Driver identification method and system based on driving style, terminal and storage medium

Abstract

The invention discloses a driver identification method and system based on a driving style, a terminal and a storage medium, and the method comprises the steps: dividing driving track data into a plurality of sub-track sequences based on a first sliding window; dividing each sub-track into a plurality of overlapped segments based on the second sliding window; based on a driving state index table, converting the driving track data in the segments into a plurality of continuous driving states, and forming a state conversion characteristic matrix of the segments by conversion intensity between every two continuous driving states; based on a statistical method, carrying out statistics on dynamic characteristics of the driving track data in the fragments, and constructing a motion statistical characteristic matrix by using statistical magnitude; fusing the state transition feature matrix and the motion statistical feature matrix according to a time step by adopting a parallel method to obtain a complex matrix; and establishing a driver classification network, training the driver classification network by taking the complex matrix and the driver identification tag as training samples, and then carrying out driver identification. The short-term driver recognition performance is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent transportation, and particularly relates to a driver recognition method, system, terminal and storage medium based on driving style. Background Art

[0002] Driver recognition can be used in practical scenarios such as insurance claims, vehicle anti-theft, and fleet management to improve driver driving behavior, reduce driving risks and operating costs. In the prior art, driver recognition based on methods such as face and fingerprint has achieved good recognition results. However, the information such as the face and fingerprint of the driver obtained by the prior art is limited in application because it is easy to violate the privacy and will of the driver. On the contrary, automotive sensor data provides fingerprint-like data that can reflect the unique driving style of the driver through data-driven methods, which helps to achieve driver recognition under the premise of privacy protection.

[0003] In the prior art, when data-driven methods are used for driver recognition, since the same driver has multiple variable-length GPS trajectories, their driving patterns and characteristics cannot be accurately obtained, and different feature sets need to be used to characterize and identify the driver. Therefore, descriptive definitions of features are proposed based on domain expert knowledge, resulting in the extracted features lacking universality and being unable to extract data features that can characterize driving activities and their corresponding spatio-temporal dynamics from sensor data, leading to driver recognition errors; in addition, the method of learning features from low-level motion statistical matrices based on convolutional neural networks and long short-term memory, this kind of movement statistics (MS) can only achieve the extraction of motion features and cannot achieve the time dependence of the vehicle state transition (ST) relationship, resulting in trucks and cars with similar motion trajectories being recognized as the same result; the method of learning driver features from the duration and frequency of vehicle state transitions cannot distinguish the intensity of adjacent state changes. For example, moderate braking and strong braking are regarded as the same braking event and cannot be effectively distinguished. To solve the above problems, the prior art also proposes the need for additional data collection devices and remote communication devices, such as inertial measurement units or cameras, resulting in increased costs and being unable to cover the driving patterns of different types of vehicles, making it difficult to apply to practical scenarios. Summary of the Invention

[0004] To solve the deficiencies in the prior art, the present invention provides a driver recognition method, system, terminal and storage medium based on driving style, which learns the driving style only from the currently available GPS trajectories of the vehicle to improve the short-term driver recognition performance.

[0005] The present invention adopts the following technical solutions.

[0006] The present invention proposes a driver recognition method based on driving style, which collects the driving trajectory data of a vehicle, including: speed, acceleration, jerk, direction angle, angular velocity and angular acceleration at time t; including:

[0007] Based on the first sliding window, the collected driving trajectory data is divided into multiple sub-trajectory sequences; based on the second sliding window, each sub-trajectory is divided into multiple overlapping segments;

[0008] Based on the driving state index table, the driving trajectory data in the segment is converted into multiple consecutive driving states, and the conversion intensity between every two consecutive driving states forms the state transition feature matrix of the segment; based on the statistical method, the dynamic characteristics of the driving trajectory data in the segment are statistically analyzed, and the statistical quantities obtained are used to construct the motion statistical feature matrix;

[0009] Using a parallel method, the state transition feature matrix and the motion statistical feature matrix are fused according to the time step to obtain a complex matrix;

[0010] A driver classification network is established, and the complex matrix and driver recognition labels are used as training samples to train the driver classification network; the trained driver classification network is used for driver recognition based on driving style.

[0011] The recognition accuracy when the second sliding window has different time lengths is recognized by using the trained driver classification network, and each sub-trajectory is divided into multiple overlapping segments according to the time length of the second sliding window corresponding to the maximum recognition accuracy;

[0012] Wherein, the time length of the second sliding window is less than the time length of the first sliding window.

[0013] There is a chronological order between two adjacent sub-trajectory sequences in time, and there is an overlapping relationship between two adjacent segments in time; wherein, the length of the overlapping part of two adjacent segments is half of the time length of the second sliding window.

[0014] Based on the driving state index table, the driving trajectory data in the segment is converted into multiple consecutive driving states, and the conversion intensity between every two consecutive driving states forms the state transition feature matrix of the segment, including:

[0015] Based on the driving state index table, the driving trajectory data in the segment is converted into multiple consecutive driving states;

[0016] The conversion intensity between every two consecutive driving states forms the state transition feature matrix of the segment.

[0017] Based on the rule method, a driving state index table is established using speed and direction angle, including:

[0018] 1) v t >v t-1 And b t >b t-1 Then the driving state at time t is accelerating right turn;

[0019] 2) v t >v t-1 And b t <b t-1 Then the driving state at time t is accelerating to turn left;

[0020] 3) v t >v t-1 and|b t -b t-1 |<Δb, the driving state at time t is accelerating straight ahead; Δb is the set direction angle threshold;

[0021] 4) v t <v t-1 And b t >b t-1 Then the driving state at time t is decelerating and turning right;

[0022] 5) v t <v t-1 And b t <b t-1 Then the driving state at time t is decelerating and turning left;

[0023] 6) v t <v t-1 and|b t -b t-1 |<Δb, then the driving state at time t is decelerating and driving straight;

[0024] 7) |v t -v t-1 |<Δv and b t >b t-1 The driving state at time t is a uniform right turn; Δv is the set speed threshold;

[0025] 8) |v t -v t-1 |<Δv and b t <b t-1 Then the driving state at time t is a uniform left turn;

[0026] 9) |v t -v t-1 |<Δv and |b t -b t-1 |<Δb, then the driving state at time t is straight ahead at a constant speed;

[0027] Among them, vt , v t-1 are the speeds at times t and t - 1, and b t , b t-1 are the direction angles at times t and t - 1.

[0028] For every two consecutive driving states, the transition intensity between each pair of consecutive driving states is constructed based on the speed difference, azimuth angle difference, and the historical transition relationship of the states, satisfying the following relational expression:

[0029]

[0030] In the formula,

[0031] s i , s j are respectively the source state and the target state in each pair of consecutive driving states;

[0032] I(s i , s j ) is the transition intensity from the source state to the target state;

[0033] s i (v), s j (v) are respectively the speed of the source state and the speed of the target state; (s i (v) - s j (v)) 2 is the speed difference;

[0034] s i (b), s j (b) are respectively the direction angle of the source state and the direction angle of the target state; (s i (b) - s j (b)) 2 is the azimuth angle difference;

[0035] m is the historical transition relationship between the source state and the target state in each pair of consecutive driving states. When the source state and the target state have the same historical transition relationship, the value is 1; otherwise, the value is 0.

[0036] Using a parallel method, the state transition feature matrix and the motion statistical feature matrix are fused according to the time step to obtain a complex matrix, including:

[0037] Taking time as the dimension of fusion, the state transition feature and the motion statistical feature form a feature complex vector, and the complex matrix is formed using the feature complex vector;

[0038] Among them, the feature complex vector satisfies the following relational expression:

[0039]

[0040] In the formula, is a characteristic complex vector, \(m_t\) is a motion statistical feature, \(s_t\) is a state transition feature, and \(i\) is the imaginary unit.

[0041] The driver classification network includes a residual RNN encoder, a residual RNN decoder, and a prediction head;

[0042] The residual RNN encoder contains two recurrent blocks, the residual RNN decoder contains two recurrent blocks symmetric to the encoder, and the prediction head includes a fully connected layer.

[0043] In the residual RNN encoder, the first recurrent block maps the input of a given shape to a ReLu activation sequence with the same shape as the input; the second recurrent block maps the addition output to a compressed state sequence.

[0044] The present invention also proposes a driver recognition system based on driving style, including: an acquisition module, a sample construction module, a network training module, an identification module, and a time length adjustment module;

[0045] The acquisition module is used to acquire the driving trajectory data of the vehicle;

[0046] The sample construction module is used to divide the acquired driving trajectory data into multiple sub-trajectory sequences based on the first sliding window; divide each sub-trajectory into multiple overlapping segments based on the second sliding window; convert the driving trajectory data in the segment into multiple consecutive driving states based on the driving state index table, and construct a state transition feature matrix of the segment with the transition intensity between every two consecutive driving states; statistically analyze the dynamic characteristics of the driving trajectory data in the segment based on a statistical method to obtain a motion statistical feature matrix; use a parallel method to fuse the state transition feature matrix and the motion statistical feature matrix according to the time step to obtain a complex matrix;

[0047] The network training module is used to establish a driver classification network and train the driver classification network with the complex matrix and the driver recognition label as training samples;

[0048] The identification module is used to perform driver identification based on driving style by using the trained driver classification network;

[0049] The time length adjustment module is used to identify the recognition accuracy when the second sliding window has different time lengths by using the trained driver classification network, and send the time length of the second sliding window corresponding to the maximum recognition accuracy to the sample construction module.

[0050] The present invention also proposes a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the method.

[0051] A computer-readable storage medium storing a computer program which, when executed by a processor, implements the steps of the method.

[0052] The beneficial effects of the present invention are at least as follows compared with the prior art. Based on the fusion mode, the present invention obtains a fusion matrix of a state transition matrix and a motion statistics matrix, which is more robust for characterizing the driving behaviors of car and truck drivers. In the driver classification network, the necessity of further learning a strong driving style representation is realized through residual connections and the regularization of the autoencoder, improving the recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 is a flowchart of the driver recognition method based on driving style proposed by the present invention;

[0054] Figure 2 is the average recognition accuracy (%) of car driving data in different modes in the embodiments of the present invention, Figure 2 (a) is for recognizing the driving data of 5 car drivers, Figure 2 (b) is for recognizing the driving data of 10 car drivers; in the figure, ResGRUARNet and ResLSTMARNet are two different neural networks, and the left bar chart corresponding to each mode is the output of ResGRUARNet, and the right bar chart is the output of ResLSTMARNet;

[0055] Figure 3 is the average recognition accuracy (%) of truck driving data in different modes in the embodiments of the present invention, Figure 3 (a) is for recognizing the driving data of 5 truck drivers, Figure 3 (b) is for recognizing the driving data of 10 truck drivers; in the figure, GRU and LSTM are two different neural networks, and the left bar chart corresponding to each mode is the output of GRU, and the right bar chart is the output of LSTM;

[0056] Figure 4 is a structural diagram of the driver recognition system based on driving style proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. The embodiments described in this application are only a part of the embodiments of the present invention, rather than all embodiments. Based on the spirit 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.

[0058] As shown Figure 1 in the figure, the driver recognition method based on driving style proposed by the present invention includes:

[0059] Step 1, collect the GPS driving trajectory data of the vehicle;

[0060] Specifically, the GPS driving trajectory data generated during the vehicle driving is represented as Tr = <p1, p2, …, p i , …>, where p i is the i-th sampling data of the GPS driving trajectory.

[0061] Specifically, derive the speed v t , acceleration a t , jerk j t , direction angle b t , angular velocity ba t and angular acceleration bj t at time t according to the GPS driving trajectory data. Therefore, the i-th sampling data of the GPS driving trajectory satisfies p i = {v t , a t , j t , b t , ba t , bj t}, enabling the sampling data to represent driving behavior and driving style in multiple dimensions.

[0062] Due to the growing privacy concerns about the exposure of location information, the present invention derives speed and direction as the main information sources based on the GPS driving trajectory data, rather than directly using GPS coordinates.

[0063] Step 2, based on the first sliding window, divide the collected GPS driving trajectory data into multiple sub-trajectory sequences; based on the second sliding window, divide each sub-trajectory into multiple overlapping segments.

[0064] Among them, the length of the overlapping part of two adjacent segments is half of the time length of the second sliding window.

[0065] The time length of the second sliding window is less than the time length of the first sliding window.

[0066] Specifically, the time length L s of the first sliding window is preferably 256 s.

[0067] Specifically, when the time length L f of the second sliding window is preferably 16 s, the i-th sampling data {v t , a t , j t , bt , ba t , bj t} It has 6 dimensions. Therefore, each segment is a matrix with 6 rows and 16 columns. The size of the segment is jointly determined by the sampling data dimension and the time length of the second sliding window.

[0068] There is a sequential order in time between two adjacent sub-trajectory sequences, and there is an overlapping relationship in time between two adjacent overlapping sub-trajectory sequences.

[0069] In the embodiment, in the first stage, the variable-length journey of any driver will be divided into several short-term sub-trajectories. After accurately identifying the sub-trajectories, the correct driver ID is predicted. Therefore, by using a sliding window, any sub-trajectory is further segmented, and obtaining multiple overlapping segments is beneficial to capturing driving style rules and improving the recognition accuracy.

[0070] In the prior art, a single sliding window method is generally used to obtain a driving segment sequence, and then a recurrent neural network is used to capture the spatio-temporal dependence relationship between segments. However, the size of the sliding window is usually an empirical value. An inappropriate size of the motion segment loses its state transition and statistical features on the one hand, and affects the classification performance of the subsequent network model on the other hand.

[0071] Driving behaviors are interdependent in time. Therefore, from the perspective of the time window, when obtaining trajectories, the sub-trajectory sequences obtained based on the first sliding window have a sequential order in time, and the overlapping segments obtained based on the second sliding window have a certain degree of correlation. Therefore, it can better identify driving states and motions, and emphasizes the relevance of previous driving behaviors to subsequent driving states and motions, thereby improving the accuracy of driving recognition.

[0072] Step 3: Based on the driving state index table, convert the GPS driving trajectory data in the segment into multiple consecutive driving states, and use the conversion intensity between every two consecutive driving states to form the state transition feature matrix of the segment.

[0073] Specifically, Step 3 includes:

[0074] Step 3.1: Based on the driving state index table, convert the GPS driving trajectory data in the segment into multiple consecutive driving states;

[0075] Specifically, based on the rule method, a driving state index table is established using speed and direction angle, including:

[0076] 1), v t > v t-1 and b t > b t-1 Then the driving state at time t is accelerating right turn;

[0077] 2), v t > v t-1 and b t < b t-1 Then the driving state at time t is accelerating left turn;

[0078] 3), v t > v t-1 and |b t - b t-1 | < Δb Then the driving state at time t is accelerating straight; Δb is a set direction angle threshold, preferably 10°;

[0079] 4), v t < v t-1 and b t > b t-1 Then the driving state at time t is decelerating right turn;

[0080] 5), v t < v t-1 and b t < b t-1 Then the driving state at time t is decelerating left turn;

[0081] 6), v t < v t-1 and |b t - b t-1 | < Δb Then the driving state at time t is decelerating straight;

[0082] 7), |v t - v t-1 | < Δv and b t > b t-1 Then the driving state at time t is uniform speed right turn; Δv is a set speed threshold, preferably 1 km / h;

[0083] 8), |v t - v t-1 | < Δv and b t < b t-1 Then the driving state at time t is uniform speed left turn;

[0084] 9), |v t - v t-1 | < Δv and |b t - b t-1 | < Δb Then the driving state at time t is uniform speed straight.

[0085] In the embodiment, based on the driving state index table shown in Table 1, the driving states of the vehicle are obtained to form a driving state sequence, which is [<Accelerating, Going straight>, <Constant speed, Going straight>, ……, <Decelerating, Turning right>]. The conversion based on the driving state index table is simple and easy to operate, and can quickly and accurately obtain the driving state sequence from the trajectory point sequence.

[0086] Table 1 Driving state index table

[0087] Index Status 1 Accelerated right turn 2 Accelerated left turn 3 Accelerated straight ahead 4 Decelerated right turn 5 Decelerated left turn 6 Decelerated straight ahead 7 Constant-speed right turn 8 Constant-speed left turn 9 Constant-speed straight ahead

[0088] Step 3.2, the state transition feature matrix of the segment is formed by the transition intensity between every two consecutive driving states.

[0089] For every two consecutive driving states, the transition intensity between each pair of consecutive driving states is constructed based on the speed difference, azimuth angle difference, and the historical transition relationship of the states, satisfying the following relational expression:

[0090]

[0091] In the formula,

[0092] s i ,s j are respectively the source state and the target state in each pair of consecutive driving states;

[0093] I(s i ,s j ) is the transition intensity from the source state to the target state;

[0094] s i (v), s j (v) are respectively the speed of the source state and the speed of the target state; (s i (v) - s j (v)) 2 is the speed difference;

[0095] s i (b), s j (b) are respectively the direction angle of the source state and the direction angle of the target state; (s i (b) - s j (b)) 2 is the azimuth angle difference;

[0096] m is the historical transition relationship between the source state and the target state in each pair of consecutive driving states. When the source state and the target state have the same historical transition relationship, the value is 1, otherwise the value is 0.

[0097] In the present invention, the conversion intensity not only considers the changes in speed and direction, but also considers whether there is the same historical conversion relationship between two consecutive driving states, so as to be able to distinguish different drivers under the same state conversion event.

[0098] Furthermore, for two consecutive driving states with the same source state and target state, the average value of the conversion intensity between the two consecutive driving states constitutes the state conversion feature matrix of the segment.

[0099] Furthermore, the MaxMinScaler method is used to normalize the conversion intensity, and the normalized conversion intensity constitutes the state conversion feature matrix of the segment; in the embodiment, the state conversion feature matrix ST is a 9×9 matrix, and each element in the matrix is between 0 and 1, as shown in Table 2.

[0100] Table 2 State conversion feature matrix

[0101]

[0102] Step 4: Based on the statistical method, statistically analyze the dynamic characteristics of each GPS driving trajectory data in the segment, and construct a motion statistical feature matrix with the obtained statistical quantities.

[0103] Specifically, the statistical quantities include: average value, maximum value, minimum value, 25% quantile, 50% quantile, 75% quantile, and standard deviation, as shown in Table 3.

[0104] Table 3 Statistical quantity table

[0105]

[0106]

[0107] In the embodiment, therefore, the time length L of the second sliding window f is preferably 16 s, and the i-th sampled data {v t , a t , j t , b t , ba t , bj t} of the GPS driving trajectory has 6 signal dimensions, and 7 statistical quantities are derived from each dimension of the signal. Therefore, the motion statistical feature matrix MS of a single segment is a 6×7 matrix, and the size of this matrix is related to the signal dimension of the sampled data and the number of statistical quantities derived from each signal.

[0108] Step 5: Adopt a parallel method to fuse the state conversion feature matrix and the motion statistical feature matrix according to the time step to obtain a complex matrix.

[0109] The state transition feature matrix ST and the motion statistics feature matrix MS each have limitations due to partial loss of information in the motion segments. Therefore, the two modes need to be fused to more comprehensively describe the driving segments. The two matrices corresponding to the two modes are combined as local features and will be fed into the proposed network to learn the driving style representation.

[0110] Specifically, step 5 includes: taking time as the fusion dimension, the state transition feature and the motion statistics feature form a feature complex vector, and a complex matrix is formed using the feature complex vector. Among them, the feature complex vector satisfies the following relational expression:

[0111]

[0112] In the formula, is the feature complex vector, mt is the motion statistics feature, st is the state transition feature, and i is the imaginary unit.

[0113] Among them, the state transition feature is obtained based on the speed difference and the azimuth angle difference, and the speed difference and the azimuth angle difference are significant features recognized by the driver. Using a parallel method is beneficial to improving the recognition rate of significant features on the basis of motion statistics features, and moreover, no other additional parameters (such as weights) need to be introduced, saving parameters and computational complexity.

[0114] The complex matrix obtained through fusion can provide more information than any single matrix alone. For example, using the average speed, we can determine whether a constant driving state reflects high speed. Statistics cannot show rapid driving behaviors such as sudden braking, which can be reflected by the transition intensity of adjacent points.

[0115] The present invention converts the original GPS data into a low-level movement mode, and preprocesses the variable-length driving trajectory to generate a series of sub-trajectories, where each sub-trajectory encodes the short-term driving style of the driver to be predicted. Then, any sub-trajectory is divided into kinematic segments of a fixed length, and each kinematic segment is transformed into a local feature vector. The local features integrate the semantic and statistical information of the ST and MS modes to improve the ability of the developed model to represent the driving style.

[0116] Step 6, establish a driver classification network, and use the fused feature matrix and the correct driver identification label as training samples to train the driver classification network.

[0117] Fuse the state transition matrix and the motion statistics matrix to more comprehensively describe the driving segment. The combination of the two matrices corresponding to the two modes is used as a local feature and will be fed into the proposed network to learn the driving style representation. The two matrices need to be unfolded into vectors for fusion. This fusion will provide more information than either mode alone. For example, using the average speed, we can determine whether a constant driving state reflects high speed. Statistics cannot show rapid driving behaviors such as sudden braking, which can be reflected by the transition intensity between adjacent points. Finally, the sequential fusion mode and the correct driver labels at the sub-trajectory level are used as training samples for training the driver recognition model.

[0118] Specifically, the driver classification network includes a residual RNN encoder, a residual RNN decoder, and a prediction head.

[0119] Specifically, the residual RNN encoder utilizes the sequentiality of the input data in time to encode the dependencies between the state transition matrix and the motion statistics matrix. The residual RNN encoder contains two recurrent blocks. The first recurrent block maps the input of a given shape to a sequence of ReLu activations with the same shape as the input. Then, the second recurrent block maps the additive output to a compressed state sequence. The last state vector in the compressed state sequence is the driver recognition result, similar to the digital fingerprint of the sub-trajectory. Among them, the residual RNN decoder contains two recurrent blocks symmetric to the encoder.

[0120] The prediction head includes a fully connected layer.

[0121] Step 7, use the trained driver classification network for driver recognition based on driving style.

[0122] In the embodiment, to solve the 5-driver and 10-driver recognition problems on two datasets, two datasets are selected, from passenger cars and logistics trucks respectively. For driving trajectory segmentation, the time length L of the first sliding window s is set to 60 seconds (1 minute). It should be noted that due to the different trip lengths of the selected drivers, the preprocessed data is unbalanced. Drivers with longer trips produce more training samples. The dimension of the driving style embedding is fixed at 100, the batch size is 256, the maximum number of iterations is set to 1500, and other hyperparameters are correctly adjusted through extensive experiments. 5-fold cross-validation is used, and the validation is repeated 5 times to produce an average accuracy metric. For each fold of training, 15% of the training data is reserved for validation and early stopping.

[0123] Let ResGRUARNet and ResLSTMARNet receive inputs with MS mode, ST mode, and MS+ST mode, and compare the recognition accuracies. Figure 2 (a) and Figure 2(b) The bar chart shows the average classification accuracy of car driving data. There is a significant gap between the bars of the MS mode and the ST mode. Among the MS mode and the ST mode, Figure 2 The MS mode shows outstanding performance in the 5-driver recognition task shown in (a), but Figure 2 The ST mode shows outstanding performance in the 10-driver recognition task shown in (b), while Figure 2 (a) and Figure 2 Both show good performance in the recognition tasks trained with the MS+ST mode in (b), even better than the recognition results of single training with the MS mode or the ST mode. Figure 3 The comparison results of truck driving data are shown, Figure 3 (a) and Figure 3 In both (a) and (b), among the MS mode and the ST mode, the MS mode shows outstanding performance. Figure 3 (a) and Figure 3 Both show good performance in the recognition tasks trained with the MS+ST (complex matrix) mode in (b). Figure 2 And Figure 3 show that the driving behaviors between trucks and cars are different. When applying the learned model to the datasets of different types of vehicles, using the MS model or the ST mode alone may lead to poor recognition performance. Therefore, by fusing ST and MS together, a more stable and powerful local representation ability can be obtained. It can be seen that the fusion of the ST and MS modes effectively captures the segment-level driving patterns of different drivers. The mode fusion is more robust for the two datasets.

[0124] The state transition feature matrix ST is a 9×9 matrix, and the size of the motion statistics feature matrix MS is related to the signal dimension of the sampled data, the number of statistics derived from each signal, and the time length of the second sliding window. It can be seen that if the computational amount is to be optimized, on the premise that the signal dimension of the sampled data and the number of statistics derived from each signal are both determined, the time length of the second sliding window needs to be optimized.

[0125] Use the trained driver classification network to identify the recognition accuracy when the second sliding window has different time lengths, and use the time length of the second sliding window corresponding to the maximum recognition accuracy to divide each sub-trajectory into multiple overlapping segments; among them, the time length of the second sliding window is less than the time length of the first sliding window.

[0126] Based on the fusion mode, a fusion matrix of the state transition matrix and the motion statistics matrix is obtained, which is more robust for characterizing the driving behaviors of car and truck drivers. In the driver classification network, the necessity of further learning strong driving style representations through residual connections and the regularization of autoencoders improves the recognition performance.

[0127] The present invention also provides a driver recognition system based on driving style, as Figure 4 shown, including: a collection module, a sample construction module, a network training module, an identification module, and a time length adjustment module;

[0128] The collection module is used to collect GPS driving trajectory data of the vehicle;

[0129] The sample construction module is used to divide the collected GPS driving trajectory data into multiple sub-trajectory sequences based on the first sliding window; divide each sub-trajectory into multiple overlapping segments based on the second sliding window; convert each GPS driving trajectory data in the segment into multiple consecutive driving states based on the driving state index table, and use the conversion intensity between every two consecutive driving states to form the state conversion feature matrix of the segment; statistically analyze the dynamic characteristics of each GPS driving trajectory data in the segment based on statistical methods to construct a motion statistical feature matrix with the obtained statistical quantities; use a parallel method to fuse the state conversion feature matrix and the motion statistical feature matrix according to the time step to obtain a complex matrix;

[0130] The network training module is used to establish a driver classification network and train the driver classification network with the complex matrix and the driver recognition label as training samples;

[0131] The identification module is used to perform driver recognition based on driving style by using the trained driver classification network;

[0132] The time length adjustment module is used to identify the recognition accuracy when the second sliding window has different time lengths by using the trained driver classification network, and send the time length of the second sliding window corresponding to the maximum recognition accuracy to the sample construction module.

[0133] The present disclosure may be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for causing a processor to implement various aspects of the present disclosure.

[0134] A computer-readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. A computer-readable storage medium may be, for example—but not limited to—an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device such as a punched card or raised structures in grooves having instructions stored thereon, and any suitable combination of the foregoing. The computer-readable storage medium as used herein is not construed to be a transitory signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagated through a waveguide or other transmission medium (e.g., an optical pulse through an optical fiber cable), or an electrical signal transmitted through a wire.

[0135] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to various computing / processing devices, or can be downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network may include a copper transmission cable, an optical fiber transmission, a wireless transmission, a router, a firewall, a switch, a gateway computer, and / or an edge server. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in a computer-readable storage medium in each computing / processing device.

[0136] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine - related instructions, microcode, firmware instructions, state - setting data, or source code or object code written in any combination of one or more programming languages, including object - oriented programming languages such as Smalltalk, C++, etc., and conventional procedural programming languages such as the "C" language or similar programming languages. The computer - readable program instructions may be executed entirely on the user's computer, partially on the user's computer, executed as a stand - alone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or it may be connected to an external computer (e.g., by using an Internet service provider to connect through the Internet). In some embodiments, by using the state information of the computer - readable program instructions to customize an electronic circuit, such as a programmable logic circuit, a field - programmable gate array (FPGA), or a programmable logic array (PLA), the electronic circuit can execute the computer - readable program instructions to implement various aspects of the present disclosure.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications or equivalent substitutions can still be made to the specific embodiments of the present invention, and any modification or equivalent substitution that does not depart from the spirit and scope of the present invention should be covered by the protection scope of the claims of the present invention.

Claims

1. A driver recognition method based on driving style, which collects the driving trajectory data of a vehicle, including: The speed, acceleration, jerk, direction angle, angular velocity, and angular acceleration at time t; it is characterized by including: Based on a first sliding window, dividing the collected driving trajectory data into multiple sub-trajectory sequences; based on a second sliding window, dividing each sub-trajectory into multiple overlapping segments; Based on a driving state index table, converting the driving trajectory data in the segment into multiple consecutive driving states, and using the conversion intensity between every two consecutive driving states to form the state transition feature matrix of the segment; based on a statistical method, statistically analyzing the dynamic characteristics of the driving trajectory data in the segment to construct a motion statistical feature matrix with the obtained statistics; Adopting a parallel method, fusing the state transition feature matrix and the motion statistical feature matrix according to the time step to obtain a complex matrix; Establishing a driver classification network, using the complex matrix and driver identification labels as training samples to train the driver classification network; using the trained driver classification network for driver identification based on driving style.

2. The driver identification method based on driving style according to claim 1, characterized in that: Using the trained driver classification network to identify the recognition accuracy when the second sliding window has different time lengths, and dividing each sub-trajectory into multiple overlapping segments with the time length of the second sliding window corresponding to the maximum recognition accuracy; Wherein, the time length of the second sliding window is less than the time length of the first sliding window.

3. The driver identification method based on driving style according to claim 2, characterized in that: There is a chronological order between two adjacent sub-trajectory sequences in time, and there is an overlapping relationship between two adjacent segments in time; wherein, the length of the overlapping part of two adjacent segments is half of the time length of the second sliding window.

4. The driver identification method based on driving style according to claim 1, characterized in that: Based on a driving state index table, converting the driving trajectory data in the segment into multiple consecutive driving states, and using the conversion intensity between every two consecutive driving states to form the state transition feature matrix of the segment, including: Based on a driving state index table, converting the driving trajectory data in the segment into multiple consecutive driving states; Using the conversion intensity between every two consecutive driving states to form the state transition feature matrix of the segment.

5. The driver identification method based on driving style according to claim 4, characterized in that: Based on a rule method, using speed and direction angle to establish a driving state index table, including: 1), v t > v t-1 and b t > b t-1 Then the driving state at time t is accelerating right turn; 2), v t > v t-1 and b t < b t-1 Then the driving state at time t is accelerating left turn; 3), v t > v t-1 and |b t - b t-1 |< Δb, then the driving state at time t is accelerating straight ahead; Δb is the set direction angle threshold; 4), v t < v t-1 and b t > b t-1 Then the driving state at time t is decelerating right turn; 5), v t < v t-1 and b t < b t-1 Then the driving state at time t is decelerating left turn; 6), v t < v t-1 and |b t - b t-1 | < Δb, then the driving state at time t is decelerating straight ahead; 7), |v t -v t-1 | < Δv and b t > b t-1 Then the driving state at time t is a uniform right turn; Δv is the set speed threshold; 8), |v t -v t-1 | < Δv and b t < b t-1 Then the driving state at time t is a uniform left turn; 9)、|v t -v t-1 | < Δv and |b t -b t-1 | < Δb, then the driving state at time t is straight-ahead at a constant speed; Among them, v t and v t-1 are the speeds at times t and t - 1, and b t and b t-1 are the direction angles at times t and t - 1.

6. The driver identification method based on driving style according to claim 4, characterized in that: For every two consecutive driving states, constructing the conversion intensity between each pair of consecutive driving states with speed difference, azimuth angle difference, and the historical transition relationship of the states, satisfying the following relational expression: In the formula, s i ,s j are respectively the source state and the target state in each pair of consecutive driving states; I(s i ,s j ) is the conversion strength from the source state to the target state; s i (v), s j (v) and s are the speeds of the source state and the target state respectively; (s i (v) - s j (v)) 2 is the speed difference; s i (b), s j (b) are the direction angles of the source state and the target state respectively; (s i (b) - s j (b)) 2 is the azimuth difference; m is the historical transition relationship between the source state and the target state in each pair of consecutive driving states. When there has been the same historical transition relationship between the source state and the target state, the value is 1, otherwise the value is 0.

7. The driver identification method based on driving style according to claim 1, characterized in that: Using a parallel method, the state transition feature matrix and the motion statistics feature matrix are fused according to the time step to obtain a complex matrix, including: Taking time as the fusion dimension, the state transition feature and the motion statistics feature form a feature complex vector, and the feature complex vector is used to form a complex matrix; Among them, the feature complex vector satisfies the following relational expression: In the formula, is the characteristic complex vector, mt is the motion statistical feature, st is the state transition feature, and i is the imaginary unit.

8. The driver recognition method based on driving style according to claim 1, wherein: The driver classification network includes a residual RNN encoder, a residual RNN decoder, and a prediction head; The residual RNN encoder contains two recurrent blocks, the residual RNN decoder contains two recurrent blocks symmetric to the encoder, and the prediction head includes a fully connected layer.

9. The driver recognition method based on driving style according to claim 8, wherein: In the residual RNN encoder, the first recurrent block maps the input of a given shape to a ReLu activation sequence with the same shape as the input; the second recurrent block maps the addition output to a compressed state sequence.

10. A driver recognition system based on driving style, characterized in that, Including: An acquisition module, a sample construction module, a network training module, an identification module, and a time length adjustment module; The acquisition module is used to acquire the driving trajectory data of the vehicle; The sample construction module is used to divide the acquired driving trajectory data into multiple sub-trajectory sequences based on the first sliding window; Based on the second sliding window, each sub-trajectory is divided into multiple overlapping segments; Based on the driving state index table, the driving trajectory data in the segment is converted into multiple consecutive driving states, and the state transition feature matrix of the segment is formed by the transition intensity between every two consecutive driving states; based on the statistical method, the dynamic characteristics of the driving trajectory data in the segment are statistically analyzed, and the motion statistics feature matrix is constructed with the obtained statistics; Using a parallel method, the state transition feature matrix and the motion statistics feature matrix are fused according to the time step to obtain a complex matrix; The network training module is used to establish a driver classification network, and use the complex matrix and the driver recognition label as training samples to train the driver classification network; The identification module is used to perform driver identification based on driving style using the trained driver classification network; The time length adjustment module is used to use the trained driver classification network to identify the recognition accuracy when the second sliding window has different time lengths, and send the time length of the second sliding window corresponding to the maximum recognition accuracy to the sample construction module.

11. A terminal, including a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1-9.

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

Images