A mine unmanned transport vehicle collision prediction method based on adaptive expansion

Abstract

The present application relates to a kind of mine unmanned transport vehicle collision prediction method based on adaptive inflation, belong to mine unmanned transport vehicle collision prediction technical field, solve the problem of insufficient accuracy and high computational complexity in prior art, including: step S1, data preprocessing module carries out the collection and pre-processing of data, obtains driving scene recognition dataset and trajectory prediction dataset, and further carries out normalization and sliding window processing, obtains driving scene recognition model training dataset and trajectory model training dataset;Step S2, trajectory prediction module constructs driving scene recognition model and trajectory prediction model, training data set in step S1 is used to complete the training of driving scene recognition model and trajectory prediction model, executes trajectory prediction, obtains predicted trajectory;Step S3, collision prediction module carries out collision prediction judgment based on predicted trajectory, obtains judgment result and outputs.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of collision prediction of mine unmanned transport vehicles, in particular to a mine unmanned transport vehicle collision prediction method based on adaptive inflation. BACKGROUND

[0002] The application demand of closed scenes such as open-pit mine unmanned driving is increasing. Compared with manned driving, unmanned driving has higher safety, more scientific transportation management and other advantages, and the rapid development of unmanned driving technology in the mine field can greatly improve the equipment production efficiency and reduce the loss caused by improper human operation. However, there is still not small resistance in the development of mine unmanned driving field at present. Due to the complexity and harshness of the production environment of open-pit mine, the mine unmanned transport vehicle may collide with obstacles such as stones, soil piles and cartons during driving, or collide with the retaining wall when deviating from the road, resulting in damage to the vehicle. Therefore, in order to realize the complete intelligentization of open-pit mine, it is still necessary to focus on solving the problem of collision prediction of mine unmanned transport vehicles driving on unstructured roads in the mine.

[0003] The existing contents related to unmanned vehicle collision are mainly divided into collision detection and collision prediction. Collision detection refers to the process of identifying whether the objects will collide at the current moment or will collide soon, which is usually based on geometric method, physical method, graphical method, etc., and the distance between objects, intersection points, etc. are detected to judge; collision prediction refers to the process of predicting whether the moving objects will collide in a period of time in the future, which is based on motion model, trajectory prediction, probability model, learning method, etc. The future motion trajectory of the object is predicted to evaluate the collision risk.

[0004] Collision detection is very effective in simple environment due to its real-time and reliability, but it cannot provide forward-looking warning, especially in complex environment, which shows poor adaptability, so it is not suitable for complex environment in open-pit mine; collision prediction provides forward-looking information to help mine unmanned transport vehicles take preventive measures in advance, improve the overall safety and decision quality. However, the complexity of risk assessment in collision prediction and the dependence on data will greatly increase the difficulty of system implementation, and also have high requirements on computing resources.

[0005] Therefore, there is a need in the art for a collision prediction method of mine unmanned transport vehicles which can provide high accuracy and low computational complexity. SUMMARY

[0006] In order to solve the above problems, an adaptive inflation-based mine unmanned transportation vehicle collision prediction method is provided according to an embodiment of the present application. Specifically, considering the daily driving stability requirements of the mine unmanned transportation vehicle, combining the complex structure of the mine unmanned transportation vehicle and the road environment of the open-pit mine, an adaptive inflation-based mine unmanned transportation vehicle collision prediction method is proposed. When the vehicle collision trend is detected, an early warning is given to the downstream modules of the system, thereby enhancing the safety and reliability of the daily operation of the mine unmanned transportation vehicle.

[0007] According to an embodiment of the present application, an adaptive inflation-based mine unmanned transportation vehicle collision prediction method is provided, which comprises the following steps:

[0008] Step S1, the data preprocessing module performs data acquisition and preprocessing to obtain a driving scene recognition data set and a trajectory prediction data set, and further performs normalization and sliding window processing to obtain a driving scene recognition model training data set and a trajectory model training data set;

[0009] Step S2, the trajectory prediction module constructs a driving scene recognition model and a trajectory prediction model, and uses the training data set in step S1 to complete the training of the driving scene recognition model and the trajectory prediction model, performs trajectory prediction, and obtains a predicted trajectory;

[0010] Step S3, the collision prediction module performs collision prediction based on the predicted trajectory, obtains a judgment result and outputs it;

[0011] Specifically, the step S1 comprises:

[0012] Step S1.1, collect historical driving sensor data of different mine unmanned transportation vehicles in different mines, classify the collected driving sensor data and add classification labels to obtain a driving scene recognition data set and a trajectory prediction data set;

[0013] Step S1.2, use Kalman filtering to clean the driving scene recognition data set and the trajectory prediction data set to obtain cleaned driving scene recognition data set and trajectory prediction data set;

[0014] Step S1.3, select features with high correlation with trajectory prediction from the cleaned driving scene recognition data set and trajectory prediction data set for Z-score normalization to obtain a normalized scene recognition data set and a normalized trajectory prediction data set. The features with high correlation with trajectory prediction include 9 features: X coordinate in UTM coordinate system, Y coordinate in UTM coordinate system, heading angle, vehicle speed V, yaw rate, longitudinal acceleration, lateral acceleration, lateral angle deviation and heading angle deviation;

[0015] Step S1.4, the normalized scene recognition dataset and the normalized trajectory prediction dataset are further processed respectively using a sliding window method to obtain a driving scene recognition model training dataset and a trajectory model training dataset;

[0016] The step S2 specifically comprises:

[0017] Step S2.1, a driving scene recognition model based on a GRU network is constructed, and the driving scene recognition model based on the GRU network is trained to obtain a trained driving scene recognition model;

[0018] Step S2.2, a trajectory prediction model is constructed based on GA-seq2seqGRU and is trained, wherein the trajectory prediction model has an encoding-decoding structure, and includes two independent GRU network loops, one of which is an encoder and the other of which is a decoder;

[0019] Step S2.3, the hyperparameters of the trajectory prediction model are optimized using a GA algorithm to obtain an optimized trajectory prediction model;

[0020] Step S2.4, a predicted trajectory is obtained through the trained driving scene recognition model and the trained and optimized trajectory prediction model;

[0021] The step S3 specifically comprises:

[0022] Step S3.1, a vehicle box is constructed based on the obtained predicted trajectory;

[0023] Step S3.2, an expansion coefficient expand is calculated to expand the vehicle box to obtain an expanded vehicle box;

[0024] Step S3.3, an obstacle box is constructed based on vertex information of the obstacle, and the expanded vehicle box is combined to determine whether the mine unmanned transport vehicle and the obstacle have a collision trend as a determination result;

[0025] Step S3.4, a barrier collision prediction is performed to determine whether the mine unmanned transport vehicle and the barrier have a collision trend as a determination result;

[0026] Step S3.5, the determination results of the mine unmanned transport vehicle and the obstacle collision prediction and the barrier collision prediction are fed back and output to a downstream safety module for decision-making.

[0027] Optionally, the step S2.1 specifically comprises:

[0028] The 7 features of the trajectory sequence in the historical time domain of the input driving scene recognition model training data set, and the lateral angle deviation and the heading angle deviation of the last frame in the time domain, are respectively processed by FC (full connection layer) and then spliced, and then processed by ReLU activation function, and taken as the input of the GRU network, wherein the 7 features include the X coordinate in the UTM coordinate system, the Y coordinate in the UTM coordinate system, the heading angle, the vehicle speed V, the yaw rate, the longitudinal acceleration, and the lateral acceleration;

[0029] The last hidden state is extracted from the output of the GRU network, and the hidden state is further extracted by FC to extract features and reduce dimensions, so as to map the complex feature representation in the network to a lower dimensional space while retaining the most relevant features, and finally the probability combination W = [q1, q2] of the normal driving and abnormal deviation events is obtained by processing through the sigmoid function, wherein q1 is the normal driving probability, and q2 is the abnormal deviation probability, and the trained driving scene recognition model is obtained.

[0030] Optionally, the step S2.2 specifically comprises:

[0031] The trajectory prediction model is constructed to have an encoding-decoding structure, wherein two independent GRU network cycles are included, one of which is used as an encoder and the other of which is used as a decoder, and the loss function is selected as shown in the following formula, wherein y i and y′ i are the observation value and the prediction value respectively, and n1 is the total number of prediction values.

[0032]

[0033] In the training stage of the trajectory prediction model, the encoder inputs the 9 features of the sequence of the trajectory model training data set in the historical time domain processed by the foregoing steps, and outputs a context vector;

[0034] The trained driving scene recognition model obtained by the step S2.1 is used to output a scene recognition result;

[0035] The obtained scene recognition result is spliced with the context vector into a trajectory prediction vector after being processed by FC, and taken as the data stream of the input of the decoder;

[0036] The input data stream in the decoder is shown in the following formula, wherein the hidden state h t-1 and the decoder output vector q t-1 of the previous time are used to update the hidden state h t of the current time, and then the hidden state h t of the current time and the decoder output vector q t-1 of the previous time are used to update the decoder output q t of the current time.

[0037] h t = f(h t-1 , q t-1 )

[0038] q t = g(h t , q t-1 )

[0039] wherein f(·) denotes a nonlinear transformation, and g(·) denotes a nonlinear transformation as well;

[0040] The output of each time of the decoder needs to be adjusted in dimension by a fully connected layer to obtain a trajectory point, and the future trajectory is gradually generated by using such a loop structure, and the number of loops is equal to the prediction time domain T o correlation, and finally the future prediction trajectory is output

[0041] O (t) = [X (t) , Y (t) , heading (t) ], t = (k + 1, k + 2,...)

[0042] wherein X (t) denotes the X coordinate of the future trajectory in the UTM coordinate system, Y (t) denotes the Y coordinate of the future trajectory in the UTM coordinate system, heading (t) denotes the heading angle of the future trajectory, k denotes the current time, and t denotes the time frame.

[0043] Optionally, the step S2.3 specifically comprises:

[0044] Step S2.3.1, population initialization, setting the parameter range and the population number, and initializing the population;

[0045] Step S2.3.2, setting the fitness function, and calculating the fitness score by using the individual after the training is completed;

[0046] Step S2.3.3, selection operation, after the parent individuals are all trained, selecting the individuals with high ranking as offspring by the fitness score;

[0047] Step S2.3.4, crossover operation, randomly selecting two individuals to perform crossover, and exchanging part of the genes to generate new offspring;

[0048] Step S2.3.5, mutation operation, randomly mutating the individuals in the population, and changing some gene values in the individual gene string as offspring;

[0049] Step S2.3.6, iteration optimization, setting the iteration round, repeating the execution of steps S2.3.3 to S2.3.5 until the maximum iteration round is met, obtaining the optimal combination, using the optimal parameter combination to perform trajectory prediction model training, and obtaining the optimized trajectory prediction model.

[0050] Optionally, the step S3.1 specifically comprises:

[0051] The obtained spatial position information includes the predicted trajectory, the reference path, the retaining wall, and the obstacle;

[0052] The base point of the vehicle body box is selected by using a downsampling method on the obtained predicted trajectory, including obtaining the X, Y, and heading of the base point, wherein X represents the X coordinate of the base point in the UTM coordinate system, Y represents the Y coordinate of the base point in the UTM coordinate system, and heading represents the heading angle of the base point;

[0053] An initial vehicle body box is constructed with the center of the rear axle of the mine unmanned transport vehicle as the initial origin, and the initial vehicle body box is converted into the vehicle body coordinate system by rotation and translation in combination with the width of the mine unmanned transport vehicle and the front and rear half axle lengths

[0054]

[0055] wherein, is the corner point of the vehicle body box in the vehicle body coordinate system, (x traj ,y traj ) is the base point coordinate; w vel is the vehicle width; l F and l R are the front half axle length and the rear half axle length of the vehicle respectively, represents the x coordinate of the corner point after translation and rotation, represents the y coordinate of the corner point after translation and rotation, and i represents the corner point number. The total number of the corner points of the vehicle body box is 4.

[0056] Optionally, the step S3.2 specifically comprises:

[0057] The inflation coefficient is set as:

[0058]

[0059] wherein, W, W min , and W max are the current road width, the minimum road width, and the maximum road width respectively; C, C min , and C max are the current road curvature, the minimum road curvature, and the maximum road curvature respectively; a and b are weight coefficients; r1 is the power index of the width, r2 is the power index of the curvature; and k is the compensation coefficient.

[0060] Join the expansion coefficient expand as a proportional scaling to expand the vehicle body box converted into the vehicle body coordinate system, to obtain the corner point coordinates of the expanded vehicle body box i=1, 2, 3, 4, so as to obtain the expanded vehicle body box.

[0061] Optionally, the step S3.3 specifically comprises:

[0062] Constructing the obstacle box based on the vertex information of the obstacle;

[0063] For the expanded vehicle body box, the separation axis theorem is used for obstacle collision prediction:

[0064]

[0065] Wherein, n is the projection axis of two polygons; is the vertex of the obstacle box, j represents the vertex number of the obstacle box, P(·) represents the projection operation, min traj represents the minimum value of the range interval of the expanded vehicle body box after projection on n, max traj represents the maximum value of the range interval of the expanded vehicle body box after projection on n, min obs represents the minimum value of the range interval of the obstacle box after projection on n, max obs represents the maximum value of the range interval of the obstacle box after projection on n;

[0066] Judge whether two intervals [min traj , max traj ] and [min obs , max obs ] overlap on axis n, if there is an axis that makes the two intervals not overlap, it is judged that the vehicle body box does not collide with the obstacle box, that is, it is judged that the mine unmanned transport vehicle does not collide with the obstacle; Otherwise, it is judged that the mine unmanned transport vehicle has a collision trend with the obstacle.

[0067] Optionally, the step S3.4 specifically comprises:

[0068] Obtain the retaining wall point information O(x bd ,y bd );

[0069] In turn, calculate the cross product between the retaining wall point O(x bd ,y bd ) and the vector formed by each vertex and its adjacent vertex of the four polygons of the expanded vehicle body box

[0070]

[0071] wherein, and represents the coordinates of the corner points of the expanded vehicle body box, and Cross(·) represents the vector cross product formula;

[0072] The cross product result obtained is analyzed, when all cross product signs are the same, it is defined that the retaining wall point is located inside the vehicle body box, and it is judged that there is a collision trend between the mine unmanned transport vehicle and the retaining wall; otherwise, it is judged that there is no collision risk between the mine unmanned transport vehicle and the retaining wall.

[0073] The technical scheme adopted by the present application is a mine unmanned transport vehicle collision prediction method based on adaptive expansion. Since the closed mine transportation path is single, there is a large amount of daily operation data, therefore a database is constructed based on historical normal data and road deviation data, a scene recognition model and a trajectory prediction model are built based on a gated recurrent unit (GRU), a vehicle body box is constructed based on down-sampling of the predicted trajectory, the road width, curvature and vehicle information are considered for adaptive expansion of the vehicle body box, then an obstacle box is constructed using obstacle information, and finally it is judged whether there is a collision with the obstacle based on the separation axis theorem, and whether there is a collision with the retaining wall point based on the point and polygon relationship, so as to realize the full-cycle health monitoring and protection function. The steps of the technical scheme are as follows:

[0074] The first part is data preprocessing: data cleaning, feature extraction and normalization are performed on the vehicle sensor data. Kalman filter is used for data cleaning; key features are selected for network training; and Z-score method is used for normalization.

[0075] The second part is trajectory prediction: a driving scene recognition module and a trajectory prediction module are constructed to predict the future trajectory of the mine unmanned transport vehicle.

[0076] The third part is collision prediction: the predicted trajectory input by the upper model is combined with vehicle information, road width, curvature, obstacle information and retaining wall point information, a vehicle body box is constructed based on the adaptive expansion coefficient along the predicted trajectory, an obstacle box is constructed, and it is judged whether there is a collision with the obstacle and the retaining wall based on the separation axis theorem and the point and polygon relationship.

[0077] Compared with the prior art, the mine unmanned transport vehicle collision prediction method based on adaptive expansion provided by the embodiment of the present application has at least the following beneficial effects:

[0078] 1. Aiming at the timeliness, high accuracy and unmanned demand of collision prediction in open-pit mine, a collision prediction algorithm based on trajectory prediction and graphic method combination is proposed, which can effectively reduce the computational complexity.

[0079] 2. In view of the high uncertainty characteristics of different scenarios for trajectory prediction, a trajectory prediction model based on driving scenario recognition is proposed, which splices the scenario recognition vector to improve the trajectory prediction accuracy of vehicles in different scenarios and indirectly improve the collision prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0080] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below, and the features and advantages of the present application can be more clearly understood by referring to the drawings. The drawings are schematic and should not be understood as any limitation on the present application. For those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0081] Figure 1 is a flowchart of a mine unmanned transportation vehicle collision prediction method based on adaptive inflation according to an embodiment of the present application.

[0082] Figure 2 is a longitudinal acceleration Kalman filter cleaning effect diagram of a mine unmanned transportation vehicle collision prediction method based on adaptive inflation according to an embodiment of the present application.

[0083] Figure 3 is a trajectory prediction module architecture diagram of a mine unmanned transportation vehicle collision prediction method based on adaptive inflation according to an embodiment of the present application.

[0084] Figure 4 is a schematic diagram of coordinate conversion, vehicle box inflation and collision prediction judgment in an example applying a mine unmanned transportation vehicle collision prediction method based on adaptive inflation according to an embodiment of the present application.

[0085] Figure 5 is a collision prediction process effect schematic diagram in ROS\Rviz (a three-dimensional visualization platform) of an example applying a mine unmanned transportation vehicle collision prediction method based on adaptive inflation according to an embodiment of the present application.

[0086] Figure 6 is a collision prediction process effect schematic diagram in ROS\Rviz (a three-dimensional visualization platform) of an example applying a mine unmanned transportation vehicle collision prediction method based on adaptive inflation according to an embodiment of the present application. DETAILED DESCRIPTION

[0087] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0088] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description.

[0089] A mine unmanned transportation vehicle collision prediction method based on adaptive inflation is provided according to an embodiment of the present application, which is described in detail below with reference to the accompanying drawings.

[0090] In view of the fact that the existing graphic method does not consider the rapid change of road environment, and in combination with the characteristics of variable road width and curvature in the open-pit mine area, the embodiment of the present application provides a graphic method considering adaptive inflation coefficient.

[0091] As shown in Figure 1 The mine unmanned transportation vehicle collision prediction method based on adaptive inflation according to the embodiment of the present application includes three processing modules: a data preprocessing module, a trajectory prediction module, and a collision prediction module. The specific description is as follows.

[0092] Step S1, the data preprocessing module performs data collection and preprocessing to obtain a driving scene recognition data set and a trajectory prediction data set, and further performs normalization and sliding window processing to obtain a driving scene recognition model training data set and a trajectory model training data set. This step specifically includes the following steps.

[0093] Step S1.1, collect historical driving sensor data of different mine unmanned transportation vehicles in different mine areas for driving scene recognition and trajectory prediction. Among them, the collected driving sensor data is classified by scene and labeled, which can be divided into two categories: normal driving and abnormal deviation, to obtain a driving scene recognition data set and a trajectory prediction data set. The driving scene recognition data set and the trajectory prediction data set after the additional classification label are the total data set.

[0094] Step S1.2, use Kalman filtering to clean the total data set to obtain a cleaned data set, and the cleaning effect is as shown in Figure 2 The cleaned data set includes a cleaned driving scene recognition data set and a cleaned trajectory prediction data set.

[0095] Step S1.3, for the cleaned data set, select the features with high correlation with trajectory prediction for Z-score normalization, here 9 feature values of X, Y, heading, V, yaw rate, longitudinal acceleration, lateral acceleration, lateral angle deviation and heading angle deviation are selected as key features. Among them, X represents the X coordinate in the UTM coordinate system (UNIVERSAL TRANSVERSE MERCATOR PROJECTION), Y represents the Y coordinate in the UTM coordinate system, heading represents the heading angle, and V represents the vehicle speed. For the cleaned driving scene recognition data set, 7 features except lateral angle deviation and heading angle deviation are normalized to obtain the normalized scene recognition data set. For the cleaned trajectory prediction data set, all 9 features are normalized to obtain the normalized trajectory prediction data set.

[0096] Step S1.4, the normalized scene recognition data set and the normalized trajectory prediction data set are further processed using the sliding window method. Let the historical time domain be T p , the prediction time domain be T o , the sliding window length be T p , and the sampling step be 1. The normalized driving scene recognition data set is processed to obtain the driving scene recognition model training data set; the sliding window length is set to T p + T o , and the sampling step is 1. The normalized trajectory prediction data set is processed to obtain the trajectory model training data set.

[0097] Step S2, the trajectory prediction module constructs a driving scene recognition model and a trajectory prediction model, adopts the training data set in step S1 to complete the training of the driving scene recognition model and the trajectory prediction model, executes trajectory prediction to obtain a predicted trajectory. Among them, first, a driving scene recognition model is constructed based on a GRU (Gated Recurrent Unit) network, the historical trajectory sequence features are input into the driving scene recognition model, and the probability combination W of two events of normal driving probability and abnormal deviation probability is output. The output result of the driving scene recognition model is combined and a trajectory prediction model is constructed based on GA-seq2seqGRU (Genetic Algorithm Sequence to Sequence Gated Recurrent Unit), the historical trajectory sequence features are still input, and the future predicted trajectory O (t) is output. The driving scene recognition model and the trajectory prediction model are as shown in Figure 3 . This step S2 specifically includes the following steps.

[0098] Step S2.1: Construct a driving scene recognition model based on a GRU network and train it to obtain the trained driving scene recognition model. During this training and usage process, the seven features of the trajectory sequence in the historical time domain and the lateral and heading angle deviations of the last frame in the time domain are input to the training dataset of the driving scene recognition model processed in the above steps. These features are then concatenated through a fully connected layer (FC) and processed by the ReLU activation function, serving as input to the GRU network. The last hidden state is extracted from the output of the GRU network, and this hidden state is further processed by FC to extract features and reduce dimensionality. The aim is to map the complex feature representations in the network to a lower-dimensional space while retaining the most relevant features. Finally, after processing by the sigmoid function (S-shaped growth curve function), the probability combination W = [q1, q2] of normal driving and abnormal deviation events is obtained, where q1 is the probability of normal driving and q2 is the probability of abnormal deviation. The cross-entropy loss function is used to train the driving scene recognition model. The trained driving scene recognition model is saved for use in trajectory prediction model training and practical prediction applications.

[0099] Step S2.2: Construct a trajectory prediction model based on GA-seq2seqGRU and train it. For example... Figure 3 As shown, the trajectory prediction model has an encoder-decoder structure, which includes two independent GRU network recurrent bodies, one as an encoder and the other as a decoder. The loss function is selected as shown in Equation (1), where y i and y′ i These are the observed values ​​and the predicted values, respectively, with n1 being the total number of predicted values.

[0100]

[0101] During the training phase of the trajectory prediction model, the encoder inputs nine features of the trajectory model training dataset processed by the above steps within the historical time domain, outputs a context vector, and outputs the scene recognition result using the driving scene recognition model trained in step S2.1. This scene recognition result is then concatenated (cat) with the context vector after passing through the FC (Full-Functional Control) to form the trajectory prediction vector, which serves as the input to the decoder. The data flow in the decoder is shown in equation (2). The decoder needs to calculate the trajectory prediction vector based on the decoder output vector q from the previous time step. t-1 and hidden state h t-1 Update the hidden state h at the current moment t Then, based on the current hidden state h... t and the decoder output vector q from the previous time step t-1 Update the decoder output q at the current time step. t :

[0102]

[0103] where f(·) denotes a nonlinear transformation and g(·) denotes a nonlinear transformation as well.

[0104] The output of each time of the decoder needs to be adjusted by a fully connected layer to get the trajectory point. The future trajectory is generated step by step using such a recurrent structure, and the number of loops is equal to the prediction time domain T o Correspondingly, the final output is the future prediction trajectory:

[0105] O (t) = [X (t) , Y (t) , heading (t) ], t = (k+1, k+2,...)

[0106] where X (t) denotes the X coordinate of the future trajectory in the UTM coordinate system, Y (t) denotes the Y coordinate of the future trajectory in the UTM coordinate system, heading (t) denotes the heading angle of the future trajectory, k denotes the current time, and t denotes the time frame.

[0107] Step S2.3, the hyperparameters of the trajectory prediction model are optimized using the GA algorithm to obtain an optimized trajectory prediction model, specifically including the following steps.

[0108] Step S2.3.1, population initialization, setting the parameter range and population size, and initializing the population.

[0109] Step S2.3.2, setting the fitness function, and calculating the fitness score using the individual after training.

[0110] Step S2.3.3, selection operation, after all parent individuals are trained, the individuals with high ranking are selected as offspring through the fitness score.

[0111] Step S2.3.4, crossover operation, randomly select two individuals for crossover, exchange part of the genes to produce new offspring, as shown in Figure 2 .

[0112] Step S2.3.5, mutation operation, randomly mutate the individuals in the population, change some gene values in the individual gene string as offspring.

[0113] Step S2.3.6, iteration optimization, set the iteration round, repeat steps S2.3.3 to S2.3.5 until the maximum iteration round is met, and the optimal combination is obtained. The fitness function of the GA algorithm is formula (1), the iteration round is set to solve the hyperparameters, the optimal parameter combination is used for model training, and the trained trajectory prediction model is saved for actual application.

[0114] Step S2.4, obtain the predicted trajectory by the trained driving scene recognition model and the trajectory prediction model after training and optimization.

[0115] Step S3, the collision prediction module performs collision prediction judgment to obtain the judgment result and output. Based on the predicted trajectory, the road information and vehicle information are considered to calculate the inflation coefficient to construct the vehicle box (two-dimensional rectangular box of vehicle body) and the obstacle box (two-dimensional rectangular box of obstacle), and the collision with the obstacle and the retaining wall is judged based on the separation axis theorem and the point and polygon position relationship, as shown in FIG. 6. This step specifically includes the following steps. Figure 4

[0116] Step S3.1, construct the vehicle box based on the obtained predicted trajectory. The obtained spatial position information, including the predicted trajectory, the reference path, the retaining wall and the obstacle, is converted from the unified UTM coordinate system (UNIVERSAL TRANSVERSE MERCATOR PROJECTION) to the vehicle body coordinate system with the rear axle center of the vehicle as the initial origin. The base point for constructing the vehicle box is selected by using the downsampling method for the obtained predicted trajectory. The vehicle box is constructed according to the X, Y and heading of the base point in combination with the width, front and rear axle lengths of the mine unmanned transport vehicle. The specific method is to construct an initial vehicle box with the rear axle center of the mine unmanned transport vehicle as the initial origin, and then convert the initial vehicle box to the vehicle body coordinate system by rotation and translation. As shown in formula (3),

[0117]

[0118] wherein θ is the base point heading angle; (x traj ,y traj ) is the base point coordinate; w vel is the vehicle width; l F and l R are the front axle length and rear axle length of the vehicle respectively, represents the x-coordinate of the corner point after translation and rotation, represents the y-coordinate of the corner point after translation and rotation, and i represents the corner point number. The total number of vehicle body box corner points is 4.

[0119] ​​Step S3.2, calculate the expansion coefficient expand to expand the vehicle body box to obtain the expanded vehicle body box. In order to ensure the safety of the mine unmanned transport vehicle driving in the open-pit mine unstructured road environment, the vehicle body box expansion concept is introduced, and the expansion coefficient expand is added to appropriately expand the vehicle body box. The expansion coefficient acts as a proportional scaling. Since the mine unmanned transport vehicle has a generally low risk in small curvature sections, and a relatively high driving speed in large curvature sections, it is necessary to maintain a low false detection rate, and the vehicle body box should be slightly expanded. When driving in large curvature sections, the vehicle faces higher risks, and the vehicle body box should continue to increase the expansion degree to improve safety. At the same time, due to the restriction of road width, the expansion area can be appropriately increased in wide road sections, and the expansion needs to be limited to ensure a low false detection rate in narrow road sections. As shown in equation (4), the expansion coefficient is:

[0120]

[0121] wherein W, W min and W max are the current road width, the minimum road width and the maximum road width, respectively; C, C min and C max are the current road curvature, the minimum road curvature and the maximum road curvature, respectively; a and b are weight coefficients; r1 is the power index of width, and r2 is the power index of curvature; k is a compensation coefficient.

[0122] The expansion coefficient expand is added as a proportional scaling to expand the vehicle body box converted to the vehicle body coordinate system to obtain the corner point coordinates of the expanded vehicle body box i = 1, 2, 3, 4, thereby obtaining the expanded vehicle body box.

[0123] Step S3.3, based on the vertex information of the obstacle, an obstacle box is constructed, and combined with the expanded vehicle body box, it is judged whether the mine unmanned transport vehicle and the obstacle have a collision trend, as the judgment result. The separation axis theorem is used as the basic method for obstacle collision prediction:

[0124]

[0125] wherein n is the projection axis of the two polygons; is the vertex of the obstacle box, j represents the vertex number of the obstacle box, P(·) represents the projection operation, min traj represents the minimum value of the range interval of the expanded vehicle body box after projection on n, max traj represents the maximum value of the range interval of the expanded vehicle body box after projection on n, min obs represents the minimum value of the range interval of the obstacle box after projection on n, maxobs represents the maximum value of the range interval of the obstacle box projected on n.

[0126] Next, it is determined whether the two intervals [min traj , max traj ] and [min obs , max obs ] overlap on the axis n. If there is an axis such that the above condition does not hold, i.e., there is an axis such that the two intervals do not overlap, it is determined that the vehicle body box does not collide with the obstacle box, i.e., the mine unmanned transport vehicle does not collide with the obstacle. Otherwise, it is determined that the mine unmanned transport vehicle has a collision trend with the obstacle.

[0127] Step S3.4, barrier collision prediction is performed to determine whether the mine unmanned transport vehicle has a collision trend with the barrier. As a result of the determination, the position relationship between the point and the polygon is calculated by using the vector cross product method. Specifically, as shown in equation (6), the cross products between the barrier point O(x bd , y bd ) and the vectors formed by the adjacent vertices of the expanded vehicle body box are calculated in sequence.

[0128]

[0129] wherein, and represent the coordinates of the corner points of the expanded vehicle body box, and Cross(·) represents the vector cross product formula. The expanded expression of the cross product function is as follows:

[0130] Cross(A,B,C)=(B x -A x )(C y -A y _-)B y -A y _(C x -A x )

[0131] wherein, A, B, C, A x , B x , C x , A y , B y , and C y represent the variables in the cross product function.

[0132] The cross product result is analyzed. When all cross product signs are the same, it is defined that the barrier point is located inside the vehicle body box, and it is judged that the mine unmanned transport vehicle has a collision trend with the barrier; otherwise, it is judged that the mine unmanned transport vehicle has no collision risk with the barrier. In the embodiment, the barrier point is existing data, which can be obtained from the upstream module.

[0133] Step S3.5, the judgment results of the mine unmanned transport vehicle and the barrier collision prediction and the barrier collision prediction are fed back and output to the downstream safety module for decision making. As shown in Figure 5 and Figure 6 The collision prediction process effect in ROS\Rviz (a three-dimensional visualization platform) is exemplarily shown, wherein the gray long curve is a reference path point set; the gray short curve is a predicted trajectory; the black curve is a barrier point set; the black solid square is an obstacle; and the black hollow square is a vehicle body box.

[0134] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application.

[0135] All the optional technical solutions described above can be combined to form optional embodiments of the present application, which will not be described one by one here.

[0136] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution. The execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0137] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered in the protection scope of the present application.

Claims

1. A collision prediction method for unmanned mining transport vehicles based on adaptive dilation, characterized in that, Includes the following steps: Step S1: The data preprocessing module performs data collection and preprocessing to obtain the driving scene recognition dataset and trajectory prediction dataset, and further performs normalization and sliding window processing to obtain the driving scene recognition model training dataset and trajectory model training dataset. Step S2: The trajectory prediction module constructs a driving scene recognition model and a trajectory prediction model. It uses the training dataset from step S1 to train the driving scene recognition model and the trajectory prediction model, performs trajectory prediction, and obtains the predicted trajectory. Step S3: The collision prediction module performs collision prediction judgment based on the predicted trajectory, obtains the judgment result, and outputs it. Specifically, step S1 includes: Step S1.1: Collect historical driving sensor data of different mining unmanned transport vehicles in different mining areas, classify the collected driving sensor data into scenarios and attach classification labels to obtain driving scenario recognition dataset and trajectory prediction dataset. Step S1.2: Use Kalman filtering to clean the driving scene recognition dataset and trajectory prediction dataset to obtain the cleaned driving scene recognition dataset and trajectory prediction dataset; Step S1.3: For the cleaned driving scene recognition dataset and trajectory prediction dataset, select the features with high correlation to trajectory prediction and perform Z-score normalization respectively to obtain the normalized scene recognition dataset and the normalized trajectory prediction dataset. The features with high correlation to trajectory prediction include 9 features: X coordinate in UTM coordinate system, Y coordinate in UTM coordinate system, heading angle, vehicle speed V, yaw rate, longitudinal acceleration, lateral acceleration, lateral angle deviation and heading angle deviation. Step S1.4: Use the sliding window method to further process the normalized scene recognition dataset and the normalized trajectory prediction dataset to obtain the driving scene recognition model training dataset and the trajectory model training dataset. Specifically, step S2 includes: Step S2.1: Construct a driving scene recognition model based on GRU network, and train the driving scene recognition model based on GRU network to obtain the trained driving scene recognition model. Step S2.2: Construct a trajectory prediction model based on GA-seq2seqGRU and train it. The trajectory prediction model has an encoder-decoder structure, which includes two independent GRU network recurrent bodies, one as an encoder and the other as a decoder. Step S2.3: Use the GA algorithm to optimize the hyperparameters of the trajectory prediction model to obtain the optimized trajectory prediction model; Step S2.4: Obtain the predicted trajectory using the trained driving scene recognition model and the trained and optimized trajectory prediction model; Specifically, step S3 includes: Step S3.1: Construct the vehicle body box based on the obtained predicted trajectory; Step S3.2: Calculate the expansion coefficient expand to expand the vehicle body box and obtain the expanded vehicle body box; Step S3.3: Construct obstacle boxes based on the vertex information of obstacles, and combine them with the expanded vehicle body boxes to determine whether there is a collision trend between the mining unmanned transport vehicle and the obstacles, and use the result as the judgment result; Step S3.4: Perform retaining wall collision prediction to determine whether there is a collision trend between the unmanned mining transport vehicle and the retaining wall, and use the judgment result as the result. Step S3.5: Feed back the judgment results of the collision prediction between the unmanned mining transport vehicle and the obstacle and the collision prediction with the retaining wall, and output them to the downstream safety module for decision-making. Step S3.2 specifically includes: Set the expansion coefficient: Among them, W, W min and W max These represent the current road width, minimum road width, and maximum road width, respectively; C, C min and C max Let be the current road curvature, the minimum road curvature, and the maximum road curvature, respectively; a and b are weighting coefficients; r1 is the power exponent of the width, r2 is the power exponent of the curvature; and k is the compensation coefficient. By adding an expansion factor (expand) as a scaling factor, the vehicle body box transformed into the vehicle body coordinate system is expanded, resulting in the corner coordinates of the expanded vehicle body box. This results in the expanded vehicle body box; The vehicle body box is a two-dimensional rectangular box, and the obstacle box is a two-dimensional rectangular box.

2. The collision prediction method for unmanned mining transport vehicles based on adaptive expansion according to claim 1, characterized in that, Step S2.1 specifically includes: The seven features of the trajectory sequence in the historical time domain of the input driving scene recognition model training dataset, as well as the lateral angle deviation and heading angle deviation of the last frame in the time domain, are concatenated after passing through a fully connected layer, and then processed by the ReLU activation function as the input of the GRU network. The seven features include the X coordinate in the UTM coordinate system, the Y coordinate in the UTM coordinate system, the heading angle, the vehicle speed V, the yaw rate, the longitudinal acceleration, and the lateral acceleration. The last hidden state is extracted from the output of the GRU network, and then further features are extracted and dimensionality is reduced by passing the hidden state through a fully connected layer. This maps the complex feature representation in the network to a lower-dimensional space while retaining the most relevant features. Finally, after processing by the sigmoid function, the probability combination W = [q1, q2] of the two events, normal driving and abnormal deviation, is obtained, where q1 is the probability of normal driving and q2 is the probability of abnormal deviation, thus obtaining the trained driving scene recognition model.

3. The collision prediction method for unmanned mining transport vehicles based on adaptive expansion according to claim 2, characterized in that, Step S2.2 specifically includes: A trajectory prediction model with an encoder-decoder structure is constructed, comprising two independent GRU network recurrent bodies, one acting as the encoder and the other as the decoder. The loss function is chosen as shown in the following equation, where y i and y′ i These are the observed values ​​and the predicted values, respectively, where n1 is the total number of predicted values; During the training phase of this trajectory prediction model, the encoder takes into input nine features of the trajectory model training dataset in the historical time domain after the aforementioned steps and outputs a context vector. The trained driving scene recognition model obtained in step S2.1 is used to output the scene recognition result; The obtained scene recognition result is concatenated with the context vector after passing through a fully connected layer to form a trajectory prediction vector, which is then used as the data stream input to the decoder. The input data stream is represented in the decoder as shown in the following equation, based on the decoder output vector q from the previous time step. t-1 and hidden state h t-1 Update the hidden state h at the current moment t Then, based on the current hidden state h t and the decoder output vector q from the previous time step t-1 Update the decoder output q at the current time step. t : h t =f(h t-1 ,q t-1 ) q t =g(h t ,q t-1 ) Where f(·) represents a nonlinear transformation, and g(·) similarly represents a nonlinear transformation; The output of the decoder at each time step needs to be adjusted in dimensionality by a fully connected layer to obtain the trajectory point. This iterative structure is used to progressively generate the future trajectory, with the number of iterations corresponding to the prediction time domain T. o The correlation ultimately outputs the predicted future trajectory. O (t) =[X (t) ,Y (t) ,heading (t) ],t=(k+1,k+2,…) Among them, X (t) The X and Y coordinates of the future trajectory in the UTM coordinate system are represented by... (t) This represents the Y-coordinate of the future trajectory in the UTM coordinate system, heading. (t) The heading angle represents the future trajectory, k represents the current time, and t represents the time frame.

4. The collision prediction method for unmanned mining transport vehicles based on adaptive expansion according to claim 3, characterized in that, Step S2.3 specifically includes: Step S2.3.1, Population initialization: Set the range of each parameter and the population size, and initialize the population; Step S2.3.2: Set the fitness function, which can be used to calculate the fitness score after the individual has finished training; Step S2.3.3: Select operation. After all parent individuals have been trained, select the individual with the highest fitness score as the offspring. Step S2.3.4, crossover operation: randomly select two individuals for crossover, exchange some genes to produce new offspring; Step S2.3.5, mutation operation: randomly mutate individuals in the population to change certain gene values ​​in the individual's gene string as offspring; Step S2.3.6: Iterative optimization. Set the number of iteration rounds and repeat steps S2.3.3 to S2.3.5 until the maximum number of iteration rounds is met to obtain the optimal combination. Use the optimal parameter combination to train the trajectory prediction model and obtain the optimized trajectory prediction model.

5. The collision prediction method for unmanned mining transport vehicles based on adaptive expansion according to claim 4, characterized in that, Step S3.1 specifically includes: The obtained spatial location information includes the predicted trajectory, reference path, retaining walls, and obstacles; The predicted trajectory is downsampled to select the base point for constructing the vehicle body box. This includes obtaining the X, Y and heading of the base point. X represents the X coordinate of the base point in the UTM coordinate system, Y represents the Y coordinate of the base point in the UTM coordinate system, and heading represents the heading angle of the base point. An initial vehicle body box is constructed with the rear axle center of the unmanned mining transport vehicle as the initial origin. Then, by combining the width of the unmanned mining transport vehicle and the lengths of the front and rear axles, the initial vehicle body box is transformed into the vehicle body coordinate system through rotation and translation. Where θ is the heading angle of the base point. Let x be the corner point of the vehicle body box in the vehicle body coordinate system. traj ,y traj ) represents the coordinates of the base point; w vel For vehicle width; l F and l R These refer to the lengths of the front and rear axles of the vehicle, respectively. This represents the x-coordinate of the corner point after translation and rotation. This represents the y-coordinate of the corner point after translation and rotation, and i represents the corner point number. The total number of corner points of the vehicle body box is 4.

6. The collision prediction method for unmanned mining transport vehicles based on adaptive dilation according to claim 5, characterized in that, Step S3.3 specifically includes: Construct the obstacle box based on the vertex information of the obstacle; For the expanded vehicle body box, obstacle collision prediction is performed using the split axis theorem: Where n represents the projection axes of the two polygons; Let j be the vertex of the obstacle box, and P(·) represent the projection operation. traj The maximum value represents the minimum range of the expanded vehicle body's box projected onto n. traj The min represents the maximum value of the range of the expanded vehicle body box projected onto n. obs The maximum value represents the minimum range of the obstacle box projected onto n. obs This represents the maximum value of the range of the obstacle box projected onto n; Determine the minimum value of two intervals [min] traj max traj ] and [min obs max obs If there exists an axis n such that the two intervals do not overlap, then it is determined that the vehicle body box and the obstacle box do not collide, that is, it is determined that the unmanned mining transport vehicle will not collide with the obstacle; otherwise, it is determined that the unmanned mining transport vehicle has a tendency to collide with the obstacle.

7. The collision prediction method for unmanned mining transport vehicles based on adaptive expansion according to claim 6, characterized in that, Step S3.4 specifically includes: Obtain retaining wall point information O(x) bd ,y bd ); Calculate the retaining wall point O(x) sequentially bd ,y bd The cross product between the vector formed by each vertex of the expanded car body box quadrilateral and its adjacent vertices. in, and This represents the coordinates of the corner points of the expanded vehicle body box, and Cross(·) represents the formula for the cross product of vectors. Analyzing the obtained cross product results, when all cross product signs are the same, the retaining wall point is defined as being located inside the vehicle body box, indicating a potential collision between the unmanned mining transport vehicle and the retaining wall; otherwise, it is determined that there is no risk of collision between the unmanned mining transport vehicle and the retaining wall.

Images