system is too conservative when the perception performance is good, which affects traffic efficiency and driving experience. Summary of the Invention

[0005] In view of this, embodiments of this application provide an intelligent driving perception safety assessment method, device, and computer program product to solve the technical problems in the prior art where the perception module lacks a dedicated SOTIF evaluation index and perception quality and collision risk cannot be dynamically integrated.

[0006] A first aspect of this application provides an intelligent driving perception safety assessment method, comprising: calculating perception quality evaluation indicators based on target detection results of a perception module, wherein the perception quality evaluation indicators include alarm coverage rate, false alarm rate, classification quality score, and uncertainty quality score; dynamically adjusting a safety threshold for collision risk assessment based on a pre-constructed threshold function and the perception quality evaluation indicators; fusing the perception quality evaluation indicators with collision time based on a risk weight allocation matrix to obtain a fused risk value; and outputting a safety assessment result based on a comparison between the fused risk value and the safety threshold.

[0007] In one embodiment, the alarm coverage is calculated according to the following formula: Where N is the total number of perceived targets. For the i-th perceived target, For the set of key objectives, Let be the detection confidence score for the i-th perceived target. Let be the confidence threshold, and let be the indicator function. The uncertainty weight is the weight for the i-th perceived target; the uncertainty weight is calculated according to the following formula: ;in, The positional uncertainty of the i-th perceived target. This is the position uncertainty attenuation coefficient.

[0008] In one embodiment, the false alarm rate is calculated according to the following formula: Where M is the total number of detected targets, o j For the j-th detected target, 𝒪 critical For the set of key objectives, τ j Let τ be the tracking duration of the j-th detected target. track is the minimum tracking time threshold, and 𝕀() is the indicator function.

[0009] In one embodiment, the classification quality score is calculated according to the following formula: Where N is the total number of perceived targets. Let be the category probability distribution vector of the i-th perceived target. The information entropy of the probability distribution vector of this category. Let be the intersection-union ratio (IoU) of the detection bounding box and the ground truth bounding box for the i-th perceived target. Let i be the predicted category of the i-th perceived target. Let be the truth class of the i-th perceived target, and let 𝕀() be the indicator function.

[0010] In one embodiment, the uncertainty quality score is calculated according to the following formula: Where B is the total number of intervals, Let b be the set of sample indices for the b-th interval. Let be the number of samples in the b-th interval. Let i be the predicted category of the i-th perceived target. Let be the truth class of the i-th perceived target. Let be the average uncertainty of the b-th interval, and let be the indicator function.

[0011] In one embodiment, the threshold function is expressed as: ;in, Let be the dynamic safety threshold at time t. α is the baseline threshold, β is the alarm coverage adjustment coefficient, β is the false alarm rate adjustment coefficient, ACR(t) is the alarm coverage at time t, and FAR(t) is the false alarm rate at time t; , The dynamic security threshold decreases when alarm coverage increases and false alarm rate decreases.

[0012] In one embodiment, the risk weight allocation matrix is ​​a 4×3 matrix used to map alarm coverage, false alarm rate, classification quality score, and uncertainty quality score to three levels: low risk, medium risk, and high risk.

[0013] In one embodiment, the fusion risk value is calculated according to the following formula: ;in, To integrate risk values, TTC stands for Collision Time. For the normalized perceived quality evaluation index vector and W is the risk weight allocation matrix, and r is the risk level coefficient vector. , The risk level coefficient is low. The risk level coefficient is medium. It represents a high-risk level coefficient.

[0014] A second aspect of this application provides an intelligent driving perception safety assessment device, comprising: an index calculation module, configured to calculate perception quality evaluation indicators based on target detection results of a perception module, the perception quality evaluation indicators including alarm coverage rate, false alarm rate, classification quality score, and uncertainty quality score; a threshold adjustment module, configured to dynamically adjust a safety threshold for collision risk assessment based on a threshold threshold function and the perception quality evaluation indicators; a fusion calculation module, configured to fuse the perception quality evaluation indicators with the collision time based on a risk weight allocation matrix to obtain a fused risk value; and an assessment output module, configured to output a safety assessment conclusion based on a comparison between the fused risk value and the safety threshold.

[0015] A third aspect of this application provides a vehicle that includes the intelligent driving perception safety assessment device provided in the second aspect of this application.

[0016] A fourth aspect of this application provides a computer program product including a computer program that, when run, causes the method described in the first aspect of this application to be performed.

[0017] The intelligent driving perception safety assessment method provided in the first aspect of this application realizes a quantitative assessment of the functional adequacy of the perception module by constructing a perception quality evaluation index system that includes alarm coverage, false alarm rate, classification quality score and uncertainty quality score; it dynamically correlates perception quality with safety threshold through a threshold threshold function, enabling the system to adaptively adjust risk tolerance according to perception credibility; and it achieves multi-dimensional fusion of perception quality index and collision time through a risk weight allocation matrix, breaking through the bottleneck of information separation between the perception layer and the decision layer in traditional methods, and improving the safety assessment capability of the intelligent driving system in long-tail scenarios.

[0018] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 is a flowchart illustrating an intelligent driving perception safety assessment method according to an embodiment of this application; Figure 2 is a flowchart illustrating an intelligent driving perception safety assessment method according to another embodiment of this application; Figure 3 is a structural schematic diagram illustrating an intelligent driving perception safety assessment device according to an embodiment of this application. Detailed Implementation

[0021] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0022] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0023] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0025] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] As shown in Figure 1, the intelligent driving perception safety assessment method provided in this application includes the following steps S101 to S104: Step S101: Calculate the perception quality evaluation index based on the target detection results of the perception module. The perception quality evaluation index includes alarm coverage rate, false alarm rate, classification quality score and uncertainty quality score.

[0028] Step S102: Dynamically adjust the safety threshold for collision risk assessment based on the threshold threshold function and the perceived quality evaluation index.

[0029] Step S103: Based on the risk weight allocation matrix, the perceived quality evaluation index and the collision time are fused and calculated to obtain the fused risk value.

[0030] Step S104: Based on the comparison result between the fusion risk value and the security threshold, output the security assessment conclusion.

[0031] In applications, calculating perception quality evaluation indicators based on target detection results from the perception module involves acquiring raw data such as the location coordinates, category labels, confidence scores, and uncertainty estimates of each detected target from the perception module, and then calculating quantitative indicators across four dimensions—alarm coverage, false alarm rate, classification quality score, and uncertainty quality score—according to a pre-defined mathematical model. The perception module typically includes a fusion perception system comprised of multiple sensors such as LiDAR, cameras, and millimeter-wave radar. Its output target detection results include the target's three-dimensional position, velocity vector, target category, and corresponding confidence score. Alarm coverage measures the completeness of the perception system's detection of critical security threats; the false alarm rate quantifies the degree to which the perception system misclassifies non-critical targets as effective threats; the classification quality score assesses the accuracy and certainty of the perception system in target category identification; and the uncertainty quality score measures whether the perception system's estimate of its own predictive uncertainty matches its actual performance.

[0032] In application, dynamically adjusting the safety threshold for collision risk assessment based on the threshold function and perception quality evaluation index means using the calculated perception quality evaluation index as the input variable of the threshold function to calculate the appropriate safety threshold at the current moment in real time. Traditional methods use fixed collision time thresholds for risk assessment, which cannot adapt to the uncertainties caused by fluctuations in perception performance. This method establishes a functional mapping relationship between perception quality and safety threshold, enabling the safety boundary to dynamically shrink or expand according to perception reliability. When perception quality is high, the system can adopt a relatively aggressive safety threshold to avoid overly conservative driving; when perception quality deteriorates, the system automatically tightens the safety boundary to provide a greater safety margin.

[0033] In application, the fusion calculation of perceived quality evaluation indicators and collision time based on a risk weight allocation matrix refers to the weighted combination of four perceived quality indicators and traditional collision time indicators using a pre-trained weight matrix to generate a fused risk value that comprehensively considers perceived credibility and collision risk. The risk weight allocation matrix captures the contribution of each perceived quality indicator to different risk levels, enabling the fusion result to reflect the nonlinear interaction between perceived quality and collision risk.

[0034] In application, the safety assessment conclusion is output based on the comparison between the fusion risk value and the safety threshold. This means comparing the calculated fusion risk value with the dynamically adjusted safety threshold. When the fusion risk value exceeds the safety threshold, it is determined to be a high-risk state and corresponding safety intervention measures are triggered; otherwise, it is determined to be a safe state. The safety assessment conclusion can be output to the downstream decision-making and planning module to adjust the vehicle's trajectory or trigger active safety functions such as emergency braking.

[0035] Specifically, as shown in Figure 2, the complete workflow of this method is illustrated using pedestrian detection in a nighttime backlight scenario as an example. The perception module receives LiDAR point cloud data, camera images, and millimeter-wave radar target data, and outputs the target detection results after fusion processing. Due to the image quality degradation caused by backlight, the confidence level of pedestrian detection drops from 0.95 in normal scenarios to 0.72, and the position uncertainty increases from 0.1 meters to 0.6 meters. Based on the target detection results, the system calculates an alarm coverage rate of 0.85, a false alarm rate of 0.12, a classification quality score of 0.78, and an uncertainty quality score of 0.15. The threshold function calculates a dynamic safety threshold of 1.8 seconds based on the above indicators, which is tighter than the fixed threshold of 2.5 seconds. The risk weight allocation matrix fuses the perception quality index vector with the collision time of 2.1 seconds to obtain a fused risk value. Since the fused risk value takes into account the impact of the perception quality degradation, the system determines that it is currently in a medium-to-high risk state and triggers a warning 0.7 seconds in advance, giving the driver or autonomous driving system more reaction time.

[0036] The calculation of perceived quality evaluation indicators can employ an end-to-end deep learning-based method to directly map raw sensor data into indicator values, or a rule-based step-by-step calculation method to calculate each sub-indicator sequentially and then summarize them. Threshold functions can utilize linear weighting functions for simple and efficient threshold adjustment, or nonlinear neural network models can be used to capture more complex indicator-threshold mapping relationships.

[0037] This embodiment constructs a perception quality evaluation index system and dynamically integrates it with collision risk indicators, achieving information flow between the perception layer and the decision-making layer. This allows safety assessments to fully consider the functional adequacy of the perception modules. Compared to traditional fixed threshold methods, this method can adaptively adjust safety boundaries based on actual perception performance, avoiding overly conservative driving while ensuring safety. This improves the adaptability of the intelligent driving system in long-tail scenarios and enhances the user experience.

[0038] In one embodiment, the alarm coverage is calculated according to the following formula: Where N is the total number of perceived targets. For the i-th perceived target, For the set of key objectives, Let be the detection confidence score for the i-th perceived target. Let be the confidence threshold, and let be the indicator function. The uncertainty weight is the weight for the i-th perceived target; the uncertainty weight is calculated according to the following formula: ;in, The positional uncertainty of the i-th perceived target. This is the position uncertainty attenuation coefficient.

[0039] In applications, Alarm Coverage Rate (ACR) is a normalized metric between 0 and 1, used to quantify the effectiveness of a sensing system in detecting critical safety hazards. The set of critical targets typically includes categories of targets that significantly impact driving safety, such as pedestrians, non-motorized vehicles, and obstacles in construction areas. This set is predefined based on the application scenario and safety standards. The indicative function 𝕀() takes a value of 1 if the condition within the parentheses is true, and 0 otherwise, used to filter targets that meet specific conditions. The confidence threshold is typically set between 0.5 and 0.7 to filter low-confidence detection results. The numerator of the ACR formula is the sum of all valid detections belonging to the critical target set, with a detection confidence exceeding the threshold, and weighted by uncertainty weights. The denominator is the total number of targets in the critical target set. Dividing the two yields the weighted coverage rate.

[0040] In applications, uncertainty weights are implemented using an exponential decay function, converting the target's location uncertainty into a weighting coefficient between 0 and 1. Location uncertainty is typically output by the uncertainty estimation network of the sensing module, representing the confidence interval of the detection result in its spatial location. When location uncertainty is low, the exponential function value is close to 1, indicating that the detection result is reliable and should be given a higher weight; when location uncertainty is high, the exponential function value approaches 0, indicating that the detection result is unreliable and its contribution to ACR calculation is significantly weakened. The location uncertainty decay coefficient controls the sensitivity of the weights to changes in uncertainty; this coefficient needs to be calibrated according to sensor characteristics and application scenarios.

[0041] Specifically, suppose the perception system detects 10 targets in a frame, 6 of which belong to the key target set. For these 6 key targets, the detection confidence of 5 exceeds the threshold of 0.6. Further calculation of the uncertainty weights for each target is performed. Assuming the position uncertainties of the 5 effectively detected targets are 0.1m, 0.2m, 0.15m, 0.3m, and 0.5m respectively, and the position uncertainty attenuation coefficient is set to 0.3m, the corresponding uncertainty weights are 0.72, 0.51, 0.61, 0.37, and 0.19 respectively. The ACR is calculated by summing all weights (2.40) and dividing by the total number of key targets (6), resulting in an ACR value of 0.40. This value reflects the actual effective coverage of the key targets by the perception system after considering detection uncertainty.

[0042] The definition of the key target set can employ a static definition method based on semantic categories, pre-specifying categories such as pedestrians and non-motorized vehicles as key targets. Alternatively, a dynamic definition method based on risk assessment can be used, dynamically determining whether a target poses a safety threat based on its relative position and motion state with the vehicle. Location uncertainty can be obtained using a direct estimation method based on Bayesian deep learning, where the neural network simultaneously outputs the detection results and uncertainty estimates. Alternatively, an indirect estimation method based on Monte Carlo sampling can be used, calculating the uncertainty through the distribution of results from multiple forward inferences.

[0043] This embodiment introduces an uncertainty weighting mechanism into the alarm coverage calculation, enabling the ACR (Audit Response Rate) metric to reflect not only the quantity of detections but also the quality of detection. Traditional coverage metrics only count binary results of detection (whether detected or not), failing to distinguish between high-confidence and low-confidence detections, potentially leading to inflated coverage figures that mask the true performance of the sensing system. This method transforms location uncertainty into a weighting coefficient using an exponential decay function, effectively suppressing the contribution of high-uncertainty detections to the coverage rate, thus making the ACR metric a more realistic and reliable reflection of the sensing system's effective detection capability.

[0044] In one embodiment, the false alarm rate is calculated according to the following formula: Where M is the total number of detected targets, o j For the j-th detected target, 𝒪 critical For the set of key objectives, τ j Let τ be the tracking duration of the j-th detected target. track is the minimum tracking time threshold, and 𝕀() is the indicator function.

[0045] In applications, the false alarm rate (FAR) is a normalized metric between 0 and 1, used to quantify the degree to which a perception system misclassifies non-critical targets as valid detection results. Non-critical targets refer to detection results that do not belong to a pre-defined set of critical targets. These may be real objects that do not affect driving safety, or they may be false targets generated by sensor noise or algorithmic misdetection. The numerator of the FAR formula is the number of detections that do not belong to the set of critical targets and whose tracking time is less than a threshold; the denominator is the total number of detected targets. Dividing the two yields the false alarm rate.

[0046] In applications, tracking duration constraints are a key innovation in FAR calculations. Tracking duration refers to the number of frames in which a detected target is continuously detected across multiple consecutive frames. Real-world objects can typically be stably tracked for multiple frames, while false targets generated by sensor noise or transient interference often only exist for one or two frames before disappearing. The minimum tracking duration threshold is usually set to 3 to 5 frames, determined based on the sensor frame rate and application scenario. Only when the tracking duration of non-critical targets is less than this threshold is it included in the false alarm statistics. This effectively filters out real background objects that appear briefly due to physical reasons, allowing the FAR metric to focus more on the actual false detection problems of the perception system.

[0047] Specifically, suppose the perception system detects a total of 15 targets in a certain frame, of which 6 belong to the key target set and 9 do not. Among these 9 non-key targets, 3 have a tracking duration of less than the minimum tracking duration threshold of 5 frames, namely 1 frame, 2 frames, and 3 frames, while the tracking duration of the remaining 6 non-key targets all exceed 5 frames. The FAR is calculated by dividing the number of non-key targets that meet the criteria (3) by the total number of detected targets (15), resulting in an FAR value of 0.20. This value indicates that 20% of the detection results are short-term false alarms, requiring attention to the stability of the perception algorithm.

[0048] The tracking duration can be calculated using a frame counting method based on target identifiers, assigning a unique identifier to each detected target and performing correlation matching in subsequent frames. Alternatively, an implicit tracking method based on spatiotemporal consistency can be used, determining whether a target is the same target based on the continuity of its position and velocity. The minimum tracking duration threshold can be set using an empirically based fixed threshold method or an adaptive threshold method based on scene characteristics, dynamically adjusting the threshold according to factors such as vehicle speed and environmental complexity.

[0049] This embodiment effectively distinguishes between instantaneous noise-related false alarms and persistent false detections by introducing a tracking duration constraint into the false alarm rate calculation. Traditional false alarm rate metrics treat all non-critical targets equally in the false alarm statistics, failing to differentiate between occasional noise and systematic false detections, and easily overestimating the false alarm level of the sensing system. This method filters short-term detections by using a tracking duration threshold, making the FAR metric more focused on persistent false alarms that have a real impact on safety assessments, thus improving the diagnostic value and practicality of the metric.

[0050] In one embodiment, the classification quality score is calculated according to the following formula: Where N is the total number of perceived targets. Let be the category probability distribution vector of the i-th perceived target. The information entropy of the probability distribution vector of this category. Let be the intersection-union ratio (IoU) of the detection bounding box and the ground truth bounding box for the i-th perceived target. Let i be the predicted category of the i-th perceived target. Let be the truth class of the i-th perceived target, and let 𝕀() be the indicator function.

[0051] In applications, the Classification Quality Score (CQS) is a comprehensive metric for evaluating the classification accuracy and certainty of a perceptual system. The category probability distribution vector is the softmax probability value output by the classification layer of a deep learning object detection network, representing the probability that the target belongs to each candidate category. Information entropy H is a classic measure of the uncertainty of a random variable in information theory, calculated as the negative sum of probabilities multiplied by their logarithms. When the classification result is highly certain, the probability distribution approaches a one-hot form, and the information entropy approaches 0; when the classification result is uncertain, the probability distribution approaches a uniform distribution, and the information entropy reaches its maximum value. The CQS formula transforms the entropy penalty into a deterministic reward by subtracting the information entropy from 1.

[0052] In applications, the Intersection over Union (IoU) is a standard evaluation metric in object detection, defined as the area of ​​the intersection of the detected bounding box and the ground truth bounding box divided by the area of ​​their union. IoU ranges from 0 to 1; a higher value indicates a greater degree of overlap between the detected and ground truth bounding boxes, resulting in higher localization accuracy. The indicator function φ() is used to determine whether the predicted category matches the ground truth category; only correctly classified targets contribute positively to the CQS (Classification Quality Score). The CQS formula multiplies the entropy penalty term, IoU, and classification correctness, achieving a multi-dimensional comprehensive evaluation of the detection results, and averaging the results across all targets to obtain the overall classification quality score.

[0053] Specifically, suppose the perception system detects four targets, and the category probability distribution vectors output by the classification network are [0.95, 0.03, 0.02], [0.70, 0.20, 0.10], [0.50, 0.30, 0.20], and [0.80, 0.15, 0.05], with corresponding information entropies of 0.25, 0.80, 1.03, and 0.60, respectively. Subtracting the entropy value from 1 yields 0.75, 0.20, -0.03 (taking 0), and 0.40, respectively. Assume the IoU of these four targets are 0.85, 0.72, 0.68, and 0.90, respectively, and that targets 1, 2, and 4 are correctly classified, while target 3 is misclassified. The contributions of each objective to the CQS are 0.75 × 0.85 × 1 = 0.64, 0.20 × 0.72 × 1 = 0.14, 0 × 0.68 × 0 = 0, and 0.40 × 0.90 × 1 = 0.36, respectively. The CQS is calculated by dividing the sum of the contributions (1.14) by the total number of objectives (4), resulting in a CQS value of 0.29.

[0054] The information entropy can be calculated using the standard method of directly using the softmax output probability, or a temperature-scaled calibrated probability can be used to improve the accuracy of entropy estimation. The intersection-over-union ratio (IoU) can be calculated using the standard IoU calculation method with axis-aligned bounding boxes, or a generalized IoU calculation method using rotated bounding boxes or polygon masks can be used to adapt to different detection task requirements.

[0055] This embodiment effectively suppresses the impact of low-confidence classification results on the overall quality assessment by introducing an information entropy penalty mechanism into the classification quality score. Traditional classification accuracy indicators only count the binary result of whether the classification is correct or not, which cannot reflect the degree of certainty of the classification result. The perception system may give a correct but uncertain classification result on some difficult samples. Although this is statistically correct, it has limited support for the reliability of subsequent decisions. This method forcibly reduces the contribution weight of uncertain classification results by penalizing entropy value, so that the CQS index more realistically reflects the classification reliability of the perception system.

[0056] In one embodiment, the uncertainty quality score is calculated according to the following formula: Where B is the total number of intervals, Let b be the set of sample indices for the b-th interval. Let be the number of samples in the b-th interval. Let i be the predicted category of the i-th perceived target. Let be the truth class of the i-th perceived target. Let be the average uncertainty of the b-th interval, and let be the indicator function.

[0057] In applications, the Uncertainty Quality Score (UQS) is a metric that measures the accuracy of a sensing system's estimation of its own predictive uncertainty. Based on the concept of expected calibration error, this metric assesses whether the uncertainty estimate output by the sensing system matches the actual prediction accuracy. Ideally, when a sensing system reports 80% certainty, its prediction accuracy should also be close to 80%, meaning the uncertainty estimate is highly consistent with actual performance. A lower UQS value indicates better calibration, and a UQS of 0 indicates that the uncertainty estimate is completely accurate.

[0058] In application, the calculation of UQS first requires dividing all predicted samples into multiple intervals based on their uncertainty estimates. Interval division can be done using equal-width binning or equal-frequency binning, with the total number of intervals B typically set to 10 to 20. For each interval, the proportion of correctly predicted samples within that interval is calculated as the actual accuracy, and the average of the uncertainty estimates for all samples within that interval is calculated as the expected accuracy. UQS calculates the average of the absolute values ​​of the differences between the actual accuracy and the expected accuracy for each interval; a smaller difference indicates better calibration.

[0059] Specifically, assuming the perception system outputs 100 prediction results, divided into 5 intervals based on the uncertainty estimate. The first interval contains 20 samples, with an average uncertainty of 0.1 (i.e., 90% certainty), an actual accuracy of 17 / 20 = 0.85, and an absolute difference of |0.85 - 0.90| = 0.05. The second interval contains 25 samples, with an average uncertainty of 0.25, an actual accuracy of 18 / 25 = 0.72, and an absolute difference of |0.72 - 0.75| = 0.03. The differences for the remaining intervals are calculated similarly, assuming the differences for the 5 intervals are 0.05, 0.03, 0.08, 0.12, and 0.07, respectively. The Uncertainty Level (UQS) is calculated by dividing the sum of the differences (0.35) by the number of intervals (5), resulting in a UQS value of 0.07. This value indicates that the uncertainty estimate of the perception system deviates from the actual performance by an average of 7%, indicating a acceptable level of calibration.

[0060] Among the methods for dividing the intervals, equal-width binning can be used to divide the uncertainty range into several equal intervals, or adaptive binning can be used to dynamically adjust the interval boundaries according to the sample distribution to ensure a relatively balanced number of samples in each interval. Uncertainty estimates can be obtained using direct estimation methods based on Bayesian neural networks, or indirect estimation methods based on model ensembles, inferring uncertainty by the degree of consistency between the prediction results of multiple models.

[0061] This embodiment assesses the degree of uncertainty calibration in a perception system using the Uncertainty Quality Score (UQS), filling a gap in existing technologies for evaluating the metacognitive capabilities of perception systems. A perception system not only needs to provide accurate predictions but also needs to accurately estimate the reliability of its own predictions. Only systems with accurate uncertainty estimations can provide reliable confidence references for subsequent decision-making and planning. This method quantifies the degree of uncertainty estimation calibration using the UQS index, enabling security assessments to identify overconfident or overly conservative perception systems, thus improving the comprehensiveness and depth of the assessment.

[0062] In one embodiment, the threshold function is expressed as: ;in, Let be the dynamic safety threshold at time t. α is the baseline threshold, β is the alarm coverage adjustment coefficient, β is the false alarm rate adjustment coefficient, ACR(t) is the alarm coverage at time t, and FAR(t) is the false alarm rate at time t; , The dynamic security threshold decreases when alarm coverage increases and false alarm rate decreases.

[0063] In application, the threshold function is the core mechanism for establishing a dynamic mapping relationship between perceived quality evaluation indicators and safety thresholds. The baseline threshold is typically set to a collision time of 2 to 3 seconds, representing the default safety boundary when perceived quality is at a standard level. The alarm coverage adjustment coefficient α and the false alarm rate adjustment coefficient β are two hyperparameters with values ​​between 0 and 1, controlling the influence of ACR and FAR on the safety threshold, respectively. These two coefficients need to be calibrated based on vehicle dynamics characteristics, driver reaction time, and safety regulatory requirements.

[0064] In application, the design of the threshold function reflects the negative correlation between perception quality and safety margin. A high Alarm Coverage Rate (ACR) indicates a strong detection capability of the perception system for critical targets, allowing the system to appropriately lower the safety threshold to avoid overly conservative driving. Conversely, a high False Alarm Rate (FAR) indicates numerous false alarms, requiring the system to raise the safety threshold to reserve more buffer space for handling false alarms. The function's multiplicative structure allows improvements in perception quality to be linearly translated into adjustments to the safety threshold.

[0065] Specifically, assuming a baseline threshold of 2.5 seconds, an alarm coverage adjustment coefficient of 0.3, and a false alarm rate adjustment coefficient of 0.2, under normal conditions, the ACR is 0.92 and the FAR is 0.05. Substituting these values ​​into the formula, the dynamic safety threshold is calculated to be 2.5 × (1 - 0.3 × 0.92 + 0.2 × 0.05) = 2.5 × 0.734 = 1.84 seconds. Under severe weather conditions, the ACR drops to 0.78 and the FAR rises to 0.15. Substituting these values ​​into the formula, the dynamic safety threshold is calculated to be 2.5 × (1 - 0.3 × 0.78 + 0.2 × 0.15) = 2.5 × 0.796 = 1.99 seconds. It is evident that when the perceived quality deteriorates, the safety threshold automatically increases, providing the system with a greater safety margin.

[0066] The calibration of the adjustment coefficients can employ offline optimization methods based on simulation testing, searching for the optimal combination of coefficients that balances safety and traffic efficiency in a virtual scenario. Alternatively, online learning methods based on real-vehicle data can be used to dynamically adjust the coefficients according to historical driving data to adapt to different drivers' risk preferences. The benchmark threshold can be set using a fixed value method based on safety regulations, or a dynamic adjustment method based on vehicle speed, with a larger benchmark threshold used for high-speed driving.

[0067] This embodiment achieves dynamic coupling between perception quality and safety threshold through a threshold threshold function, overcoming the limitations of traditional fixed threshold methods. Traditional methods use a uniform collision time threshold for risk assessment, which cannot adapt to fluctuations in perception performance under different scenarios. This method enables the safety threshold to adaptively adjust based on real-time perception quality. When perception confidence is high, the threshold is appropriately relaxed to improve traffic efficiency, and when perception confidence is low, the threshold is automatically tightened to ensure safety, achieving a dynamic balance between safety and efficiency.

[0068] In one embodiment, the risk weight allocation matrix is ​​a 4×3 matrix used to map alarm coverage, false alarm rate, classification quality score, and uncertainty quality score to three levels: low risk, medium risk, and high risk. The sum of the elements in each column of the risk weight allocation matrix is ​​1. The fused risk value is calculated according to the following formula: ;in, To integrate risk values, TTC stands for Collision Time. For the normalized perceived quality evaluation index vector and W is the risk weight allocation matrix, and r is the risk level coefficient vector. , The risk level coefficient is low. The risk level coefficient is medium. It represents a high-risk level coefficient.

[0069] In application, the risk weight allocation matrix is ​​a 4x3 real number matrix. The four rows correspond to four perceived quality indicators: alarm coverage, false alarm rate, classification quality score, and uncertainty quality score. The three columns correspond to three risk levels: low risk, medium risk, and high risk. Each element of the matrix represents the contribution weight of the corresponding perceived quality indicator to the corresponding risk level. The sum of the elements in each column is normalized to 1 to ensure that the total weight allocation for each risk level is consistent. This matrix is ​​obtained through offline training using a fuzzy C-means clustering algorithm based on historical accident scenario data.

[0070] In application, the normalized perceived quality evaluation index vector transforms the four original indicators into a unified positive index form. Alarm Coverage Rate (ACR) and Classification Quality Score (CQS) are inherently positive indicators; higher values ​​indicate better quality, and they are directly incorporated into the vector. False Alarm Rate (FAR) and Uncertainty Quality Score (UQS) are negative indicators; lower values ​​indicate better quality, so they are inverted by subtracting the original value from the value by 1. The risk level coefficient vector defines the numerical scales for low, medium, and high risk levels, typically set to 0.3, 0.6, and 1.0, indicating that the weight of high risk is more than three times that of low risk.

[0071] In application, the calculation of the fusion risk value is a multi-level matrix operation process. First, the normalized perception quality index vector is multiplied by the risk weight allocation matrix to obtain a three-dimensional risk level distribution vector, representing the membership degree of the current perception state at each risk level; then, this distribution vector is multiplied by the risk level coefficient vector to obtain the weighted risk score; finally, it is multiplied by the collision time TTC to add the perception quality impact factor to the traditional collision risk assessment.

[0072] Specifically, assuming the current perceived quality indicators are ACR=0.85, FAR=0.10, CQS=0.78, and UQS=0.12, the normalized indicator vector is [0.85, 0.90, 0.78, 0.88]. Assuming the risk weight allocation matrix is ​​[[0.4,0.3,0.2],[0.2,0.3,0.4],[0.2,0.2,0.2],[0.2,0.2,0.2]], and the risk level coefficient vector is [0.3,0.6, 1.0]. Multiplying the index vector by the weight matrix yields [0.85×0.4+0.90×0.2+0.78×0.2+0.88×0.2,...]=[0.72, 0.78, 0.85]. Then, the dot product of this with the coefficient vector yields 0.72×0.3+0.78×0.6+0.85×1.0=1.53. If the current collision time TTC is 2.5 seconds, the fusion risk value is 2.5×1.53=3.83.

[0073] The training of the risk weight allocation matrix can employ fuzzy C-means clustering, using unsupervised learning based on perceived quality indicators and risk level labels from historical accident scenarios. Alternatively, supervised learning methods such as support vector machines or neural networks can be used to directly learn the mapping relationship between indicators and risk levels. The risk level coefficient vector can be set using a fixed-value method based on safety standards, or an adjustable method based on risk preferences, allowing users to adjust the relative weights of each level according to their needs.

[0074] This embodiment achieves a soft mapping from perception quality indicators to risk levels through a risk weight allocation matrix, and then multiplies and fuses them with collision time to construct a fused risk value that comprehensively considers perception reliability and collision risk. Traditional methods directly transmit the perception layer output to the decision layer, and perception quality information cannot affect the risk assessment results. This method compresses multidimensional perception quality indicators into a single risk adjustment factor through matrix operations, coupling it with collision time. This allows the fused risk value to simultaneously reflect information from both the "how close the collision is" and "how reliable the perception is," improving the comprehensiveness of the safety assessment.

[0075] In one embodiment, before calculating the perception quality evaluation index, the method further includes: receiving multi-source sensor data and performing spatiotemporal alignment processing; generating an uncertainty tensor for each perception target through an uncertainty quantization engine, wherein each dimension of the uncertainty tensor corresponds to the uncertainty of position, category, velocity, and orientation, respectively.

[0076] In applications, receiving and spatiotemporally aligning multi-source sensor data refers to acquiring raw sensing data from various sensors such as LiDAR, cameras, and millimeter-wave radar, and unifying the data from each sensor into the same spatiotemporal reference frame through time synchronization and spatial calibration. Time synchronization employs a hardware timestamp mechanism to ensure that the acquisition time deviation of each sensor's data is controlled within 1 millisecond; spatial calibration uses an extrinsic parameter matrix to transform the coordinate systems of each sensor to the vehicle coordinate system, ensuring that the spatial alignment error is controlled within 5 centimeters. Spatiotemporal alignment is the foundation of multi-sensor fusion, and the alignment accuracy directly affects the performance of subsequent sensing algorithms.

[0077] In application, the uncertainty quantization engine is a Bayesian deep learning-based neural network module that outputs an uncertainty tensor for each perceived target. The dimensions of this tensor correspond to the target attribute dimensions, including four components: position uncertainty, category uncertainty, velocity uncertainty, and orientation uncertainty. Position uncertainty represents the size of the confidence interval for the target's 3D coordinate estimation; category uncertainty represents the degree of certainty in the classification result; velocity uncertainty represents the reliability of the target's motion state estimation; and orientation uncertainty represents the accuracy of the target's heading angle estimation. The uncertainty tensor provides the foundational data for subsequent calculations of perception quality indicators.

[0078] Specifically, the data preprocessing process is illustrated using an autonomous vehicle driving on urban roads as an example. The vehicle is equipped with one 128-line LiDAR, six surround-view cameras, and five millimeter-wave radars. The LiDAR outputs point cloud data at a frequency of 10Hz, with approximately 100,000 points per frame; the cameras output image data at a frequency of 30Hz, with a resolution of 1920×1080 pixels; and the millimeter-wave radars output a target list at a frequency of 20Hz, with a maximum of 64 targets per frame. The spatiotemporal alignment module first timestamps the data from each sensor, using nearest-neighbor interpolation to unify data from different frequencies to a fusion frequency of 10Hz; then, it transforms the sensor data to the vehicle coordinate system using a pre-calibrated extrinsic parameter matrix. The aligned data is then fed into the fusion perception network, which outputs the target detection results and the corresponding uncertainty tensor.

[0079] Time synchronization can be achieved using a hardware synchronization method based on GPS time synchronization, using PPS signals to achieve microsecond-level time alignment. Alternatively, a post-processing synchronization method based on software timestamps can be used, compensating for the acquisition time deviations of each sensor through interpolation. Uncertainty quantification can be achieved using a Bayesian neural network method, learning the posterior distribution of weights through variational inference. Alternatively, a deep ensemble method can be used, estimating uncertainty by predicting the distribution of results from multiple independently trained networks.

[0080] This embodiment lays the data foundation for calculating perception quality evaluation indicators through spatiotemporal alignment and uncertainty quantification preprocessing of multi-source sensor data. Spatiotemporal alignment ensures that the data from each sensor are fused under a unified reference frame, avoiding fusion errors caused by spatiotemporal inconsistencies; uncertainty quantification enables the perception system to output not only detection results but also a self-assessment of the reliability of the results, providing crucial input for subsequent perception quality evaluation. These two preprocessing techniques together support the method's ability to quantitatively evaluate the adequacy of the perception module's functionality.

[0081] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0082] This application also provides an intelligent driving perception safety assessment device for performing the steps described in the intelligent driving perception safety assessment method embodiments. The intelligent driving perception safety assessment device can be a virtual appliance within an electronic device, run by the electronic device's processor, or it can be the electronic device itself.

[0083] As shown in Figure 3, the intelligent driving perception safety assessment device 100 provided in this embodiment includes: an index calculation module, used to calculate perception quality evaluation indicators based on the target detection results of the perception module, the perception quality evaluation indicators including alarm coverage rate, false alarm rate, classification quality score, and uncertainty quality score; a threshold adjustment module, used to dynamically adjust the safety threshold for collision risk assessment according to a threshold threshold function and the perception quality evaluation indicators; a fusion calculation module, used to fuse the perception quality evaluation indicators with the collision time based on a risk weight allocation matrix to obtain a fused risk value; and an assessment output module, used to output a safety assessment conclusion based on the comparison result between the fused risk value and the safety threshold.

[0084] In application, the index calculation module is the core computing unit of the device. It is responsible for receiving target detection result data from the sensing module and calculating four indicators—alarm coverage rate, false alarm rate, classification quality score, and uncertainty quality score—according to preset mathematical formulas. This module internally includes a data parsing submodule, an index calculation submodule, and a result caching submodule, which are responsible for format conversion of raw data, calculation of index values, and storage of historical results, respectively. The output of the index calculation module is a four-dimensional sensing quality evaluation index vector.

[0085] In the application, the threshold adjustment module is responsible for dynamically generating safety thresholds based on perceived quality evaluation indicators. This module has a built-in threshold function and its parameter configuration. It receives the perceived quality evaluation indicator vector output by the indicator calculation module as input and calculates the appropriate safety threshold for the current moment in real time. The threshold adjustment module is also responsible for online parameter updates and can fine-tune and optimize the adjustment coefficients based on system operation feedback.

[0086] In the application, the fusion computing module is responsible for fusing the perceived quality indicators with the collision time. This module has a built-in risk weight allocation matrix and risk level coefficient vector, receives the perceived quality evaluation indicator vector and the collision time as input, and outputs the fused risk value through matrix operations. The fusion computing module has low computational complexity and can meet real-time requirements.

[0087] In the application, the evaluation output module is responsible for making the final security judgment and outputting the evaluation conclusion. This module receives the fused risk value and dynamic security threshold as input, determines the current security status through numerical comparison, and outputs the evaluation conclusion to the downstream decision-making and planning module or the human-computer interaction interface. The evaluation output module is also responsible for recording evaluation logs, providing data support for offline analysis and continuous optimization of the system.

[0088] Specifically, in actual deployment, the device can be implemented using a modular software approach, with each module running on an independent thread and interacting with data via message queues. The index calculation module subscribes to the target detection result messages published by the perception module, and after calculation, publishes the perception quality evaluation index vector to the shared data bus; the threshold adjustment module and the fusion calculation module subscribe to this vector data, respectively calculating the dynamic safety threshold and the fusion risk value; the evaluation output module subscribes to the above two calculation results, compares and judges them, and then publishes a safety assessment conclusion message. The data transmission latency between modules is controlled within 1 millisecond, and the overall processing latency is controlled within 20 milliseconds, meeting the real-time requirements of the autonomous driving system.

[0089] Inter-module communication can employ a zero-copy transmission method based on shared memory for efficient data exchange, or a distributed deployment method based on network sockets to support module distribution across hosts. Module scheduling can use a time-triggered periodic execution method to ensure a fixed processing cycle, or an event-triggered on-demand execution method to immediately start computation when new data arrives.

[0090] The device provided in this embodiment adopts a modular design, with clearly defined responsibilities and standardized interfaces for each functional module, facilitating independent development, testing, and maintenance. The modular architecture also supports flexible functional expansion, allowing the addition of new evaluation index calculation modules or risk fusion modules as needed without modifying the implementation of other modules. The device as a whole possesses high real-time performance and high reliability, meeting the stringent requirements of intelligent driving systems for safety assessment functions.

[0091] This application also provides a vehicle that includes the intelligent driving perception and safety assessment device provided in the above embodiments.

[0092] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0093] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0094] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps described in the various method embodiments above.

[0095] This application provides a computer program product, including a computer program, which, when run on an electronic device, enables the electronic device to perform the steps described in the various method embodiments above.

[0096] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium can include at least: any entity or device capable of carrying computer program code to a device / electronic device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0097] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0098] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0099] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0101] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for assessing the safety of intelligent driving perception, characterized in that, include: Based on the target detection results of the perception module, a perception quality evaluation index is calculated, which includes alarm coverage rate, false alarm rate, classification quality score and uncertainty quality score. The safety threshold for collision risk assessment is dynamically adjusted based on the pre-constructed threshold function and the perceived quality evaluation index. Based on the risk weight allocation matrix, the perceived quality evaluation index and the collision time are fused together to obtain the fused risk value; Based on the comparison between the fusion risk value and the security threshold, a security assessment result is output.

2. The intelligent driving perception safety assessment method as described in claim 1, characterized in that, The alarm coverage rate is calculated according to the following formula: Where N is the total number of perceived targets. For the i-th perceived target, For the set of key objectives, Let be the detection confidence score for the i-th perceived target. Let be the confidence threshold, and let be the indicator function. The uncertainty weight is the weight for the i-th perceived target; the uncertainty weight is calculated according to the following formula: ;in, The positional uncertainty of the i-th perceived target. This is the position uncertainty attenuation coefficient.

3. The intelligent driving perception safety assessment method as described in claim 1, characterized in that, The false alarm rate is calculated according to the following formula: Where M is the total number of detected targets, o j For the j-th detected target, 𝒪 critical For the set of key objectives, τ j Let τ be the tracking duration of the j-th detected target. track is the minimum tracking time threshold, and 𝕀() is the indicator function.

4. The intelligent driving perception safety assessment method as described in claim 1, characterized in that, The classification quality score is calculated according to the following formula: Where N is the total number of perceived targets. Let be the category probability distribution vector of the i-th perceived target. The information entropy of the probability distribution vector of this category. Let be the intersection-union ratio (IoU) of the detection bounding box and the ground truth bounding box for the i-th perceived target. Let i be the predicted category of the i-th perceived target. Let be the truth class of the i-th perceived target, and let 𝕀() be the indicator function.

5. The intelligent driving perception safety assessment method as described in claim 1, characterized in that, The uncertainty mass fraction is calculated according to the following formula: Where B is the total number of intervals, Let b be the set of sample indices for the b-th interval. Let be the number of samples in the b-th interval. Let i be the predicted category of the i-th perceived target. Let be the truth class of the i-th perceived target. Let be the average uncertainty of the b-th interval, and let be the indicator function.

6. The intelligent driving perception safety assessment method as described in claim 1, characterized in that, The threshold function is expressed as follows: ;in, Let be the dynamic safety threshold at time t. α is the baseline threshold, β is the alarm coverage adjustment coefficient, β is the false alarm rate adjustment coefficient, ACR(t) is the alarm coverage at time t, and FAR(t) is the false alarm rate at time t; 、 The dynamic security threshold decreases when alarm coverage increases and false alarm rate decreases.

7. The intelligent driving perception safety assessment method as described in claim 1, characterized in that, The risk weight allocation matrix is ​​a 4×3 matrix used to map alarm coverage, false alarm rate, classification quality score and uncertainty quality score to three levels: low risk, medium risk and high risk.

8. The intelligent driving perception safety assessment method as described in claim 1, characterized in that, The fusion risk value is calculated according to the following formula: ;in, To integrate risk values, TTC stands for Collision Time. For the normalized perceived quality evaluation index vector and W is the risk weight allocation matrix, and r is the risk level coefficient vector. ， The risk level coefficient is low. The risk level coefficient is medium. It represents a high-risk level coefficient.

9. An intelligent driving perception and safety assessment device, characterized in that, include: The indicator calculation module is used to calculate the perception quality evaluation index based on the target detection results of the perception module. The perception quality evaluation index includes alarm coverage rate, false alarm rate, classification quality score and uncertainty quality score. The threshold adjustment module is used to dynamically adjust the safety threshold for collision risk assessment based on the threshold threshold function and the perceived quality evaluation index. The fusion calculation module is used to perform fusion calculation on the perceived quality evaluation index and the collision time based on the risk weight allocation matrix to obtain the fusion risk value; The evaluation output module is used to output a security evaluation conclusion based on the comparison result between the fusion risk value and the security threshold.

10. A computer program product, characterized in that, Includes a computer program, which, when run, causes the method as described in any one of claims 1-8 to be performed.

Images