Complex scene edge quantification evaluation method based on multi-class risk field model

Abstract

This invention discloses a method for quantitatively evaluating the edge properties of complex scenarios based on a multi-category risk field model, belonging to the field of autonomous driving technology. The method includes: analyzing and classifying temporal scene elements based on the PEGASUS system; constructing a large-scale risk field for unidentified scene elements; constructing a medium-scale risk field for low-confidence identified scene elements; constructing an anisotropic driving behavior risk field for high-confidence identified scene elements; superimposing multiple risk fields within the interaction area; and quantifying the edge properties of the scenario using the minimum travel cost of the main vehicle's path. This invention achieves accurate quantification of the edge properties of complex scenarios by differentially modeling multiple risk sources and superimposing field strengths, utilizing the optimal path cost. It solves the problem that existing methods struggle to uniformly characterize multiple risk categories and overall travel costs, providing strong support for risk assessment and edge scenario mining for autonomous vehicles.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and specifically relates to a method for quantitative evaluation of the edge properties of complex scenarios based on a multi-category risk field model. Background Technology

[0002] With the continuous development of core technologies for autonomous driving, the perception, prediction, and decision-making capabilities of systems in conventional scenarios have been significantly improved. However, in complex scenarios with severe occlusion, abnormal behavior of traffic participants, uncertain identification of scene elements, and the presence of unknown obstacles, autonomous driving systems are still prone to performance degradation, leading to high-risk interactions and even traffic accidents. These scenarios, characterized by low probability, high risk, and tightly coupled interactions, are considered edge scenarios. Accurate evaluation of these scenarios is crucial for the testing, verification, risk assessment, and functional improvement of autonomous driving systems.

[0003] Existing scenario risk assessment methods are mostly based on a single indicator or a single type of target, such as measuring scenario risk by collision time, minimum distance, or single target potential field. Although such methods can reflect the local danger trend in the scenario, they are difficult to uniformly characterize multiple risk sources such as unidentified scenario elements, low-confidence identified scenario elements, and high-confidence elements with abnormal behavior. In particular, for obscured traffic participants, unknown objects on the road, and vehicles with sudden cut-in, sudden deceleration, and abnormal behavior, existing methods are difficult to achieve accurate assessment of the edge of such scenarios.

[0004] Furthermore, the marginality of complex scenarios is not only related to the instantaneous risk of individual traffic participants, but also closely related to the composite field strength distribution formed by multiple risk sources within the interaction area between the driver and vehicle, as well as the vehicle's accessibility. Describing a scenario using only local indicators is insufficient to reflect the overall passage cost faced by the vehicle when choosing a safe path throughout the entire interaction area. Therefore, there is an urgent need for an evaluation method that can establish differentiated risk fields for multiple types of risk objects and further quantify the marginality of scenarios using the minimum passage cost of the driving path. Summary of the Invention

[0005] To address the above problems, this invention provides a method for quantitatively evaluating the marginality of complex scenarios based on a multi-category risk field model, comprising the following steps:

[0006] Step 1: Analyze and classify the components of the time-series scene;

[0007] Based on the PEGASUS system, the elements of time-series scenes are deeply analyzed. The elements of scenes are expressed in a structured manner according to the road layer, road facilities and rules layer, temporary change layer, dynamic object layer, environmental condition layer and digital information layer. Visual modal and trajectory modal information are integrated to form a scene element representation for complex scenes. The scene elements are classified into unidentified scene elements, low-confidence identified scene elements and high-confidence identified scene elements.

[0008] Step 2: Construct a large-scale risk field for unidentified scene elements;

[0009] For unidentified scene elements, an impact score for unidentified scene elements is constructed, and a risk field for unidentified scene elements is established, assigning a range of the general trend field and a high field strength weight.

[0010] Step 3: Construct a medium-range risk field for low-confidence identification scene elements;

[0011] For low-confidence recognition scene elements, the field strength amplitude and potential field range of the low-confidence recognition scene elements are constructed. Based on the recognition probability and the interaction relationship with the main vehicle, a risk field of the low-confidence recognition scene elements is constructed, and a medium-range risk field is assigned to achieve quantitative modeling of uncertain risks.

[0012] Step 4: Construct an anisotropic driving behavior risk field for high-confidence scene elements;

[0013] For high-confidence scene elements, the field strength amplitude and potential field range of the high-confidence scene elements are constructed based on their speed, acceleration, heading changes and relative proximity trends. A driving behavior risk field of the high-confidence scene elements is constructed. When the high-confidence scene elements have abnormal and dangerous behaviors, a large field strength and potential field range are assigned. When the high-confidence scene elements drive safely, a small field strength and potential field range are assigned.

[0014] Step 5: Overlay multiple risk fields and quantify the edge of the scenario;

[0015] Within the vehicle interaction area, the risk fields of multiple types of traffic participants are superimposed to form a composite field strength. The marginality of the scenario is measured based on the minimum passage cost when the main vehicle selects the optimal path. A higher minimum passage cost indicates higher marginality, and vice versa.

[0016] Furthermore, in step one, the method for deeply analyzing the components of a temporal scene and expressing them in a structured manner is as follows:

[0017] Based on the PEGASUS system, a hierarchical analysis of complex scenes is performed. The road layer, road facilities and rules layer, temporary change layer, dynamic object layer, environmental condition layer and digital information layer are uniformly mapped to the temporal scene state. Visual modality and trajectory modality information are synchronously written into the state description. The scene state representation at time t is shown in Equation (1):

[0018] (1)

[0019] In the formula, Let t be the scene state at time t. This refers to the time-series information of the road layer. For road infrastructure and rule layer time sequence information, Temporary change layer timing information, For dynamic object layer timing information, This refers to the temporal information of the environmental conditions layer. For digital information layer time sequence information, For visual modal temporal information, This refers to the trajectory mode temporal information.

[0020] Furthermore, in step one, the method for fusing and classifying the visual and trajectory-based dual-modal scene elements is as follows:

[0021] For the i-th scene element in the scene, its visual features and trajectory features are extracted and linearly fused to obtain a bimodal element representation, as shown in Equation (2):

[0022] (2)

[0023] In the formula, Let be the bimodal fusion feature of the i-th scene element at time t. Let be the visual feature vector of the i-th scene element. Let be the trajectory feature vector of the i-th scene element. and These are the visual feature mapping matrix and the trajectory feature mapping matrix, respectively. For fusion bias term;

[0024] After obtaining the bimodal feature representation, the feature category probability distribution is output by the classifier, as shown in Equation (3):

[0025] (3)

[0026] In the formula, Let be the category probability distribution of the i-th scene element. For normalized classification functions, The classifier weight matrix is... For classifier bias terms;

[0027] To classify objects into different risk categories, a classification strategy based on confidence thresholds is adopted, as shown in equations (4)(5)(6):

[0028] (4)

[0029] (5)

[0030] (6)

[0031] In the formula, Let be the category label for the i-th scene element. This indicates taking the category label that maximizes the category probability. Let t be the set of low-confidence scene elements identified at time t. Let be the set of scene elements for high-confidence recognition at time t. This represents the maximum value among the probabilities of the i-th scene element category. For low confidence threshold, For high confidence threshold, and Greater than .

[0032] Furthermore, in step two, the method for constructing the score for the impact of unidentified scene elements is as follows:

[0033] The unidentified scene elements refer to areas that are not recognized by the autonomous driving system but may affect the driver vehicle due to occlusion, unknown obstacles, or abnormal occupation of the area in front of the driver vehicle. An impact score for unidentified scene elements is constructed, as shown in Equation (7):

[0034] (7)

[0035] In the formula, The impact score for the j-th unidentified scene element at time t, with a value between 0 and 1. , and These are the weighted coefficients corresponding to the degree of occlusion, abnormal movement, and relevance to the interaction with the main vehicle, respectively. Let represent the occlusion degree of the j-th unidentified scene element. Let be the abnormal motion quantity of the j-th unidentified scene element, used to measure whether there is motion information within this unidentified scene element that is difficult to directly identify but continuously affects the driver's decision. Let be the main vehicle interaction relevance of the j-th unidentified scene element, representing the degree of coupling between the unidentified scene element and the main vehicle path.

[0036] Furthermore, in step two, the method for establishing the risk field of unidentified scene elements is as follows:

[0037] The field strength amplitude of unidentified scene elements is determined based on the impact score of unidentified scene elements. As shown in equation (8):

[0038] (8)

[0039] In the formula, The base field strength representing unidentified scene elements. The representative factor is the adjustment coefficient for the impact score. The impact score for the j-th unidentified scene element at time t;

[0040] The potential field range of unidentified scene elements is determined based on the influence score of the unidentified scene elements, and larger radius of influence is given priority, as shown in equation (9):

[0041] (9)

[0042] In the formula, Let j be the potential field range of the j-th unidentified scene element. The radius of the basic potential field for unidentified scene elements. This is the potential field range adjustment coefficient;

[0043] A Gaussian risk field is established for the unidentified scene elements, as shown in Equation (10):

[0044] (10)

[0045] In the formula, Let be the risk field value generated by the j-th unidentified scene feature at the scene coordinate point (x, y). and These are the center coordinates of the j-th unidentified scene element. and These represent the field diffusion scales of the j-th unidentified scene element in the horizontal and vertical directions, respectively. , , , These represent the horizontal and vertical diffusion ratios, respectively.

[0046] Furthermore, in step three, the method for constructing the field strength amplitude and potential field range of low-confidence scene elements is as follows:

[0047] The low-confidence recognition scene element refers to an uncertain scene element that is detected but whose recognition probability is lower than a certain threshold. First, its maximum recognition probability is extracted, and it is determined to belong to the low-confidence recognition object according to the threshold, as shown in Equation (11):

[0048] (11)

[0049] In the formula, Let be the maximum recognition probability of the i-th recognized scene element at time t. Let be the category probability distribution of the i-th scene element. For low confidence threshold, For high confidence threshold, and Greater than ;

[0050] The field strength amplitude of low-confidence scene elements is determined based on the maximum recognition probability of the low-confidence scene elements. As shown in equation (12):

[0051] (12)

[0052] In the formula, The basic field strength representing elements in a low-confidence identification scenario. The coefficient represents the adjustment factor for uncertainty.

[0053] The potential field range of low-confidence recognition scene elements is determined based on the recognition probability, as shown in equation (13):

[0054] (13)

[0055] In the formula, Let be the potential field range of the i-th low-confidence scene element. The radius of the basic potential field for elements in low-confidence scenarios. This is the confidence level inverse adjustment coefficient;

[0056] Furthermore, in step three, the method for constructing the risk field of low-confidence identification scene elements is as follows:

[0057] The risk field for low-confidence scene elements is shown in equation (14):

[0058] (14)

[0059] In the formula, Let be the risk field value generated by the i-th low-confidence identified scene feature at the scene coordinate point (x, y). Let be the field strength amplitude of the i-th low-confidence scene element. and These are the location coordinates of the i-th low-confidence scene element. and , respectively, represent the field diffusion scales of the i-th low-confidence scene element in the horizontal and vertical directions. and , , These represent the horizontal and vertical diffusion ratios, respectively. Let be the potential field range of the i-th low-confidence scene element;

[0060] The lower the maximum recognition probability of scene elements in low-confidence recognition, the better. The larger the value, the larger the potential field range of the low-confidence scene elements. However, the overall potential field range of the low-confidence scene elements is smaller than that of the unidentified scene elements.

[0061] Furthermore, in step four, the method for constructing the field strength amplitude and potential field range of the scene elements for high-confidence recognition is as follows:

[0062] The high-confidence recognition scene element refers to a scene element that has been detected and whose recognition probability is higher than a certain threshold. A risk field is established by combining the driving behavior of the high-confidence recognition scene element, and the risk level is increased when the element has abnormal and dangerous behavior.

[0063] The behavioral anomaly index is constructed as shown in equation (15):

[0064] (15)

[0065] In the formula, For time t, the first A high-confidence indicator for identifying the behavioral anomalies of scene elements. , , , and These are the weighting coefficients for longitudinal acceleration, angular velocity, relative approach speed, collision time, and lane behavior, respectively. For the first Longitudinal acceleration of each traffic participant The normalized constant for longitudinal acceleration, For the first The heading angular velocity of each traffic participant Let be the normalized constant for the heading angular velocity. For the first The relative approach speed of each traffic participant to the host vehicle. Let be the velocity normalization constant. For the first The collision time between a traffic participant and the main vehicle Let be the collision time normalization constant. For the first The abnormal lane behavior indicator value for each traffic participant increases when the element exhibits dangerous cutting in, continuous lane crossing, abnormal lane changing or other high-risk behaviors.

[0066] The field strength amplitude and potential field range of the high-confidence recognition scene elements are determined based on the degree of behavioral anomaly, as shown in Equation (16):

[0067] (16)

[0068] In the formula, For the first The field strength amplitude of a high-confidence scene element. The basic field strength for high-confidence identification of scene elements. For the first The potential field range of a high-confidence recognition scene element. The radius of the basic potential field for high-confidence identification of scene elements. and These are the adjustment coefficients for the anomaly index.

[0069] Furthermore, in step four, the method for constructing a high-confidence recognition scene element driving behavior risk field is as follows:

[0070] An anisotropic driving behavior risk field model in a local coordinate system is adopted, as shown in equations (17) and (18):

[0071] (17)

[0072] (18)

[0073] In the formula, For the first A high-confidence identification of the risk field value generated by scene elements at scene coordinate point (x,y). and These represent the vertical and horizontal coordinates of the coordinate point (x, y) relative to the feature's position in the feature's local coordinate system. and These represent the field diffusion scales of scene elements in the vertical and horizontal directions, respectively, for high-confidence recognition. and , , These represent the longitudinal and lateral diffusion ratios, respectively. For heading angle Constructed two-dimensional rotation matrix, For the first The heading angle for identifying scene elements with high confidence. For the first The field strength amplitude of a high-confidence scene element. For the first The potential field range of a high-confidence recognition scene element. Represents the transpose operator. , Representing the first High-confidence identification elements in Always in the scene's global coordinate system , The position coordinates of the direction.

[0074] Furthermore, the method for step five is as follows:

[0075] First, construct the composite field strength of the interactive area:

[0076] Within the designated interaction area of ​​the main vehicle, this invention superimposes the risk fields generated by the behaviors of unidentified scene elements, low-confidence identified scene elements, and high-confidence identified scene elements to form a composite field strength, as shown in equation (19):

[0077] (19)

[0078] In the formula, Let be the combined field strength value at the scene coordinate point (x, y) at time t. Let t be the set of unidentified scene elements. Let t be the set of low-confidence scene elements identified at time t. Let t be the set of scene elements for high-confidence identification. Let be the risk field value generated by the j-th unidentified scene feature at the scene coordinate point (x, y). Let be the risk field value generated by the i-th low-confidence identified scene feature at the scene coordinate point (x, y). For the first A high-confidence identification of the risk field value generated by scene elements at scene coordinate point (x,y);

[0079] Secondly, the edge case of a scenario is quantified based on the minimum passage cost of the main vehicle's travel path:

[0080] Let γ be the candidate path of the main vehicle in the interaction area at time t. In this invention, the cumulative travel cost of the path in the composite field strength is used as the path evaluation index, as shown in equation (20):

[0081] (20)

[0082] In the formula, Let be the travel cost of candidate path γ at time t. Let γ be the position of the candidate path γ under path parameter s. Let γ be the first derivative of the candidate path with respect to the path parameter s. The Euclidean norm is two-dimensional; μ is the field strength cost weighting coefficient. express Time Scene Coordinates The combined field strength at that location, This indicates that the composite field strength function is along the candidate path Values, i.e., path parameters Corresponding location point Composite field strength at the location, path parameters Obtained by geometric parameterization of the vehicle's travel path;

[0083] The minimum travel cost is calculated for all candidate paths to obtain the optimal path cost for the main vehicle, as shown in equation (21):

[0084] (twenty one)

[0085] In the formula, Let be the minimum travel cost of the main vehicle in the candidate path set at time t. Let be the set of candidate routes for the main vehicle at time t;

[0086] Based on this, the optimal path cost at the current moment is normalized into a time-margin index, as shown in equation (22):

[0087] (twenty two)

[0088] In the formula, This is the edge index at time t; the larger the value, the more edge-like the scene is at that time. The reference cost for free passage represents the path cost when there is no significant risk field impact within the interaction area. The reference cost for blocked passage represents the upper bound of the path cost when the interaction region is in a high-risk blocked state. For normalized stable terms;

[0089] The entire temporal scene is aggregated to obtain the overall edge score of the scene, as shown in Equation (23):

[0090] (twenty three)

[0091] In the formula, The overall edge score for the entire temporal scene. This represents the number of discrete moments corresponding to the total duration of the time-series scene. This is the balance coefficient between peak marginality and average marginality, and its value is between 0 and 1.

[0092] The beneficial effects of this invention are:

[0093] This invention provides a method for quantitatively evaluating the edge characteristics of complex scenarios based on a multi-category risk field model. It differentiates and models multiple risk sources, including unidentified, low-confidence identified, and high-confidence abnormal behaviors, and superimposes them to form a composite risk field. Furthermore, it achieves accurate quantification of the edge characteristics of complex scenarios based on the minimum travel cost of the main vehicle's path. This invention solves the problem that existing methods struggle to uniformly characterize multiple risk categories and overall travel costs, providing support for risk assessment of autonomous vehicles, edge scenario discovery, and performance testing improvements. Attached Figure Description

[0094] Figure 1 A schematic diagram of a composite risk field for multiple categories of scenario elements;

[0095] In the diagram: 1. Risk field of unidentified scene elements; 2. Risk field of low-confidence identified scene elements; 3. Risk field of high-confidence identified scene elements; 4. Main vehicle; 5. Driving path.

[0096] Figure 2 This is a flowchart for edge quantization of scenarios based on minimum passage cost. Detailed Implementation

[0097] This embodiment provides a method for quantitatively evaluating the marginality of complex scenarios based on a multi-category risk field model, including the following steps:

[0098] Step 1: Analyze and classify the components of the time-series scene;

[0099] Based on the PEGASUS system, the elements of time-series scenes are deeply analyzed. The elements of scenes are expressed in a structured manner according to the road layer, road facilities and rules layer, temporary change layer, dynamic object layer, environmental condition layer and digital information layer. Visual modal and trajectory modal information are integrated to form a scene element representation for complex scenes. The scene elements are classified into unidentified scene elements, low-confidence identified scene elements and high-confidence identified scene elements.

[0100] Step 2: Construct a large-scale risk field for unidentified scene elements;

[0101] For unidentified scene elements that are not recognized by the autonomous driving system but have a significant impact on the driver vehicle due to obstruction, unknown obstacles, or abnormal occupation of the area in front of the driver vehicle, an impact score for unidentified scene elements is constructed, and a risk field for unidentified scene elements is established, assigning a general field range and a high field strength weight.

[0102] Step 3: Construct a medium-range risk field for low-confidence identification scene elements;

[0103] For the low-confidence recognition scene elements that have been detected, the field strength amplitude and potential field range of the low-confidence recognition scene elements are constructed. Based on the recognition probability and the interaction relationship with the main vehicle, a risk field of the low-confidence recognition scene elements is constructed, and a medium-range risk field is assigned to achieve quantitative modeling of uncertain risks.

[0104] Step 4: Construct an anisotropic driving behavior risk field for high-confidence scene elements;

[0105] For high-confidence scene elements that have been detected, the field strength amplitude and potential field range of the high-confidence scene elements are constructed based on their speed, acceleration, heading changes and relative proximity trends. A driving behavior risk field of the high-confidence scene elements is constructed. When the high-confidence scene elements have abnormal and dangerous behaviors, a large field strength and potential field range are assigned, and when the high-confidence scene elements drive safely, a small field strength and potential field range are assigned.

[0106] Step 5: Overlay multiple risk fields and quantify the edge of the scenario;

[0107] Within the vehicle interaction area, the risk fields of multiple types of traffic participants are superimposed to form a composite field strength. The marginality of the scenario is measured based on the minimum passage cost when the main vehicle selects the optimal path. A higher minimum passage cost indicates higher marginality, and vice versa.

[0108] Furthermore, the specific method of step one described in this embodiment is as follows:

[0109] 11) Deeply analyze the components of a time-series scene and express them in a structured manner;

[0110] This invention performs hierarchical analysis of complex scenes based on the PEGASUS system. The PEGASUS system includes a road layer, a road facility and rule layer, a temporary change layer, a dynamic object layer, an environmental condition layer, and a digital information layer. Visual modality and trajectory modality information are synchronously written into the state description, and the scene state representation at time t is shown in Equation (1):

[0111] (1)

[0112] In the formula, Let t be the scene state at time t. This includes temporal information for the road layer, such as road geometry, number of lanes, and road boundaries. This provides time-series information on road infrastructure and regulations, including lane markings, traffic signs, traffic signals, and traffic rules. This includes time-series information for temporary changes, such as construction areas, temporary obstacles, debris, and temporary road closures. This refers to the temporal information of dynamic objects, including vehicles, pedestrians, non-motorized vehicles, and other moving targets within the scene. This provides time-series information on environmental conditions, including weather, lighting, and visibility. This refers to time-series information at the digital information layer, including high-precision maps and vehicle-road cooperative information. For visual modal temporal information, For trajectory mode temporal information;

[0113] 12) Perform fusion and classification of scene elements in both visual and trajectory modes;

[0114] For the i-th scene element in the scene, the present invention extracts its visual features and trajectory features and performs linear fusion to obtain a bimodal element representation, as shown in equation (2):

[0115] (2)

[0116] In the formula, Let be the bimodal fusion feature of the i-th scene element at time t. Let be the visual feature vector of the i-th scene element. Let be the trajectory feature vector of the i-th scene element. and These are the visual feature mapping matrix and the trajectory feature mapping matrix, respectively. For fusion bias term;

[0117] After obtaining the bimodal feature representation, the feature category probability distribution is output by the classifier, as shown in Equation (3):

[0118] (3)

[0119] In the formula, Let be the category probability distribution of the i-th scene element. For normalized classification functions, The classifier weight matrix is... For classifier bias terms;

[0120] To classify objects into different risk categories, this invention employs a classification strategy based on confidence thresholds, as shown in equations (4)(5)(6):

[0121] (4)

[0122] (5)

[0123] (6)

[0124] In the formula, Let be the category label for the i-th scene element. This indicates taking the category label that maximizes the category probability. Let t be the set of low-confidence scene elements identified at time t. Let be the set of scene elements for high-confidence recognition at time t. This represents the maximum value among the probabilities of the i-th scene element category. For low confidence threshold, For high confidence threshold, and Greater than .

[0125] Furthermore, the specific method for step two in this embodiment is as follows:

[0126] 21) Construct a score to assess the impact of unidentified scene elements;

[0127] The unidentified scene elements refer to areas that are not recognized by the autonomous driving system but may affect the driver vehicle due to occlusion, unknown obstacles, or abnormal occupation of the area in front of the driver vehicle. The present invention constructs an impact score for unidentified scene elements, as shown in Equation (7):

[0128] (7)

[0129] In the formula, The impact score for the j-th unidentified scene element at time t, with a value between 0 and 1. , and These are the weighted coefficients corresponding to the degree of occlusion, abnormal movement, and relevance to the interaction with the main vehicle, respectively. Let represent the occlusion degree of the j-th unidentified scene element. Let be the abnormal motion quantity of the j-th unidentified scene element, used to measure whether there is motion information within this unidentified scene element that is difficult to directly identify but continuously affects the driver's decision. Let be the main vehicle interaction relevance of the j-th unidentified scene element, representing the degree of coupling between the unidentified scene element and the main vehicle path;

[0130] 22) Establish a risk field for unidentified scene elements;

[0131] The field strength amplitude of unidentified scene elements is determined based on the impact score of unidentified scene elements. As shown in equation (8):

[0132] (8)

[0133] In the formula, The base field strength representing unidentified scene elements. This represents the adjustment coefficient for the impact score;

[0134] The potential field range of unidentified scene elements is determined based on the influence score of the unidentified scene elements, and larger radius of influence is given priority, as shown in equation (9):

[0135] (9)

[0136] In the formula, Let j be the potential field range of the j-th unidentified scene element. The radius of the basic potential field for unidentified scene elements. This is the potential field range adjustment coefficient;

[0137] A Gaussian risk field is established for the unidentified scene elements, as shown in Equation (10):

[0138] (10)

[0139] In the formula, Let be the risk field value generated by the j-th unidentified scene feature at the scene coordinate point (x, y). and These are the center coordinates of the j-th unidentified scene element. and These represent the field diffusion scales of the j-th unidentified scene element in the horizontal and vertical directions, respectively. , , , These represent the horizontal and vertical diffusion ratios, respectively.

[0140] Furthermore, the specific method for step three in this embodiment is as follows:

[0141] 31) Construct the field strength amplitude and potential field range of low-confidence identification scene elements;

[0142] The low-confidence recognition scene element refers to an uncertain scene element that is detected but whose recognition probability is lower than a certain threshold. The present invention first extracts its maximum recognition probability and determines that it belongs to the low-confidence recognition object based on the threshold, as shown in equation (11):

[0143] (11)

[0144] In the formula, Let be the maximum recognition probability of the i-th recognized scene element at time t. Let be the category probability distribution of the i-th scene element. For low confidence threshold, For high confidence threshold, and Greater than ;

[0145] The field strength amplitude of low-confidence scene elements is determined based on the maximum recognition probability of the low-confidence scene elements. As shown in equation (12):

[0146] (12)

[0147] In the formula, The basic field strength representing elements in a low-confidence identification scenario. The coefficient represents the adjustment factor for uncertainty.

[0148] The potential field range of low-confidence recognition scene elements is determined based on the recognition probability, as shown in equation (13):

[0149] (13)

[0150] In the formula, Let be the potential field range of the i-th low-confidence scene element. The radius of the basic potential field for elements in low-confidence scenarios. This is the confidence level inverse adjustment coefficient;

[0151] 32) Construct a risk field for low-confidence identification scene elements;

[0152] The risk field for low-confidence scene elements is shown in equation (14):

[0153] (14)

[0154] In the formula, Let be the risk field value generated by the i-th low-confidence identified scene feature at the scene coordinate point (x, y). Let be the field strength amplitude of the i-th low-confidence scene element. and These are the location coordinates of the i-th low-confidence scene element. and , respectively, represent the field diffusion scales of the i-th low-confidence scene element in the horizontal and vertical directions. and , , These represent the horizontal and vertical diffusion ratios, respectively;

[0155] The lower the maximum recognition probability of scene elements in low-confidence recognition, the better. The larger the value, the larger the potential field range of the low-confidence scene elements. However, the overall potential field range of the low-confidence scene elements is smaller than that of the unidentified scene elements.

[0156] Furthermore, the specific method for step four in this embodiment is as follows:

[0157] 41) Construct the field strength amplitude and potential field range of scene elements for high-confidence identification;

[0158] The high-confidence recognition scene element refers to a scene element that has been detected and whose recognition probability is higher than a certain threshold. This invention combines the driving behavior of the high-confidence recognition scene element to establish a risk field and increases the risk level when the element has abnormal and dangerous behavior.

[0159] First, construct the behavioral anomaly index, as shown in equation (15):

[0160] (15)

[0161] In the formula, For time t, the first A high-confidence indicator for identifying the behavioral anomalies of scene elements. , , , and These are the weighting coefficients for longitudinal acceleration, angular velocity, relative approach speed, collision time, and lane behavior, respectively. For the first Longitudinal acceleration of each traffic participant The normalized constant for longitudinal acceleration, For the first The heading angular velocity of each traffic participant Let be the normalized constant for the heading angular velocity. For the first The relative approach speed of each traffic participant to the host vehicle. Let be the velocity normalization constant. For the first The collision time between a traffic participant and the main vehicle Let be the collision time normalization constant. For the first The abnormal lane behavior indicator value for each traffic participant increases when the element exhibits dangerous cutting in, continuous lane crossing, abnormal lane changing or other high-risk behaviors.

[0162] The field strength amplitude and potential field range of the high-confidence recognition scene elements are determined based on the degree of behavioral anomaly, as shown in Equation (16):

[0163] (16)

[0164] In the formula, For the first The field strength amplitude of a high-confidence scene element. The basic field strength for high-confidence identification of scene elements. For the first The potential field range of a high-confidence recognition scene element. The radius of the basic potential field for high-confidence identification of scene elements. and These are the adjustment coefficients for the anomaly index;

[0165] 42) Construct a high-confidence recognition scenario element driving behavior risk field;

[0166] To reflect the non-uniform influence of the vehicle along the heading direction, this invention adopts an anisotropic driving behavior risk field model in a local coordinate system, as shown in equations (17) and (18):

[0167] (17)

[0168] (18)

[0169] In the formula, For the first A high-confidence identification of the risk field value generated by scene elements at scene coordinate point (x,y). and These represent the vertical and horizontal coordinates of the coordinate point (x, y) relative to the feature's position in the feature's local coordinate system. and These represent the field diffusion scales of scene elements in the vertical and horizontal directions, respectively, for high-confidence recognition. and , , These represent the longitudinal and lateral diffusion ratios, respectively. For heading angle Constructed two-dimensional rotation matrix, For the first The heading angle for identifying scene elements with high confidence. For the first The field strength amplitude of a high-confidence scene element. For the first The potential field range of a high-confidence recognition scene element. Represents the transpose operator. , Representing the first High-confidence identification elements in Always in the scene's global coordinate system , The position coordinates of the direction.

[0170] Furthermore, such as Figure 1-2 As shown, Figure 1The data displays the risk field 1 for unidentified scene elements, the risk field 2 for low-confidence identified scene elements, the risk field 3 for high-confidence identified scene elements, and the location and driving path of the main vehicle 4.

[0171] The specific method for step five in this embodiment is as follows:

[0172] 51) Construct the composite field strength of the interaction region;

[0173] Within the designated interaction area of ​​the main vehicle, this invention superimposes the risk fields generated by the behaviors of unidentified scene elements, low-confidence identified scene elements, and high-confidence identified scene elements to form a composite field strength, as shown in equation (19):

[0174] (19)

[0175] In the formula, Let be the combined field strength value at the scene coordinate point (x, y) at time t. Let t be the set of unidentified scene elements. Let t be the set of low-confidence scene elements identified at time t. Let t be the set of scene elements for high-confidence identification. Let be the risk field value generated by the j-th unidentified scene feature at the scene coordinate point (x, y). Let be the risk field value generated by the i-th low-confidence identified scene feature at the scene coordinate point (x, y). For the first A high-confidence identification of the risk field value generated by scene elements at scene coordinate point (x,y);

[0176] 52) Quantify scenario edge cases based on the minimum passage cost of the main vehicle's travel path;

[0177] Let γ be the candidate path of the main vehicle in the interaction area at time t. In this invention, the cumulative travel cost of the path in the composite field strength is used as the path evaluation index, as shown in equation (20):

[0178] (20)

[0179] In the formula, Let be the travel cost of candidate path γ at time t. Let γ be the position of the candidate path γ under path parameter s. Let γ be the first derivative of the candidate path with respect to the path parameter s. The Euclidean norm is two-dimensional; μ is the field strength cost weighting coefficient. express Time Scene Coordinates The combined field strength at that location, Indicates the composite field strength Along candidate path Values, path parameters Obtained by geometric parameterization of the vehicle's travel path;

[0180] The minimum travel cost is calculated for all candidate paths to obtain the optimal path cost for the main vehicle, as shown in equation (21):

[0181] (twenty one)

[0182] In the formula, Let be the minimum travel cost of the main vehicle in the candidate path set at time t. Let be the set of candidate routes for the main vehicle at time t;

[0183] Based on this, the optimal path cost at the current moment is normalized into a time-margin index, as shown in equation (22):

[0184] (twenty two)

[0185] In the formula, This is the edge index at time t; the larger the value, the more edge-like the scene is at that time. The reference cost for free passage represents the path cost when there is no significant risk field impact within the interaction area. The reference cost for blocked passage represents the upper bound of the path cost when the interaction region is in a high-risk blocked state. For normalized stable terms;

[0186] The entire temporal scene is aggregated to obtain the overall edge score of the scene, as shown in Equation (23):

[0187] (twenty three)

[0188] In the formula, The overall edge score for the entire temporal scene. This represents the number of discrete moments corresponding to the total duration of the time-series scene. This is the balance coefficient between peak marginality and average marginality, and its value is between 0 and 1.

Claims

1. A method for quantitatively evaluating the marginality of complex scenarios based on a multi-category risk field model, characterized in that: Includes the following steps: Step 1: Based on the PEGASUS system, deeply analyze the components of the time-series scene, and express the components of the scene in a structured manner according to the road layer, road facilities and rules layer, temporary change layer, dynamic object layer, environmental condition layer and digital information layer. Integrate visual modal and trajectory modal information to form a scene element representation for complex scenes, and classify the scene elements into unrecognized scene elements, low-confidence recognized scene elements and high-confidence recognized scene elements. Step 2: For unidentified scene elements, construct an impact score for unidentified scene elements and establish a risk field for unidentified scene elements, assigning a range of the general trend field and a high field strength weight. Step 3: For low-confidence recognition scene elements, construct the field strength amplitude and potential field range of the low-confidence recognition scene elements. Based on the recognition probability and the interaction relationship with the main vehicle, construct the risk field of the low-confidence recognition scene elements and assign a medium-range risk field to achieve quantitative modeling of uncertain risks. Step 4: For high-confidence scene elements, construct the field strength amplitude and potential field range of the high-confidence scene elements based on their speed, acceleration, heading changes and relative proximity trends, and construct the driving behavior risk field of the high-confidence scene elements. When the high-confidence scene elements have abnormal and dangerous behaviors, assign a large field strength and potential field range, and when the high-confidence scene elements drive safely, assign a small field strength and potential field range. Step 5: Within the vehicle interaction area, the risk fields of multiple types of traffic participants are superimposed to form a composite field strength. The marginality of the scenario is measured based on the minimum passage cost when the main vehicle selects the optimal path. A higher minimum passage cost indicates higher marginality, and vice versa.

2. The method for quantitative evaluation of the marginality of complex scenarios based on a multi-category risk field model according to claim 1, characterized in that: In step one, the method for deeply analyzing the components of a time-series scene and expressing them in a structured manner is as follows: Based on the PEGASUS system, a hierarchical analysis of complex scenes is performed. The road layer, road facilities and rules layer, temporary change layer, dynamic object layer, environmental condition layer and digital information layer are uniformly mapped to the temporal scene state. Visual modality and trajectory modality information are synchronously written into the state description. The scene state representation at time t is shown in Equation (1): (1) In the formula, Let t be the scene state at time t. This refers to the time-series information of the road layer. For road infrastructure and rule layer time sequence information, Temporary change layer timing information, For dynamic object layer timing information, This refers to the temporal information of the environmental conditions layer. For digital information layer time sequence information, For visual modal temporal information, This refers to the trajectory mode temporal information.

3. A method for quantitatively evaluating the marginality of complex scenarios based on a multi-category risk field model, as described in claim 1 or 2, characterized in that: In step one, the method for fusing and classifying scene elements in both visual and trajectory modes is as follows: For the i-th scene element in the scene, its visual features and trajectory features are extracted and linearly fused to obtain a bimodal element representation, as shown in Equation (2): (2) In the formula, Let be the bimodal fusion feature of the i-th scene element at time t. Let be the visual feature vector of the i-th scene element. Let be the trajectory feature vector of the i-th scene element. and These are the visual feature mapping matrix and the trajectory feature mapping matrix, respectively. For fusion bias terms; After obtaining the bimodal feature representation, the feature category probability distribution is output by the classifier, as shown in Equation (3): (3) In the formula, Let be the category probability distribution of the i-th scene element. For normalized classification functions, The classifier weight matrix is... For classifier bias terms; To classify objects into different risk categories, a classification strategy based on confidence thresholds is adopted, as shown in equations (4)(5)(6): (4) (5) (6) In the formula, Let be the category label for the i-th scene element. This indicates taking the category label that maximizes the category probability. Let t be the set of low-confidence scene elements identified at time t. Let be the set of scene elements for high-confidence recognition at time t. This represents the maximum value among the probabilities of the i-th scene element category. For low confidence threshold, For high confidence threshold, and Greater than .

4. The method for quantitative evaluation of the marginality of complex scenarios based on a multi-category risk field model according to claim 1, characterized in that: In step two, the method for constructing the score for the impact of unidentified scene elements is as follows: The unidentified scene elements refer to areas that are not recognized by the autonomous driving system but may affect the driver vehicle due to occlusion, unknown obstacles, or abnormal occupation of the area in front of the driver vehicle. An impact score for unidentified scene elements is constructed, as shown in Equation (7): (7) In the formula, The impact score for the j-th unidentified scene element at time t, with a value between 0 and 1. , and These are the weighted coefficients corresponding to the degree of occlusion, abnormal movement, and relevance to the interaction with the main vehicle, respectively. Let represent the occlusion degree of the j-th unidentified scene element. The abnormal motion quantity of the j-th unidentified scene element is used to measure whether there is motion information within this unidentified scene element that is difficult to directly identify but continuously affects the driver's decision. Let be the main vehicle interaction relevance of the j-th unidentified scene element, representing the degree of coupling between the unidentified scene element and the main vehicle path.

5. A method for quantitatively evaluating the marginality of complex scenarios based on a multi-category risk field model, as described in claim 1 or 4, characterized in that: In step two, the method for establishing the risk field of unidentified scene elements is as follows: The field strength amplitude of unidentified scene elements is determined based on the impact score of unidentified scene elements. As shown in equation (8): (8) In the formula, The base field strength representing unidentified scene elements. The representative factor is the adjustment coefficient for the impact score. The impact score for the j-th unidentified scene element at time t; The potential field range of unidentified scene elements is determined based on the influence score of the unidentified scene elements, and larger radius of influence is given priority, as shown in equation (9): (9) In the formula, Let j be the potential field range of the j-th unidentified scene element. The radius of the basic potential field for unidentified scene elements. This is the potential field range adjustment coefficient; A Gaussian risk field is established for the unidentified scene elements, as shown in Equation (10): (10) In the formula, Let be the risk field value generated by the j-th unidentified scene feature at the scene coordinate point (x, y). and These are the center coordinates of the j-th unidentified scene element. and These represent the field diffusion scales of the j-th unidentified scene element in the horizontal and vertical directions, respectively. , , , These represent the horizontal and vertical diffusion ratio coefficients, respectively.

6. The method for quantitative evaluation of the marginality of complex scenarios based on a multi-category risk field model according to claim 1, characterized in that: In step three, the method for constructing the field strength amplitude and potential field range of low-confidence scene elements is as follows: The low-confidence recognition scene element refers to an uncertain scene element that is detected but whose recognition probability is lower than a certain threshold. First, its maximum recognition probability is extracted, and it is determined to belong to the low-confidence recognition object according to the threshold, as shown in Equation (11): (11) In the formula, Let be the maximum recognition probability of the i-th recognized scene element at time t. Let be the category probability distribution of the i-th scene element. For low confidence threshold, For high confidence threshold, and Greater than ; The field strength amplitude of low-confidence scene elements is determined based on the maximum recognition probability of the low-confidence scene elements. As shown in equation (12): (12) In the formula, The basic field strength representing elements in a low-confidence identification scenario. The coefficient represents the adjustment factor for uncertainty. The potential field range of low-confidence recognition scene elements is determined based on the recognition probability, as shown in equation (13): (13) In the formula, Let be the potential field range of the i-th low-confidence scene element. The radius of the basic potential field for elements in low-confidence scenarios. This is the confidence level inverse adjustment coefficient.

7. A method for quantitatively evaluating the marginality of complex scenarios based on a multi-category risk field model, as described in claim 1 or 6, characterized in that: In step three, the method for constructing the risk field of low-confidence identification scene elements is as follows: The risk field for low-confidence scene elements is shown in equation (14): (14) In the formula, Let be the risk field value generated by the i-th low-confidence identified scene feature at the scene coordinate point (x, y). Let be the field strength amplitude of the i-th low-confidence scene element. and These are the location coordinates of the i-th low-confidence scene element. and , respectively, represent the field diffusion scales of the i-th low-confidence scene element in the horizontal and vertical directions. and , , These represent the horizontal and vertical diffusion ratios, respectively. Let be the potential field range of the i-th low-confidence scene element; The lower the maximum recognition probability of scene elements in low-confidence recognition, the better. The larger the value, the larger the potential field range of the low-confidence scene elements. However, the overall potential field range of the low-confidence scene elements is smaller than that of the unidentified scene elements.

8. The method for quantitative evaluation of the marginality of complex scenarios based on a multi-category risk field model according to claim 1, characterized in that: In step four, the method for constructing the field strength amplitude and potential field range of high-confidence scene elements is as follows: The high-confidence recognition scene element refers to a scene element that has been detected and whose recognition probability is higher than a certain threshold. A risk field is established by combining the driving behavior of the high-confidence recognition scene element, and the risk level is increased when the element has abnormal and dangerous behavior. The behavioral anomaly index is constructed as shown in equation (15): (15) In the formula, For time t, the first A high-confidence indicator for identifying the behavioral anomalies of scene elements. , , , and These are the weighting coefficients for longitudinal acceleration, angular velocity, relative approach speed, collision time, and lane behavior, respectively. For the first The longitudinal acceleration of each traffic participant The normalized constant for longitudinal acceleration, For the first The heading angular velocity of each traffic participant Let be the normalized constant for the heading angular velocity. For the first The relative approach speed of each traffic participant to the host vehicle. Let be the velocity normalization constant. For the first The collision time between a traffic participant and the main vehicle Let be the collision time normalization constant. For the first The abnormal lane behavior indicator value for each traffic participant increases when the element exhibits dangerous cutting in, continuous lane crossing, abnormal lane changing or other high-risk behaviors. The field strength amplitude and potential field range of the high-confidence recognition scene elements are determined based on the degree of behavioral anomaly, as shown in Equation (16): (16) In the formula, For the first The field strength amplitude of a high-confidence scene element. The basic field strength for high-confidence identification of scene elements. For the first The potential field range of a high-confidence recognition scene element. The radius of the basic potential field for high-confidence identification of scene elements. and These are the adjustment coefficients for the anomaly index.

9. A method for quantitatively evaluating the marginality of complex scenarios based on a multi-category risk field model, as described in claim 1 or 8, characterized in that: In step four, the method for constructing a high-confidence recognition scene element driving behavior risk field is as follows: An anisotropic driving behavior risk field model in a local coordinate system is adopted, as shown in equations (17) and (18): (17) (18) In the formula, For the first A high-confidence identification of the risk field value generated by scene elements at scene coordinate point (x,y). and These represent the vertical and horizontal coordinates of the coordinate point (x, y) relative to the feature's position in the feature's local coordinate system. and These represent the field diffusion scales of scene elements in the vertical and horizontal directions, respectively, for high-confidence recognition. and , , These represent the longitudinal and lateral diffusion ratios, respectively. For heading angle Constructed two-dimensional rotation matrix, For the first The heading angle for identifying scene elements with high confidence. For the first The field strength amplitude of a high-confidence scene element. For the first The potential field range of a high-confidence recognition scene element. Represents the transpose operator. , Representing the first High-confidence identification elements in Always in the scene's global coordinate system , The position coordinates of the direction.

10. The method for quantitative evaluation of the marginality of complex scenarios based on a multi-category risk field model according to claim 1, characterized in that: The method for step five is as follows: First, construct the composite field strength of the interactive area: Within the designated interaction area of ​​the main vehicle, the risk fields generated by the behaviors of unrecognized scene elements, low-confidence recognized scene elements, and high-confidence recognized scene elements are superimposed to form a composite field strength, as shown in equation (19): (19) In the formula, Let be the combined field strength value at the scene coordinate point (x, y) at time t. Let t be the set of unidentified scene elements. Let t be the set of low-confidence scene elements identified at time t. Let t be the set of scene elements identified with high confidence. Let be the risk field value generated by the j-th unidentified scene feature at the scene coordinate point (x, y). Let be the risk field value generated by the i-th low-confidence identified scene feature at the scene coordinate point (x, y). For the first A high-confidence identification of the risk field value generated by scene elements at scene coordinate point (x,y); Secondly, the edge case of a scenario is quantified based on the minimum passage cost of the main vehicle's travel path: Let γ be the candidate path of the main vehicle in the interaction area at time t. The cumulative travel cost of the path in the composite field strength is used as the path evaluation index, as shown in Equation (20): (20) In the formula, Let be the travel cost of candidate path γ at time t. Let γ be the position of the candidate path γ under path parameter s. Let γ be the first derivative of the candidate path with respect to the path parameter s. The Euclidean norm is two-dimensional; μ is the field strength cost weighting coefficient. express Time Scene Coordinates The combined field strength at that location, Indicates the composite field strength Along candidate path Values, path parameters Obtained by geometric parameterization of the vehicle's travel path; The minimum travel cost is calculated for all candidate paths to obtain the optimal path cost for the main vehicle, as shown in equation (21): (21) In the formula, Let be the minimum travel cost of the main vehicle in the candidate path set at time t. Let be the set of candidate routes for the main vehicle at time t; Based on this, the optimal path cost at the current moment is normalized into a time-margin index, as shown in equation (22): (22) In the formula, This is the edge index at time t; the larger the value, the more edge-like the scene is at that time. The reference cost for free passage represents the path cost when there is no significant risk field impact within the interaction area. The reference cost for blocked passage represents the upper bound of the path cost when the interaction region is in a high-risk blocked state. For normalized stable terms; The entire temporal scene is aggregated to obtain the overall edge score of the scene, as shown in Equation (23): (23) In the formula, The overall edge score for the entire temporal scene. This represents the number of discrete moments corresponding to the total duration of the time-series scene. This is the balance coefficient between peak marginality and average marginality, and its value is between 0 and 1.

Images