A human-vehicle collision risk analysis method based on autonomous driving vehicles

Abstract

The present invention discloses a method for analyzing the risk of human-vehicle collision based on an autonomous driving vehicle. First, according to the crossing characteristics in the pedestrian crossing history data set, the crossing behaviors are divided into different habit types through Gaussian clustering, and the joint probability distribution function is used to help the autonomous driving vehicle obtain the possible future state set of the pedestrian according to the current motion state of the pedestrian. The dynamic spatiotemporal relationship between the pedestrian and the vehicle is considered through the GCN graph convolutional neural network to obtain the possible trajectory set of the pedestrian in the future. Finally, considering the trajectory collision probability and the minimum encounter distance, the feature dimension reduction is performed using the matter-element extension theory, and the risk function is established to realize the real-time judgment of the risk of human-vehicle collision. This method focuses on the characteristics of sudden changes in the motion state of pedestrians, and more specifically considers the impact of the spatiotemporal relationship between the pedestrian and the vehicle on the trajectory. The accuracy and credibility of the human-vehicle collision judgment results are improved through the fusion of multiple indicators, further improving the intelligence of autonomous driving, and improving riding comfort and safety.

Description

Technical Field

[0001] The present invention belongs to the field of autonomous driving decision-making algorithms, and in particular relates to a method for analyzing the risk of collision between an autonomous driving vehicle and pedestrians crossing the street. Background Art

[0002] The rapid development of science and technology is pushing the whole society towards a more intelligent direction. In recent years, the technology of autonomous driving vehicles has developed rapidly, and companies such as Google, Tesla, and Baidu have been making continuous attempts in this field. Autonomous driving is getting closer and closer to people's lives, but at the same time, people still have great doubts about driving safety. Human-vehicle conflict identification is a core module in autonomous driving technology. Improving the accuracy of human-vehicle conflict assessment and reducing computational complexity will more effectively ensure the safety of autonomous driving vehicles during driving and enhance accident prevention and handling capabilities.

[0003] At present, most studies on pedestrian-vehicle conflict warning follow the method of vehicle conflict warning and adopt a deterministic model, that is, it is assumed that pedestrians and vehicles continue to move according to the current motion state, and the collision risk is quantified by indicators such as collision time. However, pedestrians are different from motor vehicles in that their speed and direction can change suddenly, especially in the game with vehicles. This phenomenon is particularly obvious. The deterministic model cannot most realistically reflect the true state of the game between pedestrians and vehicles. It is more reasonable to judge pedestrian-vehicle conflicts based on the possible future trajectories of pedestrians. On the other hand, since the speed of pedestrians can change greatly, there are many possible trajectories, speeds and directions. However, the current pedestrian-vehicle conflict judgment methods that consider pedestrian trajectories do not consider the multiple trajectory sets that pedestrians may take in the trajectory prediction stage. These inaccurate trajectory prediction methods will further affect the judgment of pedestrian-vehicle conflicts. Summary of the invention

[0004] The purpose of the present invention is to provide a method for determining the risk of a human-vehicle collision based on an autonomous driving vehicle, which can avoid the low accuracy of collision conflict prediction caused by the traditional single-mode prediction of pedestrian trajectories and the inability to predict possible sudden changes in crossing behavior. The method considers the future trajectory set of pedestrians, constructs an interactive graph network of pedestrians and vehicles, improves the accuracy of collision risk prediction, enriches the prediction scenarios, and provides a highly reliable and accurate human-vehicle collision risk determination result, providing an accurate data basis for the decision-making module and path planning module of the autonomous driving vehicle, further improving the intelligence of autonomous driving, enhancing ride comfort and safety, and promoting the advancement of autonomous driving technology.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] A method for analyzing the risk of a collision between a driver and a vehicle based on an autonomous driving vehicle comprises the following steps:

[0007] Step 1: Obtain the pedestrian crossing characteristics in the pedestrian crossing history dataset, including speed, acceleration and crossing heading angle, and use the Gaussian mixture clustering GMM method to construct a Gaussian mixture distribution function to represent three different pedestrian crossing habits;

[0008] Step 2: Obtain the motion parameters of the pedestrian in front of the autonomous driving vehicle, and match the corresponding pedestrian crossing habits determined in step 1 based on the motion characteristics of the pedestrian;

[0009] Step 3: Taking the pedestrian’s real speed, acceleration and heading angle as the center and taking positive and negative σ as the value interval, obtain all possible speed, acceleration and heading angle sets and joint probabilities in the joint Gaussian distribution, and obtain n by uniform resampling. 3 A set of group state vectors;

[0010] Step 4: Preprocess the motion state data of pedestrians crossing the street acquired by the on-board sensor, the motion state data of the autonomous driving vehicle itself, and the spatial position relationship between pedestrians and vehicles, import the preprocessed data into the graph convolutional neural network GCN model, and train the structural weights and bias parameters of the GCN model;

[0011] Step 5: Use the trained GCN model to predict the trajectory set of pedestrians under different data combinations, and determine the future trajectory of the vehicle while keeping the motion state unchanged;

[0012] Step 6: Determine the collision situation and minimum encounter distance between the pedestrian's future trajectory set and the vehicle's future trajectory;

[0013] Step 7: Use the matter-element extension theory to comprehensively consider the collision situation and the minimum encounter distance to obtain the risk of human-vehicle collision.

[0014] Compared with the prior art, the present invention has the following significant advantages:

[0015] (1) The technical solution of the present invention takes into account the characteristic that the pedestrian's speed direction may change suddenly during the trajectory prediction stage, constructs a Gaussian mixture distribution function to represent three different pedestrian crossing habits, so that the method of the present invention can obtain a more realistic set of pedestrian future trajectories;

[0016] (2) The technical solution of the present invention adopts the GCN graph convolutional neural network to predict pedestrian trajectories. Different from the prior art, it specifically considers the spatial relationship between vehicles and pedestrians and establishes a graph network composed of pedestrians and vehicles, so that the method can extract more accurate spatiotemporal features and thus obtain more accurate future trajectories. In addition, the method can also predict the trajectories of multiple pedestrians and vehicles at the same time.

[0017] (3) The technical solution of the present invention comprehensively considers the collision risk and the shortest distance risk of the two in the conflict risk estimation method, and obtains a one-dimensional risk variable through the matter-element extension algorithm, which is more specific and accurate in estimating the conflict than the existing technology.

[0018] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The present invention is a flowchart of the steps of the method for analyzing the risk of human-vehicle collision based on an autonomous driving vehicle.

[0020] Figure 2 Schematic diagram of the relationship between pedestrians and vehicles in an embodiment of the present invention.

[0021] Figure 3 It is a schematic diagram of a two-dimensional extension set in an embodiment of the present invention.

[0022] Figure 4 Schematic diagram of a one-dimensional hazard function in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] A method for analyzing the risk of a collision between a driver and a vehicle based on an autonomous driving vehicle comprises the following steps:

[0024] Step 1: Obtain the pedestrian crossing characteristics from the historical data set, including speed, acceleration, and crossing heading angle, and use the Gaussian mixture clustering GMM method to construct a Gaussian mixture distribution function to represent three different pedestrian crossing habits, specifically:

[0025] The motion parameters of pedestrians crossing the street in the presence of a human-vehicle game are obtained through the existing data set, and the crossing characteristics of pedestrians in the historical data set are extracted, including speed, acceleration and crossing heading angle.

[0026] The pedestrian trajectory heading angle refers to: the angle formed by the pedestrian trajectory and the curb of the road. Since China adopts the right-hand driving system, the heading angle is acute when the trajectory is consistent with the prescribed direction of travel, and it is obtuse when the trajectory is opposite to the prescribed direction of travel.

[0027] The vehicle motion state information includes: the vehicle's heading angle θ, the vehicle's travel speed γ, and the vehicle's travel acceleration ω in the past few seconds.

[0028] The vehicle heading angle refers to: the angle between the vehicle trajectory and the curb of the road. Since China adopts the right-hand driving system, the heading angle is acute when the trajectory is consistent with the prescribed direction of travel, and it is obtuse when the trajectory is opposite to the prescribed direction of travel.

[0029] The spatial relationship information between the pedestrian and the vehicle includes: the absolute position of the autonomous driving vehicle and the absolute position of the pedestrian.

[0030] Step 1-1, construct probability density function:

[0031]

[0032] Where μ represents the 3D mean vector, Σ represents the 3*3 covariance matrix determined by the pedestrian's heading angle, velocity and acceleration, and p(x) represents the probability density function of the random vector χ in the 3D sample space χ that obeys the Gaussian distribution;

[0033] Step 1-2: Determine the Gaussian mixture distribution function based on the three pedestrian crossing behavior habits:

[0034]

[0035] Among them, μ i and∑ i is the parameter of the i-th Gaussian mixture component, k = 3, α i >0 is the corresponding mixing coefficient,

[0036] Step 1-3, use the EM algorithm to obtain the optimal Gaussian distribution parameters after iteration. The EM algorithm optimization solution means that in each iteration, the posterior probability of each sample belonging to each Gaussian component is calculated according to the current parameters (E step), and then the model parameters are updated by maximum likelihood estimation and Laarrange multiplier method (M step), specifically:

[0037] Step 1-3-1, initialize Gaussian parameters μ i ,∑ i and α i , for each sample Z j , determine the posterior probability of belonging to each Gaussian distribution:

[0038]

[0039] Step 1-3-2: Determine each sample x j The cluster label λ j :

[0040]

[0041] Step 1-3-3: Label λ by cluster j Divide the sample set into 3 clusters C = {C1, C2, C3};

[0042] Step 1-3-4: Use the maximum likelihood method to construct the Lagrangian function to update the Gaussian distribution parameters of each cluster: μ i ,∑ i and α i , for m samples and k clusters, the Lagrangian function formula is:

[0043]

[0044] The mean vector μ of each Gaussian distribution is i The update formula is:

[0045]

[0046] The covariance matrix ∑ of each Gaussian distribution is i The update formula is:

[0047]

[0048] The mixing coefficient α of each Gaussian distribution i The update formula is:

[0049]

[0050] Next, end the current iteration and convert the newly obtained μ′ i ,∑′ i and α′ i As the initial parameters for the next iteration, until the set iteration stop requirements are met.

[0051] Step 2: Obtain the motion parameters of the pedestrian in front of the autonomous driving vehicle, and match the corresponding pedestrian crossing habits determined in step 1 based on the motion characteristics of the pedestrian, specifically:

[0052] Step 2-1: Using the laser radar, camera and / or millimeter wave radar sensors carried by the autonomous driving vehicle, the pedestrian's heading angle, historical trajectory, current speed and acceleration parameters are collected in real time through multi-sensor information fusion to form a state vector η;

[0053] The camera carried by the autonomous driving vehicle is used to collect the posture and head movement of the pedestrian, and thereby obtain the pedestrian's heading angle.

[0054] The laser radar device carried by the autonomous driving vehicle is used to collect the position and kinematic parameters of the pedestrian, and thereby obtain the pedestrian's historical trajectory and current speed and acceleration parameters;

[0055] Step 2-2: Substitute the pedestrian’s state vector η into the Gaussian mixture distribution and calculate the posterior probability that the vector η belongs to the three Gaussian distributions, p M (C=1|η), p M (C=2|η), p M (C=3|η), the Gaussian distribution with the largest posterior probability is the Gaussian distribution to which the pedestrian’s current state vector belongs, and the pedestrian’s crossing habit is then determined.

[0056] Step 3: Taking the actual speed, acceleration and heading angle of the current pedestrian as the center and the positive and negative σ as the value interval, obtain all possible speed, acceleration and heading angle sets and joint probabilities in the joint Gaussian distribution, and obtain n3 sets of state vectors by uniform resampling, specifically:

[0057] Taking the three-dimensional state vector η composed of the pedestrian's actual movement speed, acceleration and heading angle as the center and the positive and negative σ as the value interval, obtain the set and joint probability of all possible speeds, accelerations and heading angles in its joint Gaussian distribution;

[0058] The obtained possible speed, acceleration and heading angle set interval and joint probability interval are divided into n parts respectively, and then n parts are determined by cross matching. 3 A set of three-dimensional vectors of possible pedestrian states and the probability of each set.

[0059] Step 4: Preprocess the motion state data of pedestrians crossing the street acquired by the on-board sensor, the motion state data of the autonomous driving vehicle itself, and the spatial position relationship between pedestrians and vehicles, import the preprocessed data into the graph convolutional neural network GCN model, and iterate the structural weights and bias parameters of the GCN model multiple times according to the gradient descent method;

[0060] The graph convolutional neural network GCN model includes a graph structure, a convolutional layer, a pooling layer, and a fully connected layer.

[0061] The graph structure refers to a directed graph composed of a space domain or a vertex domain, and is used to extract the spatial features of a topological graph.

[0062] The convolution layer refers to a network layer composed of several convolution units after convolution kernel transformation, and the parameters of each convolution unit are optimized through the back propagation algorithm.

[0063] The pooling layer divides the input image into several rectangular areas and outputs the maximum value for each sub-area, which is a form of downsampling.

[0064] The fully connected layer means that each node is connected to all nodes in the previous layer.

[0065] Step 5: Use the trained GCN model to predict the trajectory set of pedestrians under different data combinations, and determine the future trajectory of the vehicle while maintaining the same state of motion, specifically:

[0066] Step 5-1: Obtain the speed, acceleration, and position of the current vehicle and pedestrians, including the speed, acceleration, and position of the pedestrians, and the speed, acceleration, and position of the vehicle. After data normalization, obtain the real-time trajectory sequence of pedestrians and vehicles, and combine it with the n obtained in step 3. 3One of the group state vectors is input into the trained GCN model and outputs n 3 The output prediction data is denormalized to obtain n 3 The movement trajectory of pedestrians crossing the street in the next 2 seconds;

[0067] Step 5-2: Determine the trajectory of the vehicle within the next 2 seconds based on vehicle dynamics.

[0068] Step 6: Determine the collision situation and minimum encounter distance between the pedestrian's future trajectory set and the vehicle's future trajectory, specifically:

[0069] Step 6-1, determine the collision situation, consider the real outline of the vehicle and simulate the pedestrian as a circle with a diameter of 0.5 meters, determine the collision situation of each pedestrian's future trajectory with the vehicle's future trajectory, and determine the comprehensive collision probability ρ based on the probability of each trajectory;

[0070] Step 6-2, determining the minimum encounter distance, refers to considering the actual contour of the vehicle and assuming that the pedestrian is a circle with a diameter of 0.5 meters, determining the minimum encounter distance y between each future trajectory of the pedestrian and the future trajectory of the vehicle without a collision, and determining the weighted average minimum encounter distance y based on the probability of each trajectory.

[0071] Step 7: Use the matter-element extension theory to comprehensively consider the collision situation and the minimum encounter distance to obtain the risk of human-vehicle collision, specifically:

[0072] Step 7-1, construct a two-dimensional space consisting of the ρ axis and the y axis, and set the boundary range of the rectangular classical domain Φ1 and the extension domain Φ2 respectively, and determine a point in the two-dimensional space according to the weighted average minimum encounter distance y and the comprehensive collision probability ρ determined in step 6;

[0073] Step 7-2: Based on the points determined in step 7-1, use the extenics theory to calculate the extenics distance to determine the correlation function value. When the correlation is greater than 1, there is no conflict; when the correlation is between 0 and 1, there is a minor conflict and the vehicle should slow down to avoid it; when the correlation is less than 0, there is a major conflict and the vehicle should stop to avoid it.

[0074] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the following steps are implemented:

[0075] Step 1: Obtain the pedestrian crossing characteristics in the pedestrian crossing history dataset, including speed, acceleration and crossing heading angle, and use the Gaussian mixture clustering GMM method to construct a Gaussian mixture distribution function to represent three different pedestrian crossing habits;

[0076] Step 2: Obtain the motion parameters of the pedestrian in front of the autonomous driving vehicle, and match the corresponding pedestrian crossing habits determined in step 1 based on the motion characteristics of the pedestrian;

[0077] Step 3: Taking the pedestrian’s real speed, acceleration and heading angle as the center and taking positive and negative σ as the value interval, obtain all possible speed, acceleration and heading angle sets and joint probabilities in the joint Gaussian distribution, and obtain n by uniform resampling. 3 A set of group state vectors;

[0078] Step 4: Preprocess the motion state data of pedestrians crossing the street acquired by the on-board sensor, the motion state data of the autonomous driving vehicle itself, and the spatial position relationship between pedestrians and vehicles, import the preprocessed data into the graph convolutional neural network GCN model, and train the structural weights and bias parameters of the GCN model;

[0079] Step 5: Use the trained GCN model to predict the trajectory set of pedestrians under different data combinations, and determine the future trajectory of the vehicle while keeping the motion state unchanged;

[0080] Step 6: Determine the collision situation and minimum encounter distance between the pedestrian's future trajectory set and the vehicle's future trajectory;

[0081] Step 7: Use the matter-element extension theory to comprehensively consider the collision situation and the minimum encounter distance to obtain the risk of human-vehicle collision.

[0082] A computer storable medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:

[0083] Step 1: Obtain the pedestrian crossing characteristics in the pedestrian crossing history dataset, including speed, acceleration and crossing heading angle, and use the Gaussian mixture clustering GMM method to construct a Gaussian mixture distribution function to represent three different pedestrian crossing habits;

[0084] Step 2: Obtain the motion parameters of the pedestrian in front of the autonomous driving vehicle, and match the corresponding pedestrian crossing habits determined in step 1 based on the motion characteristics of the pedestrian;

[0085] Step 3: Taking the pedestrian’s real speed, acceleration and heading angle as the center and taking positive and negative σ as the value interval, obtain all possible speed, acceleration and heading angle sets and joint probabilities in the joint Gaussian distribution, and obtain n by uniform resampling. 3 A set of group state vectors;

[0086] Step 4: Preprocess the motion state data of pedestrians crossing the street acquired by the on-board sensor, the motion state data of the autonomous driving vehicle itself, and the spatial position relationship between pedestrians and vehicles, import the preprocessed data into the graph convolutional neural network GCN model, and train the structural weights and bias parameters of the GCN model;

[0087] Step 5: Use the trained GCN model to predict the trajectory set of pedestrians under different data combinations, and determine the future trajectory of the vehicle while keeping the motion state unchanged;

[0088] Step 6: Determine the collision situation and minimum encounter distance between the pedestrian's future trajectory set and the vehicle's future trajectory;

[0089] Step 7: Use the matter-element extension theory to comprehensively consider the collision situation and the minimum encounter distance to obtain the risk of human-vehicle collision.

[0090] The present invention will be further described below in conjunction with the embodiments and drawings.

[0091] Example

[0092] Combination Figure 1 , a method for analyzing the risk of human-vehicle collision based on an autonomous driving vehicle, comprising the following steps:

[0093] Step 1: Obtain the pedestrian crossing characteristics from the historical data set, including speed, acceleration, and crossing heading angle, and use the Gaussian mixture clustering GMM method to construct a Gaussian mixture distribution function to represent three different pedestrian crossing habits, specifically:

[0094] The motion parameters of pedestrians crossing the street in the presence of a human-vehicle game are obtained through the existing data set, and the crossing characteristics of pedestrians in the historical data set are extracted, including speed, acceleration and crossing heading angle.

[0095] The pedestrian trajectory heading angle refers to: the angle formed by the pedestrian trajectory and the curb of the road. Since China adopts the right-hand driving system, the heading angle is acute when the trajectory is consistent with the prescribed direction of travel, and it is obtuse when the trajectory is opposite to the prescribed direction of travel.

[0096] The vehicle motion state information includes: the vehicle's heading angle θ, the vehicle's travel speed γ, and the vehicle's travel acceleration ω in the past few seconds.

[0097] The vehicle heading angle refers to: the angle between the vehicle trajectory and the curb of the road. Since China adopts the right-hand driving system, the heading angle is acute when the trajectory is consistent with the prescribed direction of travel, and it is obtuse when the trajectory is opposite to the prescribed direction of travel.

[0098] The spatial relationship information between the pedestrian and the vehicle includes: the absolute position of the autonomous driving vehicle and the absolute position of the pedestrian.

[0099] Step 1-1, construct probability density function:

[0100]

[0101] Where μ represents the 3D mean vector, Σ represents the 3*3 covariance matrix determined by the pedestrian's heading angle, velocity and acceleration, and p(x) represents the probability density function of the random vector χ in the 3D sample space χ that obeys the Gaussian distribution;

[0102] Step 1-2: Determine the Gaussian mixture distribution function based on the three pedestrian crossing behavior habits:

[0103]

[0104] Among them, μ i and∑ i is the parameter of the i-th Gaussian mixture component, k = 3, α i >0 is the corresponding mixing coefficient,

[0105] Step 1-3, use the EM algorithm to obtain the optimal Gaussian distribution parameters after iteration. The EM algorithm optimization solution means that in each iteration, the posterior probability of each sample belonging to each Gaussian component is calculated according to the current parameters (E step), and then the model parameters are updated by maximum likelihood estimation and Laarrange multiplier method (M step), specifically:

[0106] Step 1-3-1, initialize Gaussian parameters μ i ,∑ i and α i , for each sample Z j , determine the posterior probability of belonging to each Gaussian distribution:

[0107]

[0108] Step 1-3-2: Determine each sample x j The cluster label λ j :

[0109]

[0110] Step 1-3-3: Label λ by cluster j Divide the sample set into 3 clusters C = {C1, C2, C3};

[0111] Step 1-3-4: Use the maximum likelihood method to construct the Lagrangian function to update the Gaussian distribution parameters of each cluster: μ i ,∑ i and α i , for m samples and k clusters, the Lagrangian function formula is:

[0112]

[0113] The mean vector μ of each Gaussian distribution is i The update formula is:

[0114]

[0115] The covariance matrix ∑ of each Gaussian distribution is i The update formula is:

[0116]

[0117] The mixing coefficient α of each Gaussian distribution i The update formula is:

[0118]

[0119] Next, end the current iteration and convert the newly obtained μ′ i ,∑′ i and α i ′ is used as the initial parameter for the next iteration until the set iteration stop requirement is met.

[0120] Combination Figure 2 , the zebra crossing section without traffic lights under free flow is selected as the implementation area. For the zebra crossing without traffic lights, the real trajectory of pedestrians crossing the road section and the trajectory of vehicles competing with them can be obtained through video shooting. By preprocessing the existing data set, the movement state information of pedestrians crossing the road in front of the vehicle, the movement state information of the vehicle, and the spatial relationship information of pedestrians and vehicles are obtained; the preprocessed data is used to extract the festival features through Gaussian mixture clustering to obtain three different crossing habits and their characteristic distributions.

[0121] Step 2: Obtain the motion parameters of the pedestrian in front of the autonomous driving vehicle, and match the corresponding pedestrian crossing habits determined in step 1 based on the motion characteristics of the pedestrian, specifically:

[0122] Step 2-1: Using the laser radar, camera and / or millimeter wave radar sensors carried by the autonomous driving vehicle, the pedestrian's heading angle, historical trajectory, current speed and acceleration parameters are collected in real time through multi-sensor information fusion to form a state vector η;

[0123] The camera carried by the autonomous driving vehicle is used to collect the posture and head movement of the pedestrian, and thereby obtain the pedestrian's heading angle.

[0124] The laser radar device carried by the autonomous driving vehicle is used to collect the position and kinematic parameters of the pedestrian, and thereby obtain the pedestrian's historical trajectory and current speed and acceleration parameters;

[0125] Step 2-2: Substitute the pedestrian’s state vector η into the Gaussian mixture distribution and calculate the posterior probability that the vector η belongs to the three Gaussian distributions, p M (C=1|η), p M (C=2|η), p M (C=3|η), the Gaussian distribution with the largest posterior probability is the Gaussian distribution to which the pedestrian’s current state vector belongs, and the pedestrian’s crossing habit is then determined.

[0126] Step 3: Taking the pedestrian’s real speed, acceleration and heading angle as the center and taking positive and negative σ as the value interval, obtain all possible speed, acceleration and heading angle sets and joint probabilities in the joint Gaussian distribution, and obtain n by uniform resampling. 3 The group state vector set is:

[0127] Taking the three-dimensional state vector η composed of the pedestrian's actual movement speed, acceleration and heading angle as the center and the positive and negative σ as the value interval, obtain the set and joint probability of all possible speeds, accelerations and heading angles in its joint Gaussian distribution;

[0128] The obtained possible speed, acceleration and heading angle set interval and joint probability interval are divided into n parts respectively, and then n parts are determined by cross matching. 3 A set of three-dimensional vectors of possible pedestrian states and the probability of each set.

[0129] Step 4: Preprocess the motion state data of pedestrians crossing the street acquired by the on-board sensor, the motion state data of the autonomous driving vehicle itself, and the spatial position relationship between pedestrians and vehicles, import the preprocessed data into the graph convolutional neural network GCN model, and iterate the structural weights and bias parameters of the GCN model multiple times according to the gradient descent method;

[0130] The graph convolutional neural network GCN model includes a graph structure, a convolutional layer, a pooling layer, and a fully connected layer.

[0131] The graph structure refers to a directed graph composed of a space domain or a vertex domain, and is used to extract the spatial features of a topological graph.

[0132] The convolution layer refers to a network layer composed of several convolution units after convolution kernel transformation, and the parameters of each convolution unit are optimized through the back propagation algorithm.

[0133] The pooling layer divides the input image into several rectangular areas and outputs the maximum value for each sub-area, which is a form of downsampling.

[0134] The fully connected layer means that each node is connected to all nodes in the previous layer.

[0135] Step 5: Use the trained GCN model to predict the trajectory set of pedestrians under different data combinations, and determine the future trajectory of the vehicle while maintaining the same state of motion, specifically:

[0136] Step 5-1: Obtain the speed, acceleration, and position of the current vehicle and pedestrians, including the speed, acceleration, and position of the pedestrians, and the speed, acceleration, and position of the vehicle. After data normalization, obtain the real-time trajectory sequence of pedestrians and vehicles, and combine it with the n obtained in step 3. 3 One of the group state vectors is input into the trained GCN model and outputs n 3 The output prediction data is denormalized to obtain n 3 The movement trajectory of pedestrians crossing the street in the next 2 seconds;

[0137] Step 5-2: Determine the trajectory of the vehicle within the next 2 seconds based on vehicle dynamics.

[0138] Step 6: Determine the collision situation and minimum encounter distance between the pedestrian's future trajectory set and the vehicle's future trajectory, specifically:

[0139] Step 6-1, determine the collision situation, consider the real outline of the vehicle and simulate the pedestrian as a circle with a diameter of 0.5 meters, determine the collision situation of each pedestrian's future trajectory with the vehicle's future trajectory, and determine the comprehensive collision probability ρ based on the probability of each trajectory;

[0140] Step 6-2, determining the minimum encounter distance, refers to considering the actual contour of the vehicle and assuming that the pedestrian is a circle with a diameter of 0.5 meters, determining the minimum encounter distance y between each future trajectory of the pedestrian and the future trajectory of the vehicle without a collision, and determining the weighted average minimum encounter distance y based on the probability of each trajectory.

[0141] Step 7: Use the matter-element extension theory to comprehensively consider the collision situation and the minimum encounter distance to obtain the risk of human-vehicle collision, specifically:

[0142] Step 7-1, construct a two-dimensional space consisting of the ρ axis and the y axis, and set the boundary range of the rectangular classical domain Φ1 and the extension domain Φ2 respectively. According to the weighted average minimum encounter distance y and the combined collision probability ρ determined in step 6, a point P3 is determined in the two-dimensional space, as shown in Figure 3 As shown;

[0143] Step 7-2: Based on the points determined in step 7-1, use the extenics theory to calculate the extenics distance to determine the correlation function value, so as to reduce the two-dimensional Figure 4 The purpose of the one dimension shown;

[0144] Then the extension distances from point P3 to the classical domain and the extension domain are expressed as ρ(P3, (P4, P1)) and ρ(P3, (P5, P2)) respectively:

[0145]

[0146]

[0147] The correlation degree K(S) is expressed as:

[0148]

[0149] When the correlation is greater than 1, there is no conflict; when the correlation is between 0 and 1, there is a minor conflict and the vehicle should slow down to avoid the conflict; when the correlation is less than 0, there is a major conflict and the vehicle should stop to avoid the conflict.

[0150] In summary, the human-vehicle collision risk analysis method based on pedestrian trajectory prediction based on an autonomous driving vehicle proposed in the present invention utilizes the high-precision detection equipment carried by the autonomous driving vehicle to obtain the pedestrian movement data in front of the vehicle, predicts the future trajectory set of pedestrians, combines the vehicle trajectory with the pedestrian trajectory set, and comprehensively considers safety and comfort, thereby achieving a more accurate human-vehicle collision risk judgment.

[0151] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A method for analyzing the risk of human-vehicle collision based on an autonomous driving vehicle, characterized in that: The following steps are involved: Step 1: Obtain the pedestrian crossing characteristics from the historical data set, including speed, acceleration, and crossing heading angle, and use the Gaussian mixture clustering GMM method to construct a Gaussian mixture distribution function to represent three different pedestrian crossing habits: Step 1-1, construct probability density function: Where μ represents the 3D mean vector, Σ represents the 3*3 covariance matrix determined by the pedestrian's heading angle, velocity and acceleration, and p(x) represents the probability density function of the random vector χ in the 3D sample space χ that obeys the Gaussian distribution; Step 1-2: Determine the Gaussian mixture distribution function based on the three pedestrian crossing behavior habits: Among them, μ i and∑ i is the parameter of the i-th Gaussian mixture component, k = 3, α i >0 is the corresponding mixing coefficient, Step 1-3, using the EM algorithm to obtain the optimal Gaussian distribution parameters after iteration; Step 2: Obtain the motion parameters of the pedestrian in front of the autonomous driving vehicle, and match the corresponding pedestrian crossing habits determined in step 1 based on the motion characteristics of the pedestrian; Step 3: Taking the pedestrian's actual speed, acceleration and heading angle as the center and the positive and negative σ as the value interval, obtain all possible speed, acceleration and heading angle sets and joint probabilities in the joint Gaussian distribution, and obtain n by uniform resampling. 3 A set of group state vectors; Step 4: Preprocess the motion state data of pedestrians crossing the street acquired by the on-board sensor, the motion state data of the autonomous driving vehicle itself, and the spatial position relationship between pedestrians and vehicles, import the preprocessed data into the graph convolutional neural network GCN model, and train the structural weights and bias parameters of the GCN model; Step 5: Use the trained GCN model to predict the trajectory set of pedestrians under different data combinations, and determine the future trajectory of the vehicle while keeping the motion state unchanged; Step 6: Determine the collision situation and minimum encounter distance between the pedestrian's future trajectory set and the vehicle's future trajectory; Step 7: Use the matter-element extension theory to comprehensively consider the collision situation and the minimum encounter distance to obtain the risk of human-vehicle collision.

2. The method for analyzing the risk of collision between a person and a vehicle based on an autonomous driving vehicle according to claim 1, characterized in that: The optimal Gaussian distribution parameters after iteration are obtained by using the EM algorithm in step 1-3, specifically: Step 1-3-1, initialize Gaussian parameters μ i ,∑ i and α i , for each sample Z j , determine the posterior probability of belonging to each Gaussian distribution: Step 1-3-2: Determine each sample x j The cluster label λ j : Step 1-3-3: Label λ by cluster j Divide the sample set into 3 clusters C = {C1, C2, C3}; Step 1-3-4: Use the maximum likelihood method to construct the Lagrangian function to update the Gaussian distribution parameters of each cluster: μ i ,∑ i and α i , for m samples and k clusters, the Lagrangian function formula is: The mean vector μ of each Gaussian distribution is i The update formula is: The covariance matrix ∑ of each Gaussian distribution is i The update formula is: The mixing coefficient α of each Gaussian distribution i The update formula is: Next, end the current iteration and convert the newly obtained μ′ i ,∑′ i and α′ i As the initial parameters for the next iteration, until the set iteration stop requirements are met.

3. The method for analyzing the risk of collision between a person and a vehicle based on an autonomous driving vehicle according to claim 1, characterized in that: The step 2 uses the motion features of pedestrians to match the corresponding pedestrian crossing habits obtained in step 1, specifically: Step 2-1, collect the pedestrian's heading angle, historical trajectory, current speed, and acceleration parameters in real time to form a state vector η; Step 2-2: Substitute the pedestrian’s state vector η into the Gaussian mixture distribution and calculate the posterior probability that the vector η belongs to the three Gaussian distributions, p M (C=1|η), p M (C=2|η), p M (C=3|η), the Gaussian distribution with the largest posterior probability is the Gaussian distribution to which the pedestrian’s current state vector belongs, and the pedestrian’s crossing habit is then determined.

4. The method for analyzing the risk of collision between a person and a vehicle based on an autonomous driving vehicle according to claim 1, characterized in that: The step 3 of obtaining a discrete state vector set is specifically as follows: Taking the three-dimensional state vector η composed of the pedestrian's actual movement speed, acceleration and heading angle as the center and the positive and negative σ as the value interval, obtain the set and joint probability of all possible speeds, accelerations and heading angles in its joint Gaussian distribution; The obtained possible speed, acceleration and heading angle set interval and joint probability interval are divided into n parts respectively, and then n parts are determined by cross matching. 3 A set of three-dimensional vectors of possible pedestrian states and the probability of each set.

5. The method for analyzing the risk of human-vehicle collision based on an autonomous driving vehicle according to claim 1, characterized in that: The step 5 of determining the set of pedestrian trajectories and the trajectory of the vehicle is specifically as follows: Step 5-1: Obtain the speed, acceleration, and position of the current vehicle and pedestrians, including the speed, acceleration, and position of the pedestrians, and the speed, acceleration, and position of the vehicle. After data normalization, obtain the real-time trajectory sequence of pedestrians and vehicles, and combine it with the n obtained in step 3. 3 One of the group state vectors is input into the trained GCN model and outputs n 3 The output prediction data is denormalized to obtain n 3 The movement trajectory of pedestrians crossing the street in the next 2 seconds; Step 5-2: Determine the trajectory of the vehicle within the next 2 seconds based on vehicle dynamics.

6. The method for analyzing the risk of collision between a person and a vehicle based on an autonomous driving vehicle according to claim 1, characterized in that: The determination of the collision situation and the minimum encounter distance in step 6 is specifically as follows: Step 6-1, determine the collision situation, consider the real outline of the vehicle and simulate the pedestrian as a circle with a diameter of 0.5 meters, determine the collision situation of each pedestrian's future trajectory with the vehicle's future trajectory, and determine the comprehensive collision probability ρ based on the probability of each trajectory; Step 6-2, determining the minimum encounter distance, refers to considering the actual contour of the vehicle and assuming that the pedestrian is a circle with a diameter of 0.5 meters, determining the minimum encounter distance y between each future trajectory of the pedestrian and the future trajectory of the vehicle without a collision, and determining the weighted average minimum encounter distance y based on the probability of each trajectory.

7. The method for analyzing the risk of collision between a person and a vehicle based on an autonomous driving vehicle according to claim 1, characterized in that: The step 7 of obtaining the risk of collision between a person and a vehicle is specifically as follows: Step 7-1, construct a two-dimensional space consisting of the ρ axis and the y axis, and set the boundary range of the rectangular classical domain Φ1 and the extension domain Φ2 respectively, and determine a point in the two-dimensional space according to the weighted average minimum encounter distance y and the comprehensive collision probability ρ determined in step 6; Step 7-2: Based on the points determined in step 7-1, use the extenics theory to calculate the extenics distance to determine the correlation function value. When the correlation is greater than 1, there is no conflict; when the correlation is between 0 and 1, there is a minor conflict and the vehicle should slow down to avoid it; when the correlation is less than 0, there is a major conflict and the vehicle should stop to avoid it.

8. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A computer storable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images