Unmanned vehicle active collision avoidance method based on preceding vehicle trajectory clustering prediction

Abstract

The invention discloses an unmanned vehicle active collision avoidance method based on front vehicle trajectory clustering prediction, which comprises the following steps: extracting observation signals in the driving process of a front vehicle to obtain a u-shape set, calculating the distance between the u-shape set and a time sequence to obtain a distance matrix, and clustering the distance matrix to obtain a clustering result; according to the obtained clustering result, a long-short-term memory network model is trained based on data collected by various vehicle-mounted sensors, prediction is carried out, and the driving track of the front vehicle is obtained; and an active collision avoidance strategy is formulated according to the obtained predicted trajectory of the preceding vehicle and the dynamic stability of the own vehicle. According to the method, active collision avoidance strategies including early warning and graded braking are formulated according to the predicted track of the front vehicle and the dynamic stability of the vehicle, so that the driving safety is ensured.

Description

Technical Field

[0001] The invention relates to the technical field of unmanned driving trajectory prediction in open-pit mines, and in particular to an active collision avoidance method for unmanned vehicles based on cluster prediction of leading vehicle trajectories. Background Art

[0002] Collision avoidance technology is one of the key core technologies of driverless vehicles. At present, most longitudinal and lateral collision avoidance strategies assume that the trajectories of surrounding vehicles are known and accurate, and rarely consider their uncertainties. Therefore, existing longitudinal and lateral collision avoidance strategies are not suitable for driverless vehicles. Driverless vehicles use various on-board sensors such as millimeter-wave radar, lidar, and cameras and high-precision maps to perceive external environmental information in real time, comprehensively judge the safety of the vehicle in the future period, and initiate active collision avoidance strategies in a timely manner to ensure its driving safety. Among them, predicting the future driving trajectory of the leading vehicle and obtaining the dynamic safety distance between the leading vehicle and the vehicle are also the prerequisites for active collision avoidance of driverless vehicles. However, due to the influence of external factors such as the hardware properties of various sensors, changes in driving conditions, and road environment parameters, a single leading vehicle trajectory prediction model is difficult to accurately predict the driving trajectory of vehicles with different driving characteristics.

[0003] At present, for the research on the method of predicting the trajectory of the leading vehicle, there are existing literatures on the high-speed vehicle motion planning algorithm based on the prediction of the motion trajectory of the leading vehicle, which considers the motion model and driving intention to predict the trajectory, but this method is only applicable to relatively ideal working conditions. There are existing literatures that use unscented Kalman filters to predict the trajectories of surrounding vehicles to ensure the real-time generation of safe trajectories, but it still cannot effectively solve the complex motion of the leading vehicle. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art. To achieve the above purpose, an unmanned vehicle active collision avoidance method based on clustering prediction of the leading vehicle trajectory is adopted to solve the problems raised in the above background technology.

[0005] An unmanned vehicle active collision avoidance method based on front vehicle trajectory clustering prediction includes the following steps:

[0006] Step S1, extracting the observation signal of the preceding vehicle during driving to obtain a u-shapelets set, calculating the distance between the u-shapelets set and the time series to obtain a distance matrix, and clustering the distance matrix to obtain a clustering result;

[0007] Step S2: According to the obtained clustering results, the long and short-term memory network models are trained based on the data collected by various vehicle-mounted sensors to make predictions and obtain the driving trajectory of the preceding vehicle;

[0008] Step S3: formulate an active collision avoidance strategy based on the predicted trajectory of the preceding vehicle and the vehicle's own dynamic stability.

[0009] As a further solution of the present invention: the specific steps of extracting the u-shapelets set in step S1 include:

[0010] Data from various typical road sections are selected to construct a training set for the deep learning model; the raw data of the data set is collected synchronously by laser radar, millimeter wave radar, and camera, and the raw data is output through image processing technology;

[0011] Set the subsequence length slen, and obtain all subsequences of length slen on the preceding vehicle driving process monitoring signal dataset D as candidate sets; secondly, calculate the distance sdist between each candidate S and each preceding vehicle driving process monitoring signal T in the preceding vehicle driving process monitoring signal dataset D, using the formula:

[0012] sdist(T,S)=min(dist(S,Ti));

[0013] Where Ti is a subsequence of length slen in the preceding vehicle driving process monitoring signal T, and i represents the starting position of the subsequence in the process monitoring signal T;

[0014] According to the preset segmentation point d, the time series data set D is divided into two subsets DL and DR; the distance data are sorted, and the gap value of the data set segmentation is calculated according to the midpoint of every two distance data, where the gap value with the largest value is the optimal segmentation point d.

[0015] As a further solution of the present invention: the specific steps of obtaining the distance matrix in step S1 and clustering the distance matrix include:

[0016] According to the obtained u-shapelets set, the distance sdist between each S and all process monitoring signals T in the driving process monitoring signal dataset D is extracted to obtain the distance matrix Dis;

[0017] Extract a subsequence of 3000 sampling points from the monitoring data within 3 seconds before each preceding vehicle, i.e., slen=3000, as a candidate set of u-shapelets, extract the set, calculate the distance between each u-shapelet and the process monitoring data in the data set, and obtain the distance matrix;

[0018] Use DBSCAN to cluster the distance matrix and get the clustering results;

[0019] The clustering results are divided into several types; Type 1 is when the distance between the two vehicles is less than the minimum lane change safety distance; Type 2 is when the distance between the two vehicles is greater than or equal to the maximum lane change safety distance; Type 3 is when the distance between the two vehicles is between the minimum lane change safety distance and the maximum lane change safety distance;

[0020] The monitoring signals of different driving safety trend processes are clustered according to the distance between the monitoring signals of the driving process of each preceding vehicle and the u-shapelets, that is, the preceding vehicles with different safety change rules are clustered to obtain the clustering results.

[0021] As a further solution of the present invention: the specific steps in step S2 include:

[0022] Step S21, by creating a random index of data, the first 80% of the data is used as a training set, and the last 20% of the data is used as a validation set. Then, the training data is normalized.

[0023] Step S22, initialize the hyperparameter set, construct the four-dimensional feature, use the multi-dimensional feature as input, use the future trajectory information as output, and use a variety of training hyperparameters to establish a data-driven model;

[0024] A preceding vehicle trajectory prediction model is constructed based on LSTM, with process monitoring signals as input. The model contains three LSTM layers, each with 512 nodes, and then connected with a Dropout layer to prevent the model from overfitting. The LSTM layer contains two fully connected layers and a RUL output layer. The two fully connected layers use linear rectification function and hyperbolic tangent function as activation functions, respectively, and the RUL output layer has no activation function.

[0025] According to the constructed front vehicle driving trajectory prediction model, the SGD optimizer is used to train it, and the training loss function is defined as the mean absolute error, and the formula is:

[0026]

[0027] In the formula, y i is the theoretical value of RUL; is the RUL prediction value; m is the number of training samples in each batch;

[0028] Step S23: randomly select multiple validation sets from all training data and perform cyclic accuracy verification on the training model. The best prediction model is obtained based on the calculated regression error;

[0029] Step S24: perform feature recursion based on historical feature data, substitute the best accuracy model, input the process monitoring signal into the preceding vehicle driving trajectory prediction model based on LSTM, and predict the preceding vehicle's future driving trajectory.

[0030] As a further solution of the present invention: the specific steps of the active collision avoidance strategy in step S3 include:

[0031] Based on the predicted trajectory of the preceding vehicle and the dynamic stability of the own vehicle; and

[0032] According to the longitudinal distance D between the adjacent leading vehicle in the future time period measured by the vehicle's own sensors and the active safety warning distance thresholds Dwmin and Dwmax, compare them to determine whether there is a collision risk, and give warnings and perform graded braking;

[0033] When D≥Dwmax, that is, the adjacent leading vehicle does not affect the normal driving of the vehicle itself, the active collision avoidance system does not intervene;

[0034] When Dwmin < D < Dwmax, the future driving trajectory of the adjacent leading vehicle will affect the driving safety of the vehicle itself, there is a collision risk, the active collision avoidance system gives a warning, and controls the vehicle itself to decelerate with the target deceleration;

[0035] When the on-vehicle sensor measures that the longitudinal distance D of the adjacent leading vehicle is ≤Dwmin, it is necessary to increase the braking intensity of the vehicle itself. At this time, the target deceleration is the maximum braking deceleration under the current road surface adhesion conditions.

[0036] Compared with the prior art, the present invention has the following technical effects:

[0037] Adopting the above technical solution, by extracting the u-shapelets set and calculating the distance matrix between it and the time series for clustering to identify vehicles with different driving safety trends; secondly, based on the clustering results and on-vehicle sensor data, using the long short-term memory network model to train and predict the driving trajectory of the leading vehicle; finally, according to the predicted driving trajectory of the leading vehicle and the dynamic stability of the vehicle itself, formulate an active collision avoidance strategy, including warning and graded braking, to ensure driving safety. Description of the Drawings

[0038] The following will describe in detail the specific implementation manners of the present invention in conjunction with the drawings:

[0039] Figure 1 It is a schematic diagram of the steps of the active collision avoidance method for an unmanned vehicle according to the disclosed embodiment of the present application;

[0040] Figure 2 It is a flow block diagram of the active collision avoidance method for an unmanned vehicle according to the disclosed embodiment of the present application. Specific Embodiment

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] Please refer to Figure 1 and Figure 2In an embodiment of the present invention, an unmanned vehicle active collision avoidance method based on preceding vehicle trajectory clustering prediction comprises the following steps:

[0043] Step S1, extracting the observation signal of the preceding vehicle during driving to obtain a u-shapelets set, calculating the distance between the u-shapelets set and the time series to obtain a distance matrix, and clustering the distance matrix to obtain a clustering result;

[0044] Step S11, the specific steps of extracting the u-shapelets set include:

[0045] Data from various typical road sections are selected to construct a training set for the deep learning model; the raw data of the data set is collected synchronously by laser radar, millimeter wave radar, and camera, and the raw data is output through image processing technology;

[0046] Set the subsequence length slen, and obtain all subsequences of length slen on the preceding vehicle driving process monitoring signal dataset D as candidate sets; secondly, calculate the distance sdist between each candidate S and each preceding vehicle driving process monitoring signal T in the preceding vehicle driving process monitoring signal dataset D, using the formula:

[0047] sdist(T,S)=min(dist(S,Ti));

[0048] Where Ti is a subsequence of length slen in the preceding vehicle driving process monitoring signal T, and i represents the starting position of the subsequence in the process monitoring signal T;

[0049] According to the preset segmentation point d, the time series data set D is divided into two subsets DL and DR; the distance data are sorted, and the gap value of the data set segmentation is calculated according to the midpoint of every two distance data, where the gap value with the largest value is the optimal segmentation point d.

[0050] Step S12, obtaining a distance matrix, and the specific steps of clustering the distance matrix include:

[0051] According to the obtained u-shapelets set, the distance sdist between each S and all process monitoring signals T in the driving process monitoring signal dataset D is extracted to obtain the distance matrix Dis;

[0052] Extract a subsequence of 3000 sampling points from the monitoring data within 3 seconds before each preceding vehicle, i.e., slen=3000, as a candidate set of u-shapelets, extract the set, calculate the distance between each u-shapelet and the process monitoring data in the data set, and obtain the distance matrix;

[0053] Step S13, clustering the distance matrix using DBSCAN to obtain a clustering result;

[0054] The clustering results are divided into several types; Type 1 is when the distance between the two vehicles is less than the minimum lane change safety distance; Type 2 is when the distance between the two vehicles is greater than or equal to the maximum lane change safety distance; Type 3 is when the distance between the two vehicles is between the minimum lane change safety distance and the maximum lane change safety distance;

[0055] The monitoring signals of different driving safety trend processes are clustered according to the distance between the monitoring signals of the driving process of each preceding vehicle and the u-shapelets, that is, the preceding vehicles with different safety change rules are clustered to obtain the clustering results.

[0056] Step S2: According to the obtained clustering results, the long and short-term memory network models are trained based on the data collected by various vehicle-mounted sensors to make predictions and obtain the driving trajectory of the preceding vehicle. The specific steps include:

[0057] Step S21, by creating a random index of data, the first 80% of the data is used as a training set, and the last 20% of the data is used as a validation set. Then, the training data is normalized.

[0058] Step S22, initialize the hyperparameter set, construct the four-dimensional feature, use the multi-dimensional feature as input, use the future trajectory information as output, and use a variety of training hyperparameters to establish a data-driven model;

[0059] A preceding vehicle trajectory prediction model is constructed based on LSTM, with process monitoring signals as input. The model contains three LSTM layers, each with 512 nodes, and then connected with a Dropout layer to prevent the model from overfitting. The LSTM layer contains two fully connected layers and a RUL output layer. The two fully connected layers use linear rectification function and hyperbolic tangent function as activation functions, respectively, and the RUL output layer has no activation function.

[0060] According to the constructed front vehicle driving trajectory prediction model, the SGD optimizer is used to train it, and the training loss function is defined as the mean absolute error, and the formula is:

[0061]

[0062] In the formula, y i is the theoretical value of RUL; is the RUL prediction value; m is the number of training samples in each batch;

[0063] Step S23: randomly select multiple validation sets from all training data, and perform cyclic accuracy verification on the training model. The calculated regression error is used as the standard to obtain the prediction model with the best accuracy;

[0064] Step S24: Perform feature recursion based on historical feature data, substitute it into the best accuracy model, input the process monitoring signal into the forward vehicle driving trajectory prediction model based on LSTM, and predict the future driving trajectory of the forward vehicle.

[0065] Step S3: Develop an active collision avoidance strategy based on the obtained predicted trajectory of the forward vehicle and the dynamic stability of the own vehicle;

[0066] Among them, the specific steps of the active collision avoidance strategy include:

[0067] Based on the obtained predicted trajectory of the forward vehicle and the dynamic stability of the own vehicle; and

[0068] Then, compare the longitudinal distance D between the own vehicle and the adjacent forward vehicle in the future time period measured by the vehicle sensor with the active safety warning distance thresholds Dwmin and Dwmax to determine whether there is a collision risk, and perform warning and graded braking;

[0069] When D ≥ Dwmax, that is, the adjacent forward vehicle does not affect the normal driving of the own vehicle, the active collision avoidance system does not intervene;

[0070] When Dwmin < D < Dwmax, the future driving trajectory of the adjacent forward vehicle will affect the driving safety of the own vehicle, there is a collision risk, the active collision avoidance system gives a warning, and controls the own vehicle to decelerate with the target deceleration;

[0071] When the vehicle sensor measures that the longitudinal distance D of the adjacent forward vehicle ≤ Dwmin, it is necessary to increase the braking intensity of the own vehicle, and the target deceleration at this time is the maximum braking deceleration under the current road surface adhesion condition.

[0072] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents, and all should be included within the protection scope of the present invention.

Claims

1. An active collision avoidance method for unmanned vehicles based on cluster prediction of leading vehicle trajectories, characterized in that: The following steps are involved: Step S1, extracting the observation signal of the preceding vehicle during driving to obtain a u-shapelets set, calculating the distance between the u-shapelets set and the time series to obtain a distance matrix, and clustering the distance matrix to obtain a clustering result; Step S2: According to the obtained clustering results, the long and short-term memory network models are trained based on the data collected by various vehicle-mounted sensors to make predictions and obtain the driving trajectory of the preceding vehicle; Step S3: formulate an active collision avoidance strategy based on the predicted trajectory of the preceding vehicle and the vehicle's own dynamic stability.

2. The method for active collision avoidance of unmanned vehicles based on cluster prediction of leading vehicle trajectories according to claim 1, characterized in that: The specific steps of extracting the u-shapelets set in step S1 include: Data from various typical road sections are selected to construct a training set for the deep learning model; the raw data of the data set is collected synchronously by laser radar, millimeter wave radar, and camera, and the raw data is output through image processing technology; Set the subsequence length slen, and obtain all subsequences of length slen on the preceding vehicle driving process monitoring signal dataset D as candidate sets; secondly, calculate the distance sdist between each candidate S and each preceding vehicle driving process monitoring signal T in the preceding vehicle driving process monitoring signal dataset D, using the formula: sdist(T,S)=min(dist(S,Ti)); Where Ti is a subsequence of length slen in the preceding vehicle driving process monitoring signal T, and i represents the starting position of the subsequence in the process monitoring signal T; According to the preset segmentation point d, the time series data set D is divided into two subsets DL and DR; the distance data are sorted, and the gap value of the data set segmentation is calculated according to the midpoint of every two distance data, where the gap value with the largest value is the optimal segmentation point d.

3. The method for active collision avoidance of unmanned vehicles based on cluster prediction of leading vehicle trajectories according to claim 2, characterized in that: The specific steps of obtaining the distance matrix in step S1 and clustering the distance matrix include: According to the obtained u-shapelets set, the distance sdist between each S and all process monitoring signals T in the driving process monitoring signal dataset D is extracted to obtain the distance matrix Dis; Extract a subsequence of 3000 sampling points from the monitoring data within 3 seconds before each preceding vehicle, i.e., slen=3000, as a candidate set of u-shapelets, extract the set, calculate the distance between each u-shapelet and the process monitoring data in the data set, and obtain the distance matrix; Use DBSCAN to cluster the distance matrix and get the clustering results; The clustering results are divided into several types; Type 1 is when the distance between the two vehicles is less than the minimum lane change safety distance; Type 2 is when the distance between the two vehicles is greater than or equal to the maximum lane change safety distance; Type 3 is when the distance between the two vehicles is between the minimum lane change safety distance and the maximum lane change safety distance; The monitoring signals of different driving safety trend processes are clustered according to the distance between the monitoring signals of the driving process of each preceding vehicle and the u-shapelets, that is, the preceding vehicles with different safety change rules are clustered to obtain the clustering results.

4. The method for active collision avoidance of unmanned vehicles based on cluster prediction of leading vehicle trajectories according to claim 1, characterized in that: The specific steps in step S2 include: Step S21: By creating a random data index, the first 80% of the data is used as the training set, and the last 20% of the data is used as the validation set. Then, the training data is normalized. Step S22: Initialize the hyperparameter set, construct four-dimensional features, use the multi-dimensional features as the input, and the future trajectory information as the output. Establish a data-driven model using a variety of training hyperparameters; Based on LSTM, construct a prediction model for the leading vehicle's driving trajectory. Use the process monitoring signal as the input. The model contains three LSTM layers, with 512 nodes in each layer, and then connect a Dropout layer to prevent the model from overfitting; after the LSTM layers, there are two fully connected layers and a RUL output layer. The two fully connected layers use the rectified linear unit function and the hyperbolic tangent function as activation functions respectively, and the RUL output layer has no activation function; According to the constructed prediction model for the leading vehicle's driving trajectory, use the SGD optimizer to train it. The loss function for training is defined as the mean absolute error, and the formula is: In the formula, y i is the theoretical value of RUL; is the RUL prediction value; m is the number of training samples in each batch; Step S23: Randomly select multiple groups of validation sets from all the training data, and perform cyclic accuracy verification on the trained model. Taking the calculated regression error as the standard, obtain the prediction model with the optimal accuracy; Step S24: Perform feature recursion based on the historical feature data, substitute it into the model with the best accuracy, input the process monitoring signal into the prediction model for the leading vehicle's driving trajectory based on LSTM, and predict the future driving trajectory of the leading vehicle.

5. The method for active collision avoidance of unmanned vehicles based on cluster prediction of leading vehicle trajectories according to claim 1, characterized in that: The specific steps of the active collision avoidance strategy in step S3 include: According to the predicted trajectory of the leading vehicle obtained and the dynamic stability of its own vehicle; and Then, compare the longitudinal distance D between the adjacent leading vehicle in the future time period measured by the vehicle's own sensor with the active safety warning distance thresholds Dwmin and Dwmax to determine whether there is a collision risk, and issue a warning and perform hierarchical braking; When D≥Dwmax, that is, the adjacent leading vehicle does not affect the normal driving of its own vehicle, the active collision avoidance system does not intervene; When Dwmin < D < Dwmax, the future driving trajectory of the adjacent leading vehicle will affect the driving safety of its own vehicle, and there is a collision risk. The active collision avoidance system issues a warning and controls its own vehicle to decelerate with the target deceleration; When the on-vehicle sensor measures that the longitudinal distance D of the adjacent leading vehicle is ≤ Dwmin, it is necessary to increase the braking intensity of its own vehicle. At this time, the target deceleration is the maximum braking deceleration under the current road surface adhesion conditions.

Images