Unmanned mine card safety protection method based on double-layer risk estimation

Abstract

The invention relates to an unmanned mine card safety protection method based on double-layer risk estimation, belongs to the technical field of unmanned driving safety, and solves the problem that a passive safety protection mode in the prior art can only relieve accident hazards after an accident occurs and cannot perform accident prevention and early warning. Obtaining a prediction track; s2, acquiring information of a task path and a sensing obstacle, and performing collision detection and safe distance calculation on the task path, the predicted trajectory and the sensing obstacle to obtain a safe distance; s3, performing safety risk field calculation according to the obtained obstacle sensing information to obtain a safety risk score; and S4, based on the obtained safety distance and the safety risk field, carrying out control quantity calculation on the vehicle speed, obtaining a control instruction, outputting the control instruction, and carrying out driving control on the vehicle.

Description

Technical Field

[0001] This invention relates to the field of unmanned driving safety technology, and in particular to a safety protection method for unmanned mining trucks based on two-layer risk estimation. Background Technology

[0002] Because mining areas are relatively enclosed and have a simple environment with limited vehicle-to-vehicle and vehicle-to-human interaction, and given the strong market demand, autonomous driving in mining areas has long been considered the easiest scenario for its implementation. However, mining trucks are extremely large and heavy; for example, the Caterpillar 797F mining truck has a rated weight of 623.7 tons. Accidents in these trucks could cause very serious safety hazards and economic losses. Therefore, high safety requirements are placed on unmanned mining trucks. Unmanned mining trucks are generally equipped with advanced sensors (such as LiDAR, millimeter-wave radar, cameras, and combined inertial navigation systems), controllers, and actuators. Sensors perceive environmental information around the unmanned mining truck, and the perception program processes this raw information, converting it into specific protocol data and sending it downstream. Downstream driving safety modules receive the processed environmental information and the vehicle's own driving status (such as positioning, speed, and turning angle), assess current and future safety hazards, and enable the vehicle to respond promptly to ensure the safety of unmanned driving in mining areas.

[0003] Currently, safety methods for autonomous driving include passive safety protection and active safety protection. Passive safety protection focuses on safety protection devices, such as airbags and collision avoidance structures, which can mitigate impacts and reduce the risk of equipment damage and property loss after a dangerous situation occurs.

[0004] Active safety protection generally refers to protection at the program level of autonomous driving. Through onboard sensors, vehicle-to-everything (V2X) networks, and algorithms, information about the vehicle's situation, environment, and other road users can be detected in a timely manner, and potential risks can be predicted in advance. This allows the vehicle to take specific actions in advance to avoid possible accidents.

[0005] Chinese invention patent application, publication number CN103295424A, entitled "Automotive Active Safety System Based on Video Recognition and Vehicle-mounted Ad Hoc Network," connects vehicles in front of and behind the vehicle in real time to form a mobile ad hoc network through a video acquisition module, processing module, short-range communication module, main control module, and output module. This network transmits information such as vehicle location, license plate, and speed, alerting the user to take appropriate action and achieving active safety protection for the vehicle. However, this method only addresses the acquisition and identification of information about surrounding vehicles on conventional roads. In application scenarios such as mines where vehicle information is not readily available, and considering obstacles caused by random events like falling ore, it cannot provide an effective active safety solution for autonomous mining trucks.

[0006] Chinese invention patent application CN103287372A, entitled "A Vehicle Collision Avoidance Safety Protection Method Based on Image Processing," utilizes an image recognition system for automatic vehicle safety protection. It compares real-time images captured by a camera with information in an image feature database, calculates the actual distance between the obstacle and the vehicle using monocular or binocular ranging methods, and then intervenes to control deceleration, alert the driver, or initiate automatic braking based on the proportional relationship between the actual distance and a preset distance. However, this method is designed for ordinary manned vehicles, and the obstacle identification relies on comparing the actual width, height, or volume data of the obstacle in front of the vehicle, along with color variations at its boundary, with information in the image feature database. This method cannot predict obstacles based on vehicle task allocation in mining environments, nor can it provide an effective proactive safety solution for mining applications.

[0007] Chinese invention patent application, publication number CN105059214A, entitled "Method, Device, and Vehicle for Suppressing Malfunctions of Vehicle Active Safety Functions," compares the vehicle's current latitude and longitude information (located by a Global Positioning System) with latitude and longitude information in a preset list of suppression locations. Based on the comparison result, it performs suppression or issues an active safety system intervention command to activate the vehicle's active safety functions, such as automatic emergency braking. However, this method cannot provide effective active safety protection in mines where Global Positioning System signal reception is difficult, or in situations where obstacles arise due to random events such as ore rolling.

[0008] Chinese invention patent application, publication number CN106696927A, entitled "Control Method and Device for Automatic Emergency Braking of Vehicle, Vehicle," determines whether an external object is within the vehicle's driving lane by using the vehicle's heading angular velocity, the relative distance between the vehicle and an external object, and the offset angle of the external object relative to the vehicle. It also obtains the braking deceleration required for safe driving based on the vehicle's speed, the relative distance and speed between the vehicle and the external object, and the safe braking distance. Based on the relationship between the braking deceleration and a preset threshold, it initiates emergency alarms, predictive alarms, or intervenes to control the vehicle to perform corresponding braking actions. However, this method primarily considers the relationship between the vehicle and other vehicles traveling on regular roads and requires information such as the relative distance and speed between the vehicle and external vehicles through a global positioning system. This makes it difficult to apply to mines where global positioning signal reception is difficult, and it fails to provide effective active safety protection in situations such as obstacles caused by random events like ore rolling in mines.

[0009] Chinese invention patent application, publication number CN107145147A, entitled "A Collision Avoidance Method and System for Low-Speed ​​Autonomous Driving," uses a road environment perception system to monitor obstacles on the road, assess collision risks, and select an avoidance strategy or intervene with active braking to decelerate the vehicle based on the assessment results. However, this method is suitable for conventional road environments with regular path planning. It fails to provide effective active safety protection for situations such as mines without fixed road routes, or for obstacles arising from random events like ore rolling down the mine.

[0010] Chinese invention patent application, publication number CN111427041A, entitled "An Emergency Stopping System for an Unmanned Vehicle in a Mining Area," acquires obstacle data based on environmental information collected by onboard radar and cameras, determines the hazard level of the obstacle based on the obstacle data, and takes corresponding measures according to the hazard level, including emergency braking in response to a level two hazard. However, this method is based on real-time detection of the vehicle's surroundings and does not provide a way to predict the driving path in advance or to anticipate the risk of obstacle collisions.

[0011] Current passive safety protection methods can only mitigate the damage of an accident after it has occurred, but cannot prevent or warn of accidents. Existing active safety protection technologies are costly and primarily designed for urban roads with traffic signals and high levels of vehicle interaction; they are not suitable for mining roads with little or no traffic signal infrastructure and minimal interaction with other vehicles.

[0012] Therefore, there are safety protection methods in this field that can provide active safety protection for unmanned mining trucks traveling on mining roads with complex boundaries, without relying on traffic signal facilities. Summary of the Invention

[0013] To address the aforementioned issues, this invention provides a safety protection method for unmanned mining vehicles based on dual-layer risk estimation. Specifically, considering the complex boundaries of mining areas, the relatively simple operating routes, and the characteristics of mining vehicles being bulky, having lower control precision compared to passenger cars, and exhibiting larger control deviations, a collision detection method integrating a preset driving path and vehicle trajectory estimation is proposed. This method assesses whether there is a collision risk between the vehicle and road boundaries or surrounding environmental obstacles, based on both the vehicle's driving path and the predicted future vehicle trajectory. Simultaneously, it calculates the driving risk field based on perceived obstacles and determines the vehicle's response action based on the distance between the vehicle and the risk location and the intensity of the risk field, thereby ensuring the safe and efficient operation of unmanned mining vehicles.

[0014] A safety protection method for unmanned mining trucks based on two-layer risk estimation, according to an embodiment of the present invention, includes the following steps:

[0015] Step S1: Perform vehicle trajectory prediction to obtain the predicted trajectory;

[0016] Step S2: Obtain information on the task path and perceived obstacles, perform collision detection and safe distance calculation between the task path and predicted trajectory and perceived obstacles, and obtain the safe distance.

[0017] Step S3: Based on the obtained information about perceived obstacles, calculate the safety risk field and obtain a safety risk score;

[0018] Step S4: Calculate the control quantity of the vehicle speed based on the obtained safe distance and safety risk field, obtain the control command and output it to control the vehicle's driving.

[0019] Specifically, step S1 includes the following steps:

[0020] Step S1.1: Vectorize the map information to obtain the map information of the mining area. Represent the elements in the map information as discrete points, and classify the coordinates of the discrete points into a vector set according to geographical location and function to obtain the vectorized map information.

[0021] Step S1.2: Encode the vectorized map information using a graph neural network to obtain the encoded map information. The vectorized map information consists of nodes and edges, where each node represents the start and end point of each vector in the vector set.

[0022] Step S1.3: Encode the vehicle's historical trajectory using a historical trajectory encoding network including an FPN network (Feature Pyramid Network) to obtain the encoded vehicle historical trajectory.

[0023] Step S1.4: The features of the encoded map information and the features of the encoded vehicle historical trajectory are processed by the prediction network to obtain k future trajectories;

[0024] Step S1.5: Score the k future trajectories output by the prediction network and select the trajectory with the highest probability as the predicted trajectory.

[0025] Optionally, step S1.3 specifically includes the following steps:

[0026] Step S1.3.1: Obtain the vehicle's historical trajectory and vectorize it to obtain the vectorized vehicle historical trajectory:

[0027] x = [x t ,x t-τ ,x t-2τ ,…] T (6a)

[0028] y = [yt ,y t-τ ,y t-2τ ,…] T (6b)

[0029] h = [h] t ,h t-τ ,h t-2τ ,…] T (6c)

[0030] Where x represents the set of x-coordinates of the historical trajectory, x t Let y represent the x-coordinate of the vehicle at time t, and y represent the set of y-coordinates of the historical trajectory. t Let y represent the vehicle's y-coordinate at time t, and y represent the set of vehicle headings in the historical trajectory. t Let τ represent the heading of the vehicle at time t, where t represents the current time and τ represents the time step.

[0031] Step S1.3.2: The vectorized historical trajectory of the vehicle obtained above is used as the input of the historical trajectory encoding network. CNN (Convolutional Neural Network) is used for training to obtain the features extracted by CNN. Then, FPN network is used to perform multi-scale feature extraction on the features extracted by CNN to obtain the encoded historical trajectory of the vehicle.

[0032] Optionally, step S2 specifically includes the following steps:

[0033] Step S2.1: Obtain the predicted trajectory and task path, and generate the vehicle envelope of the vehicle.

[0034] Step S2.2: Obtain information about perceived obstacles, filter obstacles, remove obstacles that are far away and obstacles that have no risk of collision, obtain the obstacles to be detected, and process the information of the obstacles to be detected to obtain the obstacle envelope.

[0035] Step S2.3: Obtain the vehicle envelope and obstacle envelope, perform collision detection, and obtain the collision detection results;

[0036] Step S2.4: Obtain and output the safe distance based on the collision detection results.

[0037] Optionally, step S2.1 specifically includes the following steps:

[0038] Step S2.1.1: Obtain the predicted trajectory and task path, traverse the path points of the predicted trajectory and task path, and generate the vehicle envelope of the path point coordinate system based on vehicle speed as the basis for envelope expansion.

[0039] Step S2.1.2: Based on the relationship between the path points and the vehicle's position, transform the coordinates of the vehicle's bounding box from the path point coordinate system to the vehicle coordinate system:

[0040]

[0041] Among them, (x a ,y a Let (x1', y1') be the coordinates of the path point in the vehicle coordinate system, and (x1', y1') be the coordinates of the corner point 1 of the vehicle's bounding box in the path point coordinate system. Let x be the angle of rotation from the path point coordinate system to the vehicle coordinate system. 11 ,y1) are the coordinates of corner point 1 of the rotated vehicle envelope in the vehicle coordinate system;

[0042] Step S2.1.3: Repeat step S2.1.2 above to calculate the coordinates of each corner point of the vehicle envelope in the vehicle coordinate system, and obtain the vehicle envelope transformed from the path point coordinate system to the vehicle coordinate system. Output the transformed vehicle envelope in the vehicle coordinate system as the vehicle envelope.

[0043] Optionally, step S2.3 specifically includes:

[0044] Step S2.3.1: Obtain the vehicle envelope, extract an edge from the vehicle envelope and obtain its normal vector. Use the normal vector as a projection axis. Iteratively obtain each edge of the obstacle envelope and the vehicle envelope, and project them onto the projection axis to obtain the projections of the vehicle envelope and the obstacle envelope.

[0045] Step S2.3.2: Check whether the projections of the vehicle envelope and the obstacle envelope overlap. If they do not overlap, it is determined that the vehicle envelope and the obstacle envelope do not collide. The collision detection result is taken as no collision, and step S2.4 is executed. Otherwise, it is considered that they overlap, and step S2.3.3 is executed.

[0046] Step S2.3.3: Check the edges in the vehicle's envelope. Select one edge from the uncompared edges and return to step S2.3.1. If all edges in the vehicle's envelope have been checked, check the edges in the obstacle's envelope. Select one edge from the uncompared edges and return to step S2.3.1. If all edges in both the vehicle's and obstacle's envelopes have been compared and there is no case of no collision, then it is determined that there is a collision between the vehicle's envelope and the obstacle's envelope, and the collision detection result is taken as the collision detection result.

[0047] Optionally, step S2.4 specifically includes:

[0048] Obtain the collision detection result. If the collision detection result is that there is a collision, that is, the vehicle's envelope is detected to have collided with an obstacle at a certain point on the task path or predicted trajectory, then the distance from the point of collision to the current position of the vehicle is taken as the safe distance.

[0049] The obtained safe distance and trajectory path are output downstream to represent the risk level of the vehicle.

[0050] Optionally, step S3 specifically includes:

[0051] Step S3.1: Select moving obstacles based on the obtained information about perceived obstacles, and perform kinetic energy field calculation to obtain the moving obstacle O for any point (x) in the driving environment. e ,y e Kinetic energy field at point )

[0052]

[0053] Among them, v o Let θ be the velocity of the moving obstacle O. o The speed of the moving obstacle O to any point (x) in the driving environment e ,y e The angle is given by k1, k2, and k3, which are constants, and r is the vector distance between the obstacle point and any point in the driving environment. R represents the direction of the field strength. o m is the road condition influencing factor. o Let O be the virtual mass of the moving obstacle.

[0054] Step S3.2: Select a static obstacle based on the obtained information about the perceived obstacles, calculate the potential energy field of the static obstacle Q, and obtain the potential energy field of the static obstacle Q:

[0055]

[0056] Among them, R Q M is the road condition influencing factor. Q Let k be the virtual mass of the static obstacle Q, and k4 and k5 be constant terms.

[0057] Step S3.3: The kinetic energy field of the moving obstacle and the potential energy field of the static obstacle are merged to obtain the safety risk score as the safety risk field.

[0058] Optionally, step S4 specifically includes:

[0059] Step S4.1: Calculate the target speed of the vehicle based on the safe distance and safety risk score;

[0060] Step S4.2: Based on the obtained target vehicle speed, the acceleration control quantity is obtained through the controller processing.

[0061] Step S4.3: Using throttle calibration, determine the final throttle and brake opening based on the obtained acceleration control quantity, and output them as control commands to control the vehicle's driving.

[0062] Optionally, step S4.1 specifically includes:

[0063] Step S4.1.1: Obtain the security risk scoring threshold T score and safe distance threshold T dis ;

[0064] Step S4.1.2, the target vehicle speed is obtained as follows:

[0065]

[0066] Where score is the obtained safety risk score, v is the current vehicle speed, dis is the obtained safe distance, and k6 and k7 are constant parameters.

[0067] Optionally, step S4.2 specifically includes:

[0068] Construct a proportional-integral controller, including its time domain:

[0069]

[0070] Where, k P k is the proportionality coefficient. I Here, t is the coefficient of the integral term, e(t′) is the speed error, which is the difference between the target vehicle speed and the actual vehicle speed, t′ is the time, τ′ is the derivative of the time, and e(τ′) is the historical speed error.

[0071] The overall transfer function is:

[0072]

[0073] Where s represents the complex frequency domain unit;

[0074] Input the target vehicle speed into the proportional-integral controller constructed by formula (12) or (16) to obtain the acceleration control quantity.

[0075] According to another embodiment of the present invention, a safety protection method for unmanned mining trucks based on dual-layer risk estimation is provided. The current task path and the predicted future trajectory of the unmanned vehicle are read or calculated. Points are uniformly sampled from these two trajectories as detection points, and the vehicle's envelope is generated with the detection points as the rear axle center. The method calculates whether the vehicle's envelope overlaps with the envelopes of surrounding obstacles; if there is overlap, a collision risk is considered. Simultaneously, based on the perceived obstacle information, the driving safety risk field intensity at the vehicle's current position is calculated. Based on the collision distance and risk score, the method determines whether the vehicle should decelerate or brake urgently. The specific steps are as follows:

[0076] The first step is vehicle trajectory prediction: trajectory prediction is performed based on deep learning methods. In order to comprehensively consider the known driving trajectory, road map boundaries, vehicle motion state and vehicle control information, a multimodal trajectory prediction model based on graph convolutional neural network and self-attention mechanism is proposed. The mining area road information is represented by vector map and encoded by convolutional neural network. Then, the vehicle's historical trajectory is encoded based on self-attention mechanism. Then, a fusion network is used to fuse map features and vehicle trajectory features to output multimodal predicted trajectory. Finally, a selection network is used to calculate the trajectory with the highest confidence.

[0077] Step 2: Collision detection outputs safe distance: Collision detection is performed using the vehicle's task path and predicted trajectory against the boundaries and obstacles of the surrounding environment. If a collision is detected, the distance from the nearest collision point to the vehicle is the safe distance.

[0078] Step 3: Construct a safety risk field based on the vehicle's surrounding environment and output a safety risk score: The safety field model divides the factors affecting traffic safety into three categories: the risk generated by moving obstacles around the vehicle is called the kinetic energy field; the risk generated by static obstacles in the environment is called the potential energy field; and the risk caused by the driver's behavior is called the behavioral field. Unifying these three fields forms the "driving safety field" unified model.

[0079] Step 4: Control vehicle speed based on the safe distance in step 2 and the safety risk field output in step 3: First, calculate the target vehicle speed based on the safe distance and the safety risk field score, and then use PID to output the longitudinal control quantity.

[0080] The first step, vehicle trajectory prediction, involves the following specific methods:

[0081] 1) Complete the vectorization of map information and historical trajectory information: map boundaries, task paths, historical trajectories, etc. can all be regarded as a series of lines. By discretizing these lines into ordered points, the image information can be vectorized.

[0082] 2) Encoding map information using graph neural networks: The connectivity information between roads in the mining area can be stored using directed graphs, and graph neural networks can capture the interaction features between roads.

[0083] 3) Encoding historical trajectories using FPN networks: The long-term trend and short-term changes of historical trajectories are important information for trajectory prediction. FPN networks can significantly improve the extraction of multi-scale information without changing the computational cost of CNN.

[0084] 4) The prediction network processes map features and trajectory features: map features can reflect the overall trend of future vehicle trajectories, while historical trajectory features retain the subtle differences in the trajectories of different vehicles at different times. These two features are fused to output a multimodal predicted trajectory, from which the trajectory with the highest confidence is extracted.

[0085] Step 2: The specific methods for collision detection and safe distance calculation are as follows:

[0086] 1) Generate vehicle envelope: Using the task path and the predicted trajectory as baselines, generate the envelope boxes of the vehicle traveling on these two paths, which serve as the coverage area of ​​the vehicle on the map.

[0087] 2) Obstacle filtering: Filtering out obstacles that pose no risk of collision can effectively improve the program's computational efficiency.

[0088] 3) Collision detection: Collision detection is performed on the vehicle's envelope and the envelopes of obstacles that may pose a collision risk.

[0089] 4) Collision distance output: If an obstacle is detected in the surrounding area that conflicts with the vehicle's envelope, output the distance from the obstacle to the vehicle.

[0090] Step 3: The specific method for calculating the safety risk field is as follows:

[0091] As mentioned earlier, the safety risk field consists of three parts: the kinetic energy field, the potential energy field, and the behavioral field. Since vehicle speeds are generally low and road conditions are relatively simple on mining roads, the impact of the behavioral field on vehicle safety will not be considered; only the kinetic and potential energy fields will be discussed. That is:

[0092] E S =E v +E R

[0093] Where E S For driving safety, E v For the kinetic energy field, E R This is the potential energy field. Therefore, it is necessary to calculate the kinetic energy field and potential energy field generated by the surrounding obstacles at the vehicle's position separately, and then add them together to obtain the risk score of the vehicle's current position, i.e., the safety field.

[0094] Step 4: The specific method for controlling the output is as follows:

[0095] The inputs are the safe distance generated in the second step and the safety risk field generated in the third step. The longitudinal speed of the vehicle is controlled based on these two risk quantification data, without considering actions such as detours.

[0096] 1) First, calculate the target speed of the current vehicle based on the collision distance and safety risk score.

[0097] 2) Based on the target vehicle speed and the current actual vehicle speed, the speed error is obtained, and then the expected acceleration is output by the PI controller and feedback control. The feedback method is set to a first-order system.

[0098] 3) Finally, based on the vehicle's throttle-acceleration calibration table, calculate the control amount of the throttle or brake.

[0099] Compared with existing technologies, the unmanned mining truck safety protection method based on two-layer risk estimation provided by the embodiments of the present invention has at least the following beneficial effects:

[0100] 1. The unmanned mining truck safety protection method based on dual-layer risk estimation of the present invention calculates driving risks in real time according to vehicle driving conditions and surrounding environmental information, and makes intelligent decisions on stopping and speed reduction measures according to the magnitude of the risks, thereby realizing active safety protection for unmanned vehicles in mining areas.

[0101] 2. The predicted trajectory is used to calculate the safe distance (the distance between the vehicle and the target object), and the surrounding environment is used to perform risk scoring. The two quantitative indicators (safe distance and risk score) are combined to perform a two-layer risk estimation, which ensures the accuracy of risk estimation.

[0102] 3. The vehicle's future trajectory is estimated based on the vectorized map and the vehicle's historical trajectory. This fully utilizes the road characteristics of the mining area and the inference of the vehicle's future trajectory to ensure the accuracy of the vehicle trajectory prediction.

[0103] 4. Use environmental kinetic and potential energy fields to calculate vehicle driving risk scores, ensuring reasonable speed planning in complex scenarios such as intersections. Attached Figure Description

[0104] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly introduced below. The features and advantages of the present invention can be more clearly understood by referring to the accompanying drawings. The accompanying drawings are schematic and should not be construed as limiting the present invention in any way. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0105] Figure 1 A flowchart illustrating a safety protection method for unmanned mining trucks based on two-layer risk estimation, provided according to an embodiment of the present invention.

[0106] Figure 2 This is a schematic diagram of a map feature extraction network for a safety protection method for unmanned mining trucks based on two-layer risk estimation, provided according to an embodiment of the present invention.

[0107] Figure 3A schematic diagram of the FPN network structure for a safety protection method for unmanned mining trucks based on two-layer risk estimation according to an embodiment of the present invention.

[0108] Figure 4 A collision detection diagram illustrating a safety protection method for unmanned mining trucks based on dual-layer risk estimation, provided according to an embodiment of the present invention.

[0109] Figure 5 A schematic diagram of a proportional-integral feedback controller for a safety protection method for unmanned mining trucks based on dual-layer risk estimation according to an embodiment of the present invention.

[0110] Figure 6 A schematic diagram of an unmanned mining truck driving system provided according to another embodiment of the present invention. Detailed Implementation

[0111] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0112] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0113] The following description, with reference to the accompanying drawings, details the safety protection method for unmanned mining trucks based on dual-layer risk estimation according to an embodiment of the present invention.

[0114] Symbol explanation:

[0115] N i The i-th node;

[0116] v i The vector of the i-th node;

[0117] vector v i The starting point;

[0118] vector v i The end point;

[0119] P j The j-th broken line;

[0120] a i The semantic label of the i-th node;

[0121] Nf i (l)The feature vector of the i-th node in the l-th layer;

[0122] The feature vectors of the neighboring nodes of the previous node;

[0123] Multilayer perceptron;

[0124] Pooling layer;

[0125] This indicates feature concatenation;

[0126] Pf j Features of the j-th polyline;

[0127] The k-th layer feature of the m-th node;

[0128] The feature set of all polylines in Pf;

[0129] Pf n Characteristics of the nth broken line;

[0130] k-subgraph feature extraction forward propagation k layers;

[0131] m-line Pf j It contains m vectors;

[0132] nn broken lines;

[0133] The l-th layer feature of the j-th polyline;

[0134] Adjacency matrix of A-line features;

[0135] The set of x-coordinates of the x-trajectory;

[0136] x t The x-coordinate of the vehicle at time t;

[0137] The set of y-coordinates of the y-trajectory;

[0138] y t The y-coordinate of the vehicle at time t;

[0139] h is the set of vehicle headings in the historical trajectory;

[0140] h t The heading of the vehicle at time t;

[0141] t is the current time;

[0142] τ is the time step;

[0143] (x a ,y aThe coordinates of the path points in the vehicle's coordinate system;

[0144] (x1′,y1′) Coordinates of corner point 1 of the vehicle's envelope in the path point coordinate system; (x1,y1) Coordinates of corner point 1 of the vehicle's envelope in the vehicle's coordinate system;

[0145] The angle of rotation from the path point coordinate system to the vehicle coordinate system;

[0146] (x o ,y o The coordinates of the moving obstacle O;

[0147] v o The velocity of the moving obstacle O;

[0148] θ o The angle between the velocity direction of the moving obstacle O and any point in the driving environment; (x) e ,y e The position coordinates of any point in the driving environment;

[0149] E v,o The kinetic energy field of the moving obstacle O at any point in the driving environment;

[0150] k1, k2, and k3 are constant terms;

[0151] r is the vector distance between the obstacle point and any point in the driving environment;

[0152] R o Factors affecting road conditions;

[0153] M o The virtual mass of the moving obstacle O;

[0154] (x Q ,y Q The coordinates of the static obstacle Q;

[0155] E R,Q The potential energy field of the static obstacle Q at a point within the vehicle's envelope;

[0156] k4 and k5 are constant terms;

[0157] R Q Factors affecting road conditions;

[0158] M Q The virtual mass of the static obstacle Q;

[0159] T score Safety risk scoring threshold;

[0160] T dis Safe distance threshold;

[0161] v Current vehicle speed;

[0162] TTC collision time;

[0163] d min The minimum distance that a vehicle must maintain from an obstacle in front;

[0164] v target Target speed;

[0165] Score the security risk rating;

[0166] dis safe distance;

[0167] k6 and k7 are constant terms;

[0168] t′ represents time;

[0169] τ′ is the time derivative;

[0170] u(t′) is the time domain of the proportional-integral controller;

[0171] k p Proportional coefficient;

[0172] k I coefficient of the integral term;

[0173] e(t′) is the velocity error.

[0174] e(τ′) represents the historical velocity error;

[0175] s represents the complex frequency domain unit;

[0176] Frequency domain characteristics of A(s) acceleration;

[0177] Frequency domain characteristics of velocity V(s);

[0178] A des (s) Frequency domain characteristics of target acceleration.

[0179] like Figure 1 As shown, the unmanned mining truck safety protection method based on dual-layer risk estimation provided by the present invention includes the following steps.

[0180] Step S1 involves predicting the vehicle's trajectory to obtain the predicted trajectory. The vehicle in question is the currently operating unmanned mining truck. In this step, trajectory prediction is performed using deep learning methods. Based on the obtained historical trajectory of the vehicle, road map boundaries, vehicle motion state, and vehicle control information, a multimodal trajectory prediction model based on a graph convolutional neural network (GNN) and a self-attention mechanism is constructed. The mining area road information is represented using a vector map, and the vector map representation of the mining area roads is encoded using a convolutional neural network. The historical trajectory of the vehicle is then encoded using a self-attention mechanism. A fusion network is then used to fuse the features of the encoded map data and the features of the historical trajectory of the vehicle, outputting the multimodal predicted trajectory. Finally, a selection network is used to calculate the trajectory with the highest confidence level as the predicted trajectory. Step S1 specifically includes the following steps.

[0181] Step S1.1: Obtain map information of the mining area. Represent the elements in the map information as discrete points, and categorize the coordinates of these discrete points into vector sets according to geographical location and function to obtain vectorized map information. In this map information, all elements are represented by discrete points. For example, road boundaries are formed by curves composed of a series of points, vehicle reference routes can be composed of a series of points, and the connection methods of different roads at intersections can be abstracted using reference trajectories composed of a series of points. By categorizing these points according to geographical location and function, and forming vector sets from their coordinates, the vectorization of the map information is completed. Optionally, the map information of the mining area can be a top-down view of the mining area.

[0182] Step S1.2: Encode the vectorized map information using a graph neural network to obtain the encoded map information. The vectorized map information consists of nodes and edges, where each node represents the start and end point of each vector in the aforementioned vector set. To distinguish different categories (such as road boundary points, reference trajectory points, etc.), each node also stores its category label, thus ensuring that semantic information can also be correctly encoded. For the connectivity between nodes, a hierarchical graph architecture is used to ensure the connection of their contextual information. Nodes with the same semantic information and belonging to the same group are connected to the same subgraph, and information exchange is ensured between different subgraphs through full connections. Specifically, for a node n... i Assuming it is associated with the vector v i Belongs to the broken line P j And the semantic tag is a i Then the content of that node is:

[0183]

[0184] in and They are vectors v iThe start and end points can be represented in the form (x, y, z), where i represents the node number and j represents the polyline number. Simultaneously, for ease of calculation, the UTM (Universal Transverse Mercator Grid System) coordinates recorded on the map are converted to a coordinate system centered on the current vehicle position. For example... Figure 2 The forward propagation method for subgraph encoding is defined as follows:

[0185]

[0186] Among them, Nf i (l) Refers to the feature vector of the i-th node in the l-th layer. This represents the feature vector of the current node's neighboring nodes. It is a multilayer perceptron (i.e., an MLP network) that extracts features from the current node. It is a pooling layer (i.e., a pooling network) used to process the features of the current node's neighboring nodes, and finally... The features of the current node are concatenated with the features of its neighboring nodes. This yields the features of each polyline in the map.

[0187]

[0188] Pf = {Pf0, Pf1, ..., Pf} n} (4)

[0189] Among them, Pf j This represents the characteristics of the j-th polyline. Pf represents the feature vector of the m-th node in the k-th layer. n Let Pf represent the feature of the nth polyline, k represent the k layers of forward propagation for subgraph feature extraction, and m represent the polyline Pf. j It contains m vectors, where n represents the number of polylines, and Pf represents the feature set of all polylines, i.e., map features. A graph convolutional neural network is used to further extract features from the geographic features concatenated using node features. The forward propagation method is as follows:

[0190]

[0191] in, Let A represent the l-th layer features of the j-th polyline, and let A be the adjacency matrix of the polyline features, which can introduce connectivity information between roads at a larger scale. Thus, by stitching together the features of each polyline at each layer, the map information is encoded, resulting in the encoded map information.

[0192] Step S1.3 involves encoding the vehicle's historical trajectory using a historical trajectory encoding network that includes an FPN (Feature Pyramid Network) to obtain the encoded historical trajectory. Step S1.3 specifically includes the following steps.

[0193] Step S1.3.1: Obtain the vehicle's historical trajectory and vectorize it to obtain the vectorized vehicle historical trajectory. The vehicle's historical trajectory can be represented by a series of points with headings, which are then vectorized:

[0194] x = [x t ,x t-τ ,x t-2τ ,…] T (6a)

[0195] y = [y t ,y t-τ ,y t-2τ ,…] T (6b)

[0196] h = [h] t ,h t-τ ,h t-2τ ,…] T (6c)

[0197] Where x represents the set of x-coordinates of the historical trajectory, x t Let y represent the x-coordinate of the vehicle at time t, and y represent the set of y-coordinates of the historical trajectory. t Let y represent the vehicle's y-coordinate at time t, and h represent the set of vehicle headings in the historical trajectory. t Let t represent the heading of the vehicle at time t, where t represents the current time, τ represents the time step, and t-τ represents the historical time.

[0198] Step S1.3.2: The vectorized historical trajectory of the vehicle obtained above is used as the input of the historical trajectory encoding network. A CNN is used for training to obtain the features extracted by the CNN. Then, the features extracted by the CNN are used to perform multi-scale feature extraction using an FPN network to obtain the encoded historical trajectory of the vehicle. Using an FPN network for feature extraction ensures that more multi-scale features of the historical trajectory are extracted, while minimizing the increased computational cost.

[0199] like Figure 3As shown, the FPN network consists of three parts: a bottom-up network and a top-down network, with data transmission between the two networks via lateral connections. For the bottom-up network, three residual blocks are used, each containing three 1D convolutional networks. After each convolution, the ReLU function (Rectified Linear Unit) is used for activation. The stride of the last convolutional kernel in the last two residual blocks is set to 2, ensuring that the output of each residual block is scaled down by half compared to the previous residual block, thus extracting smaller-scale features. For the top-down network, interpolation is used to upsample the high-level feature map results by a factor of 2, ensuring that the feature tensor scale is the same as that of the bottom-up network output. Then, the upsampled mapping is merged with the corresponding bottom-up mapping by element-wise addition, resulting in a merged mapping. This merged mapping result is then processed using convolutional blocks, and the output serves as the input to the next top-down network layer. The final top-down network layer is the final output of the FPN network. A convolutional block may consist of a 1D convolutional network, a linear network, and a ReLU activation function to reduce aliasing during upsampling. Lateral connections may use a 1D convolutional network.

[0200] Step S1.4: The predictive network processes the features of the encoded map information and the features of the encoded historical trajectory of the autonomous vehicle to obtain the predicted trajectory. Since the autonomous vehicle is an unmanned mining truck, it is assumed that, without external intervention or safety procedures, the future operating state of the vehicle is only related to the task path, and external vehicles have no impact on the vehicle's control output. The predictive network fuses the features of the encoded map information and the features of the encoded historical trajectory of the autonomous vehicle to predict k possible future trajectories.

[0201] Step S1.5: The k future trajectories output by the prediction network are scored by the scoring network, and the trajectory with the highest probability is selected as the prediction trajectory.

[0202] Step S2: Obtain the task path and obstacle information. Use the task path and predicted trajectory to perform collision detection and calculate the safe distance against the perceived obstacles. The safe distance is then determined. Collision detection is performed between the vehicle's task path and predicted trajectory and the boundaries and obstacles of the surrounding environment. If a collision is detected, the distance from the nearest collision point to the vehicle is the safe distance. The task path is issued by the scheduling platform in the autonomous driving system and can be received using the TCP protocol. The obstacle information is issued by the perception module in the autonomous driving system and can be received using the TCP protocol.

[0203] Step S2.1, as follows Figure 4As shown, the predicted trajectory and task path are obtained, and the vehicle's envelope is generated. The trajectory points generated from the predicted trajectory and the read task path points both record the x, y coordinates and heading information of the path points. Using these path points as the rear axle center, the vehicle envelope is generated. The specific process for generating the vehicle envelope includes the following steps.

[0204] Step S2.1.1: Obtain the predicted trajectory and task path, traverse the path points of the predicted trajectory and task path, and generate the vehicle envelope box of the path point coordinate system based on vehicle speed as the basis for envelope expansion.

[0205] Step S2.1.2: Based on the relationship between the path points and the vehicle's position, transform the coordinates of the vehicle's envelope generated in step S2.2.1 from the path point coordinate system to the vehicle coordinate system. The specific coordinate transformation method is as follows: Take the coordinates of the path points in the vehicle coordinate system as (x...). a ,y a The coordinates of corner point 1 of the vehicle's bounding box in the path point coordinate system are (x1′, y1′), and the angle of rotation from the path point coordinate system to the vehicle coordinate system is... After rotation, the coordinates of corner point 1 of the vehicle's envelope in the vehicle's coordinate system are (x1, y1). Therefore, the coordinate transformation formula is:

[0206]

[0207] Step S2.1.3: Repeat step S2.1.2 above to calculate the coordinates of each corner point of the vehicle envelope in the vehicle coordinate system, and obtain the vehicle envelope transformed from the path point coordinate system to the vehicle coordinate system.

[0208] Step S2.2: Obtain information about perceived obstacles, filter obstacles, remove obstacles that are far away or pose no collision risk, obtain the obstacles to be detected, and process the information of the obstacles to be detected to obtain the obstacle envelope. To improve program efficiency, collision detection is not performed on obstacles that are obviously not a risk or that are far away. Obstacles without collision risk can include, for example, filtering out obstacles behind the rear of the vehicle when moving forward, and filtering out obstacles in front of the vehicle when reversing. Obstacles without collision risk can also include, for each trajectory point, first enclosing the vehicle at that point with a circle; if all corners of the obstacle are outside the circle, then the obstacle is considered to have no conflict risk with the trajectory point and is skipped directly.

[0209] Step S2.3: Obtain the vehicle and obstacle envelopes, perform collision detection, and obtain the collision detection results. Collision detection is performed using the separating axis theorem. Since both the vehicle and obstacle envelopes are convex polygons, if these two convex polygons do not collide, there will always be a straight line that separates the two objects; this line is called the separating line. That is, the projections of the two polygons onto the separating axis perpendicular to this separating line do not overlap. Step S2.3 specifically includes the following steps.

[0210] Step S2.3.1: Obtain the vehicle envelope, extract an edge from the vehicle envelope and obtain its normal vector. Use this normal vector as a projection axis. Iteratively obtain each edge of the obstacle envelope and the vehicle envelope, and project them onto the projection axis to obtain the projections of the vehicle envelope and the obstacle envelope.

[0211] Step S2.3.2: Check if the projections of the vehicle envelope and the obstacle envelope overlap. If they do not overlap, it is determined that the vehicle envelope and the obstacle envelope do not collide. The collision detection result is taken as no collision, and step S2.4 is executed. If they overlap, step S2.3.3 is executed.

[0212] Step S2.3.3: Check the edges in the vehicle's envelope, select one edge from the uncompared edges and return to step S2.3.1; If all edges in the vehicle's envelope have been checked, then check the edges in the obstacle's envelope, select one edge from the uncompared edges and return to step S2.3.1; If all edges in both the vehicle's envelope and the obstacle's envelope have been compared and there is still no collision, then it is determined that there is a collision between the vehicle's envelope and the obstacle's envelope, and the collision detection result is taken as the collision detection result.

[0213] Step S2.4: Obtain and output the safe distance based on the collision detection results. If the collision detection result indicates a collision, meaning the vehicle's bounding box is detected to have collided with an obstacle at a point on the task path or predicted trajectory, then the distance from the point of collision to the vehicle's current position is taken as the safe distance. The obtained safe distance and trajectory path are output downstream to represent the vehicle's risk level. Optionally, in other embodiments, if the detected collision detection result indicates no collision, return to step S2.1. Optionally, in other embodiments, if the detected collision detection result indicates no collision, a preset value can be provided to assess the vehicle's risk level, such as a no-risk level.

[0214] Step S3 involves calculating the safety risk field based on the obtained information about perceived obstacles, resulting in a safety risk score. The safety risk field is constructed based on the vehicle's surrounding environment, and the safety risk score is output. In this implementation, the safety field model categorizes traffic safety influencing factors into two types: the risk generated by moving obstacles around the vehicle is called the kinetic energy field; the risk generated by static obstacles in the environment is called the potential energy field. Unifying these two fields creates the unified "driving safety field" model. Step S3 specifically includes the following steps.

[0215] Step S3.1: Select moving obstacles based on the obtained information about perceived obstacles, and calculate the kinetic energy field of the moving obstacles. The kinetic energy field is mainly determined by the moving obstacles in the vehicle's driving environment. If a moving obstacle O has position coordinates (x... o ,y o ), with a speed of v o Calculate the effect of the moving obstacle on any point (x) in the driving environment. e ,y e The kinetic energy field at point (x) is such that the velocity direction of the moving obstacle is towards that point (x). e ,y e The angle is θ o The kinetic energy field of the moving obstacle is:

[0216]

[0217] Where k1, k2, and k3 are constants, k1 uniformly affects the intensity of the entire kinetic energy field, k2 represents the magnitude of the impact of the distance to the obstacle on the risk, and k3 represents the magnitude of the impact of the obstacle's velocity on the risk field. r is the vector distance between the obstacle point and any point in the driving environment. The direction of the electric field strength is represented by the exp function, which reflects the exponential increase in risk as the obstacle's velocity increases, and ensures that the electric field strength remains positive regardless of the angle. cosθ o The magnitude of the field strength is related to the direction of the obstacle's velocity; the closer it matches the direction of the obstacle's velocity, the higher the risk. R o M is the road condition influencing factor. o R is the virtual mass of the moving obstacle O. o M is related to road conditions, gradient, coefficient of friction, etc. o It is related to the type, quality, and size of the obstacle.

[0218] Step S3.2: Based on the obtained information about perceived obstacles, select static obstacles and calculate their potential energy fields to obtain the potential energy fields of the static obstacles. The potential energy field is mainly determined by the static obstacles in the environment. If a static obstacle Q has coordinates (x... Q ,y QThe static obstacle is located at any point (x) in the driving environment. e ,y e The potential energy field generated at point () is:

[0219]

[0220] Compared to the kinetic energy field formula, the influence of obstacle velocity is reduced. Q M Q R represents the virtual mass of the road condition influence factor and the static obstacle Q, respectively. Q M is related to road conditions, gradient, coefficient of friction, etc. Q It is related to the type, mass, and size of static obstacles, and k4 and k5 are constant terms.

[0221] Step S3.3: The kinetic energy field of the moving obstacle and the potential energy field of the static obstacle are merged to obtain a safety risk score as the safety risk field. The driving risk at the vehicle's position is calculated using the above formulas (8) and (9) to obtain the safety risk score.

[0222] Step S4 involves calculating the control input for the vehicle speed based on the obtained safe distance and safety risk field, generating and outputting control commands to control the vehicle's movement. First, the target vehicle speed is calculated based on the safe distance and safety risk score. Then, a PID controller is used to output the longitudinal control input. Step S4 specifically includes the following steps.

[0223] Step S4.1: Calculate the target speed of the vehicle based on the safe distance and safety risk score. This step specifically includes the following steps.

[0224] Step S4.1.1: Obtain the security risk scoring threshold T score and safe distance threshold T dis .

[0225] Among them, the safety risk scoring threshold T score The safe distance threshold T can be set based on empirical values. dis It can be set to be related to TTC (Time To Collision):

[0226] T dis =max(v·TTC,d min (9)

[0227] Where v is the current vehicle speed, TTC is the collision time, which can be set to 3 to 4 seconds, and d minThis is the minimum distance the vehicle must maintain from the obstacle ahead. Optionally, in another implementation, if a collision is detected between the vehicle's envelope and an obstacle at a point on the task path or predicted trajectory in step S2, the collision time can be obtained by combining the safe distance from the point of collision to the vehicle's current position with the current vehicle speed.

[0228] Step S4.1.2, the target vehicle speed is obtained as follows:

[0229]

[0230] Where score is the safety risk score obtained in step S3, v is the current vehicle speed, dis is the safe distance obtained in step S2, and k6 and k7 are constant parameters. The minimum value between the calculated safe distance and the safety risk score is taken as the target vehicle speed.

[0231] Step S4.2: Based on the obtained target vehicle speed, the acceleration control quantity is obtained through controller processing. In this embodiment, the controller is a PID (proportional-integral-derivative) controller. This PID controller includes two layers: one layer is a PI (proportional-integral) controller that calculates the desired acceleration from the target vehicle speed, and the other layer is a controller that converts the acceleration into throttle and brake control information. Specifically, it includes the following steps.

[0232] Construct a proportional-integral controller.

[0233] The time-domain expression of the proportional-integral controller is:

[0234]

[0235] Where, k P k is the proportionality coefficient. I Here, e(t′) is the coefficient of the integral term, e(t′) is the speed error (i.e., the obtained target speed minus the actual speed), t′ is time, e(τ′) is the historical speed error, and τ′ is the derivative of time. Alternatively, in another implementation, in the absence of collision risk, the speed error can be obtained by subtracting the current actual speed from a preset desired road speed or a desired road speed calculated by the onboard control module of the mining truck.

[0236] Therefore, the transfer function of the controller is:

[0237]

[0238] Where s represents the complex frequency domain unit;

[0239] Using a first-order system to simulate the response process of an acceleration system, the following conditions are met:

[0240]

[0241] Where A(s) represents the frequency domain characteristic of acceleration, V(s) represents the frequency domain characteristic of velocity, and A des (s) represents the frequency domain characteristics of the target acceleration;

[0242] The feedback transfer function is then:

[0243]

[0244] Then as Figure 5 The overall transfer function is:

[0245]

[0246] Discretize the above transfer function to meet the program calculation requirements, input the target vehicle speed into the proportional-integral controller constructed by equation (12) or (16), and obtain the acceleration control quantity.

[0247] Step S4.3: Using throttle calibration, determine the final throttle and brake opening based on the obtained acceleration control quantity, and output them as control commands to control the vehicle's driving.

[0248] Optionally, according to another embodiment of the present invention, a safety protection method for unmanned mining trucks based on dual-layer risk estimation is provided. The difference from the method provided in the above embodiment is that the target vehicle speed is obtained in step S4.1.2 as follows:

[0249] v target =min(v score ,v dis )

[0250]

[0251] Among them, v target For the target vehicle speed, v score To adjust the vehicle speed based on the safety risk score, v dis To adjust the vehicle speed based on the safe distance, v max The maximum speed is the preset value, and the score is the obtained safety risk rating. (T) score T represents the safety risk scoring threshold, dis represents the obtained safety distance, and T represents the safety risk score threshold. dis The safe distance threshold is defined by k6 and k7, which are constant parameters.

[0252] like Figure 1As shown, according to another embodiment of the present invention, a safety protection system for unmanned mining trucks based on dual-layer risk estimation is provided. This system can be connected to an unmanned mining truck automatic driving system to execute the unmanned mining truck safety protection method based on dual-layer risk estimation described above. It may include a vehicle trajectory prediction module, a collision detection module, a driving safety field module, and a control output module.

[0253] like Figure 6 As shown, according to another embodiment of the present invention, an unmanned mining truck automatic driving system is provided, including a remote cloud control platform (i.e., a scheduling module), a vehicle-side planning module, a vehicle-side onboard control module, and a vehicle-side unmanned mining truck safety protection system based on dual-layer risk estimation. The remote cloud control platform is used for task planning, distributing map information of the mining area and planned tasks to each unmanned mining truck; the vehicle-side planning module receives the planned tasks from the cloud control platform and processes them to obtain information such as the task path; the vehicle-side onboard control module calculates the desired vehicle speed and performs control based on the task path received from the planning module (e.g., path curvature, whether the path has ended, speed limit, etc.) and the road conditions received from the perception module (e.g., mining area road surface slope, unevenness, road surface adhesion rate, etc.) to execute automatic driving control and operation; the vehicle-side unmanned mining truck safety protection system based on dual-layer risk estimation... The system uses information received from the perception module, such as the vehicle's historical trajectory, road map boundaries, vehicle motion status, and perceived obstacle information, as well as the task path received from the planning module, to predict the vehicle's trajectory. Based on the predicted trajectory, it performs obstacle perception and collision detection, and generates control commands, including control quantities, based on the collision detection results. These commands are then provided to the onboard control module to intervene and execute deceleration control. The unmanned mining truck safety protection system based on dual-layer risk estimation includes a prediction module for predicting the vehicle's trajectory and a risk assessment module for assessing and handling collision risks. The risk assessment module includes a collision detection module, a driving safety field module, and a control output module.

[0254] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0255] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0256] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A safety protection method for unmanned mining trucks based on two-layer risk estimation, characterized in that, Includes the following steps: Step S1: Perform vehicle trajectory prediction to obtain the predicted trajectory; Step S2: Obtain information on the task path and perceived obstacles, perform collision detection and safe distance calculation between the task path and predicted trajectory and perceived obstacles, and obtain the safe distance. Step S3: Based on the obtained information about perceived obstacles, calculate the safety risk field and obtain a safety risk score; Step S4: Calculate the control quantity of the vehicle speed based on the obtained safe distance and safety risk field, obtain the control command and output it to control the vehicle's driving. Specifically, step S1 includes: Step S1.1: Vectorize the map information to obtain the map information of the mining area. Represent the elements in the map information as discrete points, and classify the coordinates of the discrete points into a vector set according to geographical location and function to obtain the vectorized map information. Step S1.2: Encode the vectorized map information using a graph neural network to obtain the encoded map information. The vectorized map information consists of nodes and edges, where each node represents the start and end point of each vector in the vector set. Step S1.3: Encode the vehicle's historical trajectory using a historical trajectory encoding network including an FPN network to obtain the encoded vehicle historical trajectory; Step S1.4: The features of the encoded map information and the features of the encoded vehicle historical trajectory are processed by the prediction network to obtain k future trajectories; Step S1.5: Score the k future trajectories output by the prediction network and select the trajectory with the highest probability as the predicted trajectory. Specifically, step S1.3 includes: Step S1.3.1: Obtain the vehicle's historical trajectory and vectorize it to obtain the vectorized vehicle historical trajectory: x=[x t ,x t-τ ,x t-2τ ,…] T (6a) y=[y t ,y t-τ ,y t-2τ ,…] T (6b) h=[h t ,h t-τ ,h t-2τ ,…] T (6c) Where x represents the set of x-coordinates of the historical trajectory, x t Let y represent the x-coordinate of the vehicle at time t, and y represent the set of y-coordinates of the historical trajectory. t Let represent the y-coordinate of the vehicle at time t, and represent the set of vehicle headings in the historical trajectory. t Let τ represent the heading of the vehicle at time t, where t represents the current time and τ represents the time step. Step S1.3.2: The vectorized historical trajectory of the vehicle obtained above is used as the input of the historical trajectory encoding network. CNN is used for training to obtain the features extracted by CNN. Then, the features extracted by CNN are used to perform multi-scale feature extraction with FPN network to obtain the encoded historical trajectory of the vehicle.

2. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 1, characterized in that, Step S2 specifically includes the following steps: Step S2.1: Obtain the predicted trajectory and task path, and generate the vehicle envelope of the vehicle. Step S2.2: Obtain information about perceived obstacles, filter obstacles, remove obstacles that are far away and obstacles that have no risk of collision, obtain the obstacles to be detected, and process the information of the obstacles to be detected to obtain the obstacle envelope. Step S2.3: Obtain the vehicle envelope and obstacle envelope, perform collision detection, and obtain the collision detection results; Step S2.4: Obtain and output the safe distance based on the collision detection results.

3. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 1, characterized in that, Step S2.1 specifically includes the following steps: Step S2.1.1: Obtain the predicted trajectory and task path, traverse the path points of the predicted trajectory and task path, and generate the vehicle envelope of the path point coordinate system based on vehicle speed as the basis for envelope expansion. Step S2.1.2: Based on the relationship between the path points and the vehicle's position, transform the coordinates of the vehicle's bounding box from the path point coordinate system to the vehicle coordinate system: Among them, (x a ,y a Let (x1', y1') be the coordinates of the path point in the vehicle coordinate system, and (x1', y1') be the coordinates of the corner point 1 of the vehicle's bounding box in the path point coordinate system. Let (x1, y1) be the angle from the path point coordinate system to the vehicle coordinate system, and (x1, y1) be the coordinates of corner point 1 of the rotated vehicle envelope in the vehicle coordinate system. Step S2.1.3: Repeat step S2.1.2 above to calculate the coordinates of each corner point of the vehicle envelope in the vehicle coordinate system, and obtain the vehicle envelope transformed from the path point coordinate system to the vehicle coordinate system. Output the transformed vehicle envelope in the vehicle coordinate system as the vehicle envelope.

4. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 3, characterized in that, Step S2.3 specifically includes: Step S2.3.1: Obtain the vehicle envelope, extract an edge from the vehicle envelope and obtain its normal vector. Use the normal vector as a projection axis. Iteratively obtain each edge of the obstacle envelope and the vehicle envelope, and project them onto the projection axis to obtain the projections of the vehicle envelope and the obstacle envelope. Step S2.3.2: Check whether the projections of the vehicle envelope and the obstacle envelope overlap. If they do not overlap, it is determined that the vehicle envelope and the obstacle envelope do not collide. The collision detection result is taken as no collision, and step S2.4 is executed. Otherwise, it is considered that they overlap, and step S2.3.3 is executed. Step S2.3.3: Check the edges in the vehicle's envelope. Select one edge from the uncompared edges and return to step S2.3.

1. If all edges in the vehicle's envelope have been checked, check the edges in the obstacle's envelope. Select one edge from the uncompared edges and return to step S2.3.

1. If all edges in both the vehicle's and obstacle's envelopes have been compared and there is no case of no collision, then it is determined that there is a collision between the vehicle's envelope and the obstacle's envelope, and the collision detection result is taken as the collision detection result.

5. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 1, characterized in that, Step S2.4 specifically includes: Obtain the collision detection result. If the collision detection result is that there is a collision, that is, the vehicle's envelope is detected to have collided with an obstacle at a certain point on the task path or predicted trajectory, then the distance from the point of collision to the current position of the vehicle is taken as the safe distance. The obtained safe distance and trajectory path are output downstream to represent the risk level of the vehicle.

6. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 1, characterized in that, Step S3 specifically includes: Step S3.1: Select moving obstacles based on the obtained information about perceived obstacles, and perform kinetic energy field calculation to obtain the moving obstacle O for any point (x) in the driving environment. e ,y e Kinetic energy field at point ) Among them, v o Let θ be the velocity of the moving obstacle O. o The speed of the moving obstacle O to any point (x) in the driving environment e ,y e The angle is given by k1, k2, and k3, which are constants, and r is the vector distance between the obstacle point and any point in the driving environment. R represents the direction of the field strength. o M is the road condition influencing factor. o Let O be the virtual mass of the moving obstacle. Step S3.2: Select a static obstacle based on the obtained information about the perceived obstacles, calculate the potential energy field of the static obstacle Q, and obtain the potential energy field of the static obstacle Q: Among them, R Q M is the road condition influencing factor. Q Let k be the virtual mass of the static obstacle Q, and k4 and k5 be constant terms. Step S3.3: The kinetic energy field of the moving obstacle and the potential energy field of the static obstacle are merged to obtain the safety risk score as the safety risk field.

7. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 1, characterized in that, Step S4 specifically includes: Step S4.1: Calculate the target speed of the vehicle based on the safe distance and safety risk score; Step S4.2: Based on the obtained target vehicle speed, the acceleration control quantity is obtained through the controller processing. Step S4.3: Using throttle calibration, determine the final throttle and brake opening based on the obtained acceleration control quantity, and output them as control commands to control the vehicle's driving.

8. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 7, characterized in that, Step S4.1 specifically includes: Step S4.1.1: Obtain the security risk score threshold T score and safe distance threshold T dis ; Step S4.1.2, the target vehicle speed is obtained as follows: Where score is the obtained safety risk score, v is the current vehicle speed, dis is the obtained safe distance, and k6 and k7 are constant parameters.

9. The safety protection method for unmanned mining trucks based on dual-layer risk estimation according to claim 7, characterized in that, Step S4.2 specifically includes: Construct a proportional-integral controller, including its time domain: Where, k P k is the proportionality coefficient. I Here, t is the coefficient of the integral term, e(t′) is the speed error, which is the difference between the target vehicle speed and the actual vehicle speed, t′ is the time, τ′ is the derivative of the time, and e(τ′) is the historical speed error. The overall transfer function is: Where s represents the complex frequency domain unit; Input the target vehicle speed into the proportional-integral controller constructed by formula (12) or (16) to obtain the acceleration control quantity.

Images