LSTM network-based driving style online identification method and system

Abstract

The invention provides a driving style online identification method and system based on an LSTM network, and the method comprises the steps: carrying out the online identification of the driving style of a driver based on a driving style online identification model which is trained, verified and tested in advance; the pre-training, verification and testing steps of the driving style on-line identification model are as follows: dividing a large amount of free driving data into a training set, a verification set and a testing set; training the LSTM network based on the training set; and based on the verification set and the test set, verifying and testing the trained LSTM network, and determining a driving style online identification model. According to the driving style online identification method and system based on the LSTM network, efficient and accurate online identification of the driving style of the driver is realized.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to an online driving style identification method and system based on LSTM network. Background Art

[0002] At present, with the rapid development of intelligent driving and autonomous driving technologies, online identification of driving style has become one of the key technologies to improve driving safety and riding comfort.

[0003] Driving style not only has a profound impact on the vehicle's handling performance, but is also directly related to the safety and experience of passengers. Specifically, driving style is reflected in acceleration, braking, steering and other aspects, reflecting the driver's decision-making habits and risk tolerance. Therefore, accurate identification of driving style can help the vehicle better adapt to the driver's behavioral characteristics, thereby optimizing the vehicle's dynamic control strategy and ensuring safe driving.

[0004] In addition, by real-time monitoring and analysis of the driver's behavior, the system can provide personalized driving assistance services, improve the overall driving experience, and reduce the risk of accidents.

[0005] In the future intelligent transportation system, online driving style recognition technology will also provide important support for more efficient and intelligent traffic management, helping to build a safe and convenient travel environment.

[0006] Therefore, an efficient and accurate online driving style identification method is urgently needed. Summary of the invention

[0007] One of the purposes of the present invention is to provide a driving style online identification method based on LSTM network, so as to realize efficient and accurate online identification of the driver's driving style.

[0008] An embodiment of the present invention provides a driving style online identification method based on an LSTM network, comprising:

[0009] Based on the pre-trained, verified and tested online driving style identification model, the driver's driving style is identified online;

[0010] The steps of pre-training, verifying and testing the online driving style identification model are as follows:

[0011] Divide a large amount of free driving data into training set, validation set and test set;

[0012] Based on the training set, train the LSTM network;

[0013] Based on the validation set and test set, the trained LSTM network was verified and tested respectively to determine the online driving style identification model.

[0014] Optionally, the online identification of the driver's driving style based on the pre-trained, verified and tested online driving style identification model includes:

[0015] When the driver is operating the vehicle, the lane-changing condition data is identified and extracted from the vehicle's driving data in real time based on the online driving style identification model;

[0016] Whenever a new lane-changing condition data is identified and extracted, the driving style represented by the new lane-changing condition data is identified based on the online driving style identification model;

[0017] As the driver's driving process continues, the collected multiple style recognition results are statistically analyzed to dynamically update the driver's driving style.

[0018] Optionally, the method of dynamically updating the driver's driving style by statistically analyzing the collected multiple style recognition results includes:

[0019] Dynamic update is performed through the following formula:

[0020]

[0021] Among them, d s Dynamically updated driver driving style s 、n n and n a are the times that the identification results are conservative, general and aggressive in multiple style identification results, n is the total number of lane change condition data extracted; the three values ​​of -1, 0, and 1 are the quantitative representations of conservative, general and aggressive driving styles respectively.

[0022] Optionally, when training the LSTM network based on the training set, the steering wheel angle δ in the training set is sw , vehicle speed V, steering wheel speed Longitudinal acceleration a x , lateral acceleration a y The time series of the heading angle ψ calculated is used as the input gate of the LSTM network.

[0023] Optionally, the free driving data includes at least: steering wheel angle, steering wheel speed, lateral and longitudinal acceleration, yaw angular velocity, vehicle speed, brake master cylinder pressure, accelerator pedal opening, longitude and latitude, and the position, speed and acceleration of surrounding obstacles.

[0024] Optionally, free driving data is collected through a driving data collection platform.

[0025] Optionally, the driving data collection platform includes:

[0026] Inertial navigation, camera, angular millimeter-wave radar and CAN acquisition equipment are connected to each of the collected vehicles.

[0027] Optionally, the online driving style identification method based on the LSTM network also includes:

[0028] When the online driving style identification result of the driver exceeds the preset number of times as aggressive within the latest preset first time, a safety reminder is given to the driver, and the online driving style identification of the driver continues to be performed based on the online driving style identification model until the new online driving style identification result of the driver is non-aggressive within the new latest preset second time, and then the safety reminder to the driver is stopped;

[0029] The safety reminder to the driver includes:

[0030] Determining a plurality of detection ranges and respective corresponding detection weights based on the driving trajectory of the driver within a third time preset before and after the current time;

[0031] Traversing each detection range in turn, and determining the detection value of the traversed detection range based on the latest driving style, number of driving vehicles and detection weight corresponding to the traversed detection range of other drivers falling within the traversed detection range during each traversal;

[0032] After traversing each detection range, the detection range whose detection value exceeds the detection threshold is taken as the target detection range;

[0033] Based on the relative position relationship between the target detection range and the vehicle driven by the driver, the driver is given a safety reminder.

[0034] Optionally, the determining of a plurality of detection ranges and their corresponding detection weights based on the driving trajectory of the driver within a third time preset before and after the current time includes:

[0035] Determine a first target position point from the driving trajectory; wherein the first target position point is a position point where the driver's driving style is historically identified online as aggressive;

[0036] Draw a first circular range with the first target position point as the center and a preset first radius length as the radius;

[0037] Determining whether the multimodal feature in the first circular range meets the feature triggering condition;

[0038] When the determination is yes, generating a range search condition based on the multimodal features of the first circular range;

[0039] Based on the range search condition, searching for a second circular range on the driving trajectory;

[0040] The center of the second circular range is used as the second target position point;

[0041] A first semicircle range is drawn with the first target position point as the center and the preset second radius length as the radius; wherein a ray starting from the first target position point in the driving direction of the first target position point on the driving trajectory bisects the first semicircle range;

[0042] A second semicircle is drawn with the second target position point as the center and the preset third radius length as the radius; wherein the ray starting from the second target position point in the driving direction of the second target position point on the driving trajectory bisects the second semicircle;

[0043] The first semicircle range and the second semicircle range are used as multiple detection ranges;

[0044] The sum of the overlaps between the same detection range and other different detection ranges is taken as the detection weight of the same detection range.

[0045] The driving style online identification system based on LSTM network provided by the embodiment of the present invention includes:

[0046] An online identification module, used for online identification of the driver's driving style based on a pre-trained, verified and tested online driving style identification model;

[0047] The steps of pre-training, verifying and testing the online driving style identification model are as follows:

[0048] Divide a large amount of free driving data into training set, validation set and test set;

[0049] Based on the training set, train the LSTM network;

[0050] Based on the validation set and test set, the trained LSTM network was verified and tested respectively to determine the online driving style identification model.

[0051] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0052] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0054] Figure 1 Schematic diagram of an online driving style identification method based on LSTM network in an embodiment of the present invention;

[0055] Figure 2 Schematic diagram of a driving behavior collection system in an embodiment of the present invention;

[0056] Figure 3 is a schematic diagram of a simplified bicycle model in an embodiment of the present invention;

[0057] Figure 4 is a curve diagram of the turning angle and heading change under different working conditions in an embodiment of the present invention;

[0058] Figure 5 is a distribution diagram of the driver's lateral position error in an embodiment of the present invention;

[0059] Figure 6 A distribution diagram of the pilot's heading error in an embodiment of the present invention;

[0060] Figure 7 A schematic diagram of an LSTM network structure in an embodiment of the present invention;

[0061] Figure 8 The curve diagram of the accuracy and loss change during the training process in the embodiment of the present invention;

[0062] Fig. 9 A flowchart of online driving style identification in an embodiment of the present invention;

[0063] Fig.10 A schematic diagram of the result of identifying the style of driver No. 1 in an embodiment of the present invention;

[0064] Fig.11 A schematic diagram of the result of identifying the style of driver No. 2 in an embodiment of the present invention;

[0065] Fig.12 A schematic diagram of the style identification result of driver No. 3 in an embodiment of the present invention;

[0066] Fig.13 A schematic diagram of an online driving style identification system based on an LSTM network in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0068] The embodiment of the present invention provides a driving style online identification method based on LSTM network, such as Figure 1 As shown, including:

[0069] S1. Based on the online driving style recognition model that has been trained, verified and tested in advance, the online driving style recognition of the driver is performed online;

[0070] The steps of pre-training, verifying and testing the online driving style identification model are as follows:

[0071] S11. Divide a large amount of free driving data into training set, validation set and test set;

[0072] S12. Train the LSTM network based on the training set;

[0073] S13. Based on the validation set and the test set, the trained LSTM network is verified and tested respectively to determine the online driving style identification model.

[0074] The online identification of the driving style of the driver based on the online identification model of the driving style that has been trained, verified and tested in advance comprises:

[0075] When the driver is operating the vehicle, the lane-changing condition data is identified and extracted from the vehicle's driving data in real time based on the online driving style identification model;

[0076] Whenever a new lane-changing condition data is identified and extracted, the driving style represented by the new lane-changing condition data is identified based on the online driving style identification model;

[0077] As the driver's driving process continues, the collected multiple style recognition results are statistically analyzed to dynamically update the driver's driving style.

[0078] The method of dynamically updating the driver's driving style by statistically analyzing the collected multiple style recognition results includes:

[0079] Dynamic update is performed through the following formula:

[0080]

[0081] Among them, d s Dynamically updated driver driving style s 、n n and n a are the times that the identification results are conservative, general and aggressive in multiple style identification results, n is the total number of lane change condition data extracted; the three values ​​of -1, 0, and 1 are the quantitative representations of conservative, general and aggressive driving styles respectively.

[0082] Based on the training set, when training the LSTM network, the steering wheel angle δ in the training set is sw , vehicle speed V, steering wheel speed Longitudinal acceleration ax , lateral acceleration a y The time series of the heading angle ψ calculated is used as the input gate of the LSTM network.

[0083] The free driving data includes at least: steering wheel angle, steering wheel speed, lateral and longitudinal acceleration, yaw angular velocity, vehicle speed, brake master cylinder pressure, accelerator pedal opening, longitude and latitude, and the position, speed, and acceleration of surrounding obstacles.

[0084] Free driving data is collected through the driving data collection platform.

[0085] The driving data collection platform includes:

[0086] Inertial navigation, camera, angular millimeter-wave radar and CAN acquisition equipment are connected to each of the collected vehicles.

[0087] The working principle and beneficial effects of the above technical solution are:

[0088] Driving data is the premise and basis for analyzing driver behavior characteristics. Although simulators can simulate similar roads and traffic environments, they cannot completely restore the real driving experience. In comparison, real vehicle data collection is carried out in real roads and traffic environments, and can obtain the most realistic driving data and vehicle status information, so it has higher authenticity, comprehensiveness and reliability. To this end, this application has built a data collection platform to collect natural driving data of drivers.

[0089] For data collection vehicles, three types of sensors are installed to obtain vehicle information and environmental information: inertial navigation, camera and angular millimeter wave radar, and VECTOR VN1630A is used as the CAN acquisition device. The overall acquisition system architecture is as follows: Figure 2 shown.

[0090] In the acquisition system, the driver controls the vehicle movement by manipulating the vehicle steering wheel, accelerator pedal and brake pedal. The vehicle's own state parameters can be read directly based on CAN signals through its OBD interface; the latitude and longitude information of the real driving scene and the vehicle acceleration information can be obtained from the inertial navigation through the RS232 serial port and the CAN network in real time; the relative position, speed and acceleration of surrounding vehicles and obstacles can be obtained through the millimeter-wave radar at the four corners of the vehicle, and the lane line information can be output from the camera, both of which can be transmitted through CAN signals. VECTOR VN1630A is a CAN transceiver with 4 CAN high-speed 1051cap transceivers and can achieve software and hardware synchronization. VECTOR receives vehicle signals, inertial navigation signals, camera signals and millimeter-wave radar signals in real time and transmits them to the computer through the USB interface for final data recording and storage. During the whole process, the following data were recorded in total for subsequent analysis of driver behavior characteristics: steering wheel angle, steering wheel speed, lateral and longitudinal acceleration, yaw angular velocity, vehicle speed, brake master cylinder pressure, accelerator pedal opening, longitude and latitude, and the position, speed and acceleration of surrounding obstacles.

[0091] After completing the comprehensive collection of 30 driver behavior data, in order to deeply analyze the differences in lateral behavior characteristics between different drivers, this application selected the most representative lane keeping and lane changing conditions from a large amount of original data as the basis for analyzing driver behavior characteristics. By comparing the distribution of characteristic parameters of drivers under these two typical conditions, the behavioral characteristics and decision-making mechanisms of drivers in different driving scenarios are revealed.

[0092] The collected original signals are shown in Table 1. In addition, the DBC protocols of the inertial navigation, camera and corner radar used in this application all carry timestamp data, and synchronization between data can be achieved by simply aligning the collected information of each sensor on the time axis.

[0093] Table 1 List of main signals for data acquisition

[0094]

[0095] By performing reasonable calculations and post-processing on the data in the above signal list, lane changing conditions and lane keeping conditions can be extracted from the free driving data, thereby deeply analyzing the driver's characteristics.

[0096] (1) Lane-changing condition extraction

[0097] Given that the driver's steering operation plays a pivotal role in distinguishing different driver behavior patterns, many studies have been devoted to separating and analyzing this specific operation from free driving data. Based on this, this application sets clear steering driving conditions, including:

[0098] ① The vehicle speed V>2m / s to ensure that the vehicle is in a dynamic driving state;

[0099] ② Steering wheel angle δ sw >5° to identify significant turning movements;

[0100] ③ Steering wheel speed To capture the driver's steering rate characteristics.

[0101] When the vehicle deviates from the stable straight-line driving state, the lane-changing condition extraction program is started. Then, if the vehicle returns to the straight-line driving state or exceeds the preset condition duration T con (For example, set T con =10s), the data extraction work of the current working condition is ended. Through the above method, a set of time series data representing the driver's steering behavior is obtained.

[0102] About 50 sets of steering condition data can be extracted from each driver's 25-minute driving data. Through two experiments on 30 drivers, a total of 3285 sets of steering condition data were obtained, all of which were generated by the driver when manipulating the steering. However, some of the data were generated due to the change in road curvature causing the driver to turn. In order to accurately filter out the data of the lane change part, this application further estimates the heading angle of these time series data, thereby effectively eliminating the non-lane change data.

[0103] The accuracy requirement for heading estimation is not particularly strict, and being able to distinguish between lane change and steering behavior of the vehicle can meet the actual needs. Therefore, this application uses a single vehicle model as the calculation framework. When applying this model, the following basic assumptions will be followed to ensure the real-time and easy-to-use model:

[0104] ① Assume that the vehicle's movement is limited to the horizontal plane, ignoring the movement in the vertical direction. This assumption is based on the movement characteristics of the vehicle in most actual scenarios, which can effectively simplify the calculation and improve the real-time performance of the model.

[0105] ② Assuming that the turning angles of the left and right wheels are equal, the left and right tires can be merged into one, further simplifying the construction of the single vehicle model. This simplification not only improves the ease of use of the model, but also helps to more accurately simulate the actual motion state of the vehicle.

[0106] ③ Assume that the vehicle speed changes slowly and ignore the load transfer between the front and rear axles. This assumption is based on the dynamic characteristics of the vehicle during stable driving, which helps to reduce the complexity of the model and maintain its accuracy.

[0107] ④ Assume that the body and suspension system are rigid, which means that the deformation of the body and suspension can be ignored during the vehicle movement. This assumption helps improve the computational efficiency of the model while ensuring the accuracy of the model in simulating vehicle movement.

[0108] The simplified model is as follows Figure 3 As shown in the figure, the left / right front wheel is combined into point A, the left / right rear wheel is combined into point B, and point C is the center of mass of the vehicle. The specific symbols are defined in Table 2.

[0109] Table 2. Parameters of the bicycle model

[0110]

[0111] Based on the principles of vehicle dynamics, the following equations can be derived to describe the vehicle's motion characteristics. First, based on the vehicle's steering geometry:

[0112]

[0113] Next, by introducing the assumption that the rate of change of the vehicle's direction is equal to the vehicle's angular velocity, we can obtain:

[0114]

[0115] Combining the above two equations, we can derive the vehicle angular velocity The expression is:

[0116]

[0117] In addition, the slip angle β can be calculated by the following formula:

[0118]

[0119] Finally, by calculating the vehicle angular velocity By integrating over time, we can get the change in the vehicle's heading angle at any given time.

[0120] Figure 4 The steering wheel angle and the calculated heading angle curve under two typical working conditions are shown. Figure 4 (a) shows the data characteristics of the lane-changing condition, that is, the overall change of the heading angle at the start and end of the vehicle is almost zero; in contrast, Figure 4 The turning condition depicted in (b) is completely different. The heading angle at the start and end times has obvious changes, indicating that the vehicle has changed its driving direction due to the curve of the road. By calculating the change in heading angle, all data have obtained clear working condition information, so that those atypical lane change data can be specifically eliminated.

[0121] (2) Lane keeping condition extraction

[0122] Lane keeping is the most common operating condition of a vehicle during driving. For the collected natural driving data, excluding the previously extracted lane changing conditions, we only need to further exclude the following two special conditions, and then the remaining data can be intercepted as lane keeping conditions.

[0123] First, eliminate the situation where the driver initiates lane change but eventually gives up. Specifically, when the vehicle has an obvious lateral deviation (set to 1m in this application) or has entered the adjacent lane, and then quickly returns to the original lane due to the influence of the surrounding environment, it is deemed that the driver has exited the lane under this condition. Secondly, the transition adjustment phase after the vehicle has just completed the lane change should also be excluded from the lane keeping condition. In this application, the 5s time interval before and after the lane change condition (including the lane change exit condition) is set as the adjustment phase, and the data segments outside this interval can be considered as lane keeping conditions.

[0124] In order to summarize the overall behavioral characteristics and individual differences of drivers, this application uses two representative parameters to analyze the behavioral characteristics of drivers during lane keeping: lateral position error e y and heading error e ψ .e y Defined as the lateral distance between the current position and the desired trajectory, e ψ Defined as the angle between the vehicle axis and the tangent of the desired trajectory.

[0125] Figure 5 The distribution of the driver's lateral position deviation under lane keeping conditions is shown. It can be seen that the driver's lateral error is maintained within ±0.5m for 90% of the driving time, and its overall law generally obeys the Gaussian distribution. The mean μ y =-0.043, standard deviation σ y =0.863.

[0126] Figure 6 The heading error distribution of the driver is shown in Figure 2. Similar to the lateral position error, the heading error can also be fitted by a Gaussian distribution, and the final mean μ ψ =0.063, standard deviation σ ψ =0.521.

[0127] According to the overall data distribution results, during the lane keeping process, the driver's lateral position deviation and heading deviation tend to remain near 0. However, compared with the heading deviation, the position deviation has a larger standard deviation, which shows that the driver is more sensitive to the heading deviation, but has a certain tolerance for the position deviation, allowing a certain degree of fluctuation. Therefore, the driver's forward direction of the vehicle is the most critical when keeping the lane, and whether it is completely in the center of the lane line can be appropriately relaxed.

[0128] Compared with lane keeping conditions, lane changing conditions require the driver to perform a wider range of lateral position error correction and heading error adjustment, contain more behavioral control characteristics, and better reflect the differences in characteristics between different drivers.

[0129] To this end, this application shows the lane change trajectories of drivers at different speeds, and lists the trajectories of three typical drivers at the same speed. In addition, in order to facilitate the comparison of trajectories, the three trajectories are aligned with the focus of the lane line in the longitudinal direction, and finally Figures 7 to 9 As shown in the figure, drivers 1 to 3 correspond to aggressive, general and conservative driving styles respectively.

[0130] The results in the figure show that human drivers can achieve more precise control of the vehicle at low speeds, and can maintain driving in the center of the lane at the start and end stages of the lane change process. When the speed increases to 80km / h, the lane change trajectory has obvious fluctuations, and the vehicle also deviates to a certain extent from the center of the lane. In addition, at different speeds, each driver's control style remains consistent, that is, the aggressive driver No. 1 always completes the lane change at a faster speed, while the conservative driver No. 3 chooses to complete the lane change smoothly and slowly.

[0131] On this basis, in order to quantitatively analyze the characteristic differences between different drivers, key information such as the lateral and longitudinal displacements and lane change time during the driver's lane changing process was recorded and counted. The specific content is shown in Table 3.

[0132] Table 3 Lane changing data statistics

[0133]

[0134] The results in the table show that, overall, it takes about 4-6 seconds for a driver to complete a complete lane change process, which is not highly correlated with vehicle speed. Vehicle speed mainly affects the lateral and longitudinal displacement of the driver's lane change, of which the longitudinal displacement is related to the driving distance, and the lateral displacement is directly related to the accuracy of lane change. When the vehicle speed increases from 40km / h to 80km / h, the average lateral displacement of the driver decreases from 3.31m to 3.00m, which means that the offset between the vehicle and the center of the lane increases by 0.31m. For drivers of different styles, the aggressive driver No. 1 quickly completed the lane change at different speeds, with an average lane change time of 4.11s; while the conservative driver No. 3 pays more attention to safety and stability when changing lanes, with an average lane change time of 5.12s; the average lane change time of the general driver No. 2 is 4.56s, which is between the two. On the other hand, aggressive drivers also have a greater tolerance for the lateral position deviation between the vehicle's current position and the center of the lane. The average distance between their final lane change position and the center line of the lane is 0.46m, while that of average and conservative drivers is only 0.18m and 0.31m.

[0135] Long Short-Term Memory (LSTM) is a special RNN that is designed to solve the difficulties encountered by traditional RNNs in dealing with long-term dependency problems. LSTM introduces a gating mechanism to control the speed of information accumulation, including selectively adding new information and selectively forgetting previously accumulated information, thereby improving the long-term dependency problem of RNN and alleviating the gradient vanishing problem during long sequence training.

[0136] The LSTM network structure is as follows Figure 2 As shown, the core components include input gate, forget gate, output gate and a cell state.

[0137] The input gate controls how much input information at the current moment needs to be saved in the cell state. Its expression is as follows:

[0138] i t =σ(W xi x t +W hi h t-1 +b i )

[0139] Among them, x t is the input of the current time step, h t-1 is the hidden state of the previous time step, W xi and W hi is the weight matrix, b i is the bias term, and σ is the sigmoid activation function.

[0140] The function of the forget gate is to determine how much information of the cell state at the previous moment needs to be forgotten and how much information is retained in the current cell state. Its expression is as follows:

[0141] f t =σ(W xf x t +W hf h t-1 +b f )

[0142] Among them, W xf and W hf is the weight matrix, b f is the bias term.

[0143] The output gate controls how much information of the current cell state needs to be output to the external state, which can be expressed as:

[0144] o t =σ(W xo x t +W ho h t-1 +b o )

[0145] Among them, W xo and W ho is the weight matrix, b o is the bias term.

[0146] The cell state is regarded as the "memory" of the network, which can pass information in the sequence chain, so that even the information of the earlier time step can be carried to the cells of the later time step, thus overcoming the influence of short-term memory. Through the "gate" structure, LSTM can learn which information needs to be saved or forgotten during training, so as to achieve effective processing of long-term dependent information. The update method is as follows:

[0147]

[0148] Among them, c t-1 is the cell state at the previous time step, W xc and W hc is the weight matrix, b c is the bias term, Represents element-wise multiplication (Hadamard product).

[0149] For each hidden state h t , can be obtained by the following formula:

[0150]

[0151] In order to be consistent with the data collected from the actual vehicle, this application uses the steering wheel angle δsw , vehicle speed V, steering wheel speed Longitudinal acceleration a x , lateral acceleration a y The time series of the heading angle ψ is calculated as the input gate x t During training, 1831 sets of lane-changing data were extracted from the free-driving data of 30 drivers. In order to ensure the generalization and robustness of the model, 1300 sets of data (accounting for 70% of the total data) were selected as the training set to train and optimize the LSTM model; the remaining 531 sets of data will be used as validation sets and test sets to verify and evaluate the performance of the model.

[0152] The training process is as follows Fig.13 As shown in Figure 1, the verification results of using the remaining data as the test set are shown in Table 1. From the results, it can be seen that in a single lane change condition, the trained LSTM network can well detect conservative, general and aggressive driving styles, and the accuracy of the three styles is more than 90%, which can meet the needs of online identification.

[0153] Table 4 Verification set identification results

[0154]

[0155]

[0156] After ensuring the accuracy of single lane change condition identification, the online identification and dynamic update of driver style will follow the following steps: Fig. 9 The process shown is as follows: While the driver is operating the vehicle, the online recognition model will continuously identify and extract lane change data from the vehicle's driving data in real time. Whenever a new piece of data is captured, the LSTM model will quickly intervene to instantly identify the driving style of the lane change, thereby revealing the behavioral style characteristics of a single operation. As the driving process continues, the style recognition results of multiple lane change conditions will be collected, and through statistical analysis, the driver's style can be dynamically updated and accurately portrayed.

[0157] This updating process aims to ensure that the description of driving style gradually matches the driver's actual driving performance as the amount of data increases. s The update formula is as follows:

[0158]

[0159] Among them, n s 、n n and n aare the times when the identification results are conservative, general, and aggressive, respectively, and n is the total number of lane change conditions extracted. In addition, in order to achieve a concise and intuitive expression of driving style, the conservative, general, and aggressive driving styles are quantified into three values: -1, 0, and 1, respectively.

[0160] In order to verify the effect of the style recognition model in real scenarios, this application installed the algorithm in Simulink, obtained vehicle operation data and environmental information through VECTOR, and carried out online recognition and verification of the styles of 5 drivers according to the above working condition extraction → single style recognition → style result statistics process. The results are shown in Table 5.

[0161] Table 5 Driver style online identification results

[0162]

[0163]

[0164] It can be seen from this that when the driver is driving the vehicle, although the style of each lane change condition is different, the overall style characteristics of the driver can be obtained by statistically analyzing the identification results of multiple lane change behaviors. This analysis method comprehensively considers the performance of the driver in different scenarios to ensure that the results are of practical significance. It is worth noting that the driver's style is not absolutely aggressive or conservative, but a relative concept. Therefore, the driver's style results calculated statistically range from -1 to 1, where the closer the value is to -1, the more conservative the driver's style is, and the closer the value is to 1, the more aggressive the driver's style is. 0 as the middle value represents the driver's general style. In this way, a unique and specific style evaluation can be provided to each driver, so as to better understand and adapt to their driving habits.

[0165] In order to more clearly demonstrate the process of identification, the following will take the style identification curves of drivers 1-3 in Table 2 as an example. Figures 10 to 12 Provide a visual explanation.

[0166] For driver No. 1, Fig.10 It can be observed from the identification process that it exhibits highly consistent driving style characteristics during multiple lane changes. In the early stages of the algorithm, due to the fluctuations in the behavioral style of driver No. 1 during one or two lane changes, its cumulative identification results briefly jumped between conservative and aggressive. However, as the driving process continued, through the algorithm's identification statistics of multiple lane change conditions, the style value of driver No. 1 finally stabilized at around -0.9. This result clearly shows that driver No. 1 prefers to maintain stable vehicle control and has a conservative driving style.

[0167] For driver No. 2, Fig.11 It can be seen from the driving style identification result curve that his driving style shows obvious repeated fluctuations between general and aggressive. This shows that during the driving process, Driver No. 2 may flexibly adjust his driving strategy according to changes in road conditions and traffic environment. In some driving periods, he tends to adopt a general driving style and maintain a moderate speed and acceleration. However, in other periods, he may choose an aggressive driving method, such as accelerating to change lanes. This fluctuation in driving style reflects the flexibility and adaptability of Driver No. 2 during the driving process. After comprehensive analysis by the algorithm, the final style value of Driver No. 2 is determined to be around 0.33, which shows that he has a general driving style overall, but will show aggressive driving behavior in specific situations.

[0168] For driver No. 3, from Fig.12 It can be clearly seen from the driving style identification result curve that his driving style presents a distinct aggressive feature. Throughout the driving process, the style value of driver No. 3 remained at a high level, reflecting his tendency to adopt a more aggressive and risky strategy when driving. In most lane change conditions, driver No. 3's actions appeared decisive and quick, and he tended to pursue faster speeds and shorter driving times. When identifying driver No. 3's driving style, the LSTM algorithm accurately captured his aggressive characteristics, and the final style value was 0.8. It can be clearly stated that driver No. 3 is an aggressive driver.

[0169] The online style recognition algorithm applied in this application can effectively extract the driver's lane change conditions and accurately capture the driving style characteristics during driving. Through online verification of real vehicle data, the real-time and accuracy of the algorithm are demonstrated. In actual application, while the driver is driving, the model dynamically displays the current driver's style type (such as: -1 is absolutely conservative, 0 is general, 1 is aggressive, and if it is 0.6, it is relatively aggressive). The algorithm of this application can allow each driver to obtain a unique style judgment value.

[0170] In one embodiment, the driving style online identification method based on the LSTM network further includes:

[0171] S2. When the online driving style identification result of the driver exceeds the preset number of times as aggressive within the latest preset first time, a safety reminder is given to the driver, and the online driving style identification of the driver continues to be performed based on the online driving style identification model until the new online driving style identification result of the driver is non-aggressive within the latest preset second time, and then the safety reminder to the driver is stopped;

[0172] Wherein, in S2, the preset first time may be 150 seconds; the preset number of times may be 3 times; when the online identification result of the driver's driving style exceeds the preset number of times in the most recent preset first time and is aggressive, it indicates that the driver may have risky driving, and a safety reminder is given to the driver; the preset second time may be 80 seconds; when the new online identification results of the driver's driving style are all non-aggressive in the new most recent preset second time, it indicates that the driver has received the reminder and has taken corresponding measures, and the safety reminder to the driver is stopped; if more and more drivers stop giving safety reminders to them in a short period of time after receiving the safety reminder, it indicates that the actual accuracy of the online identification model of the driving style for the driver's driving style is higher, and the system can actually verify the online identification model of the driving style based on this, and perform corresponding optimization training adjustments to further improve the working ability of the online identification model of the driving style; wherein, the safety reminder to the driver includes:

[0173] S21, determining a plurality of detection ranges and their corresponding detection weights based on the driving trajectory of the driver within a third time period preset before and after the current time;

[0174] Among them, in S21, the preset third time can be 300 seconds; the detection range is the road range to be selected that the driver needs to pay attention to when driving currently; the detection weight represents the degree to which the corresponding detection range requires the driver to pay attention to when driving currently; S22, traverse each detection range in turn, and each time traversal, determine the detection value of the traversed detection range based on the latest driving style of other drivers falling into the traversed detection range, the number of driving vehicles and the detection weight corresponding to the traversed detection range;

[0175] Among them, in S22, the latest driving style of other drivers falling within the traversed detection range refers to the latest result of online driving style identification of the other drivers using the driving style online identification model; the number of driving vehicles refers to the total number of vehicles driven by different drivers falling within the traversed detection range; the detection value determined in each traversal comprehensively represents the degree to which the driver needs to pay attention to the traversed detection range when driving at present; S23, after traversing each detection range, the detection range whose detection value exceeds the detection threshold is used as the target detection range;

[0176] Among them, in S23, the detection threshold may be, for example, 80; when the detection value exceeds the detection threshold, it indicates that the corresponding detection range requires the driver to pay more attention to the current driving, and is used as the target detection range;

[0177] S24: Providing a safety reminder to the driver based on the relative position relationship between the target detection range and the vehicle driven by the driver.

[0178] In S24, the relative position relationship between the target detection range and the vehicle driven by the driver can be a straight-line distance, a relative direction, etc.; the relative position relationship between the target detection range and the vehicle driven by the driver is used to give a safety reminder to the driver, for example: it is reminded that there may be risky driving vehicles within the target detection range having the relative position relationship with the vehicle driven by the driver, and if the driver also drives riskily, a safety accident may occur; it is convenient for the driver to understand the actual situation of the road he is currently driving on and give a safety reminder, which greatly improves the effect of the safety reminder; wherein, the detection value of the traversed detection range is determined based on the latest driving style of other drivers falling within the traversed detection range, the number of driving vehicles and the detection weight corresponding to the traversed detection range, including:

[0179] The detection value of the traversed detection range is calculated by the following formula:

[0180]

[0181] Among them, Dv is the detection value of the traversed detection range, G i is the quantitative representation of the latest driving style of the ith other driver within the traversed detection range (3 for conservative, 6 for general, and 9 for aggressive), W is the number of driving vehicles within the traversed detection range, that is, the total number of other drivers within the traversed detection range, K is the detection weight corresponding to the traversed detection range, and Q1 and Q2 are preset weight values.

[0182] In the above calculation formula, G i The larger the value is, the greater the degree of attention the driver needs to pay to the detection range traversed during the current driving. i It is positively correlated with Dv; the larger W is, the greater the degree to which the driver needs to pay attention to the traversed detection range during current driving, and W is positively correlated with Dv; the larger K is, the greater the degree to which the driver needs to pay attention to the traversed detection range during current driving, and K is positively correlated with Dv; therefore, setting the weight value in advance and calculating the detection value of the traversed detection range improves the efficiency and accuracy of determining the target detection range.

[0183] In one embodiment, the step S21, based on the driving trajectory of the driver within a third time preset before and after the current time, determines a plurality of detection ranges and their corresponding detection weights, including:

[0184] S2101, determining a first target position point from the driving trajectory; wherein the first target position point is a position point at which the driver's driving style is historically identified online as aggressive;

[0185] In S2101, the historical part of the driving trajectory can be obtained based on the driving record system, and the future part can be obtained based on the navigation planning system; when the driver is historically identified as having an aggressive driving style by the online driving style identification model, the position of the vehicle being driven is marked as the first target position point;

[0186] S2102, drawing a first circular range with the first target position point as the center and a preset first radius length as the radius;

[0187] Wherein, in S2102, the first radius length may be, for example, 1.2 meters;

[0188] S2103, determining whether the multimodal feature in the first circular range meets the feature triggering condition;

[0189] In S2103, the multimodal features of the first circular range include at least: road environment features, the number of vehicles contained in the first circular range at the current time, and the vehicle density, etc.;

[0190] S2104: when the determination is yes, generating a range search condition based on the multimodal features of the first circular range;

[0191] In S2104, the range search condition generated is to search for a second circular range that has some / all features similar / identical to the first circular range; the range search condition is generated based on the multimodal feature of the first circular range, for example: if the multimodal feature is that there is no traffic monitoring facility within the range, then the range search condition generated is to search for a range without traffic monitoring facilities within other ranges; S2105, based on the range search condition, search for a second circular range on the driving trajectory;

[0192] In S2105, based on the range search condition, when a second circular range is searched on the driving trajectory, the searched second circular range does not completely overlap with the first circular range; the center of the second circular range is also on the driving trajectory;

[0193] S2106, taking the center of the second circular range as the second target position point;

[0194] S2107, drawing a first semicircle range with the first target position point as the center and the preset second radius length as the radius; wherein a ray starting from the first target position point in the driving direction of the first target position point on the driving trajectory bisects the first semicircle range;

[0195] Among them, in S2107, the length of the second radius can be 2 meters; the first target position point on the driving trajectory has a driving direction, that is, the instantaneous driving direction of the driver driving the vehicle at the first target position point (which can be calculated as the direction from the first target position point to the trajectory point position 0.01 meters away from the first target position point in the future on the driving trajectory); under the constraint of the condition that the ray from the first target position point to the driving direction of the first target position point on the driving trajectory bisects the first semicircle range, the detection range can be accurately determined in combination with the driving state of the driver driving the vehicle; S2108, with the second target position point as the center of the circle and the preset third radius length as the radius, draw the second semicircle range; wherein, the ray from the second target position point to the driving direction of the second target position point on the driving trajectory bisects the second semicircle range;

[0196] Among them, in S2108, the length of the second radius can be 2.5 meters; the second target position point on the driving trajectory also has a driving direction, which is similarly the instantaneous driving direction of the driver driving the vehicle at the second target position point (which can be calculated as the direction from the second target position point to the trajectory point position 0.01 meters away from the second target position point in the future on the driving trajectory); under the constraint of the condition that the ray of the driving direction from the second target position point to the second target position point on the driving trajectory bisects the second semicircle range, it is further possible to accurately determine the detection range in combination with the driving state of the driver driving the vehicle; S2109, the first semicircle range and the second semicircle range are used as multiple detection ranges;

[0197] S2110. The sum of the overlaps between the same detection range and other different detection ranges is used as the detection weight of the same detection range.

[0198] Among them, the greater the sum of the overlaps between a detection range and other different detection ranges, the greater the degree of repeated reflection that requires the driver to pay attention to the current driving, and the greater the degree of the detection range that reflects the need for the driver to pay attention to the current driving; therefore, the sum of the overlaps is used as the detection weight of the same detection range. The first circular range is introduced, the second circular range is searched based on the first circular range, the first semicircular range and the second semicircular range are planned based on the first circular range and the second circular range, and the first semicircular range and the second semicircular range are used as multiple detection ranges, which improves the efficiency and accuracy of determining the detection range; secondly, the sum of the overlaps between the same detection range and other different detection ranges is used as the detection weight of the same detection range, which improves the efficiency, accuracy and comprehensiveness of determining the detection weight.

[0199] The feature trigger conditions include:

[0200] The maximum vector similarity between the situation vector formed by the multimodal feature and each of the multiple standard situation vectors exceeds the similarity threshold; wherein, the multimodal feature can be constructed as a vector based on the vector representation technology; the standard situation vector is pre-set, and the standard situation vector is constructed by multimodal features representing that the driver needs to pay attention to when driving currently, for example: there are no safety monitoring facilities within the range, etc.; the similarity threshold can be 80%; when this feature trigger condition is met, it means that the first circular range needs the driver to pay attention to when driving currently;

[0201] and / or,

[0202] The mapping value corresponding to the multimodal feature in the target knowledge mapping map exceeds the mapping threshold; wherein the target knowledge mapping map is the knowledge mapping map corresponding to the feature type set of the multimodal feature in the knowledge mapping map library. wherein the feature type set of the multimodal feature consists of different feature types of the multimodal feature, and the feature type set corresponds to a knowledge mapping map in the knowledge mapping map library, and the mapping values ​​corresponding to different multimodal features in the knowledge mapping map, the larger the mapping value, the more relevant knowledge that the driver needs to pay attention to when driving currently; the relevant knowledge can be pre-set by the technician; the mapping threshold can be 8; when this feature trigger condition is met, it means that the first circular range needs the driver to pay attention to when driving currently.

[0203] By setting the feature trigger condition, it is convenient to quickly trigger and determine whether it is necessary to generate a range search condition based on the multimodal feature of the first circular range, thereby improving the work efficiency of the system and reducing the condition generation resources of the system.

[0204] The embodiment of the present invention provides a driving style online identification system based on LSTM network, such as Fig.13 Including:

[0205] An online identification module 1, used for performing online identification of the driver's driving style based on a driving style online identification model that has been trained, verified and tested in advance;

[0206] The steps of pre-training, verifying and testing the online driving style identification model are as follows:

[0207] Divide a large amount of free driving data into training set, validation set and test set;

[0208] Based on the training set, train the LSTM network;

[0209] Based on the validation set and test set, the trained LSTM network was verified and tested respectively to determine the online driving style identification model.

[0210] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A driving style online identification method based on LSTM network, characterized in that: include: Based on the pre-trained, verified and tested online driving style identification model, the driver's driving style is identified online; The steps of pre-training, verifying and testing the online driving style identification model are as follows: Divide a large amount of free driving data into training set, validation set and test set; Based on the training set, train the LSTM network; Based on the validation set and test set, the trained LSTM network was verified and tested respectively to determine the online driving style identification model.

2. The driving style online identification method based on LSTM network as claimed in claim 1, characterized in that: The online identification of the driving style of the driver based on the online identification model of the driving style that has been trained, verified and tested in advance comprises: When the driver is operating the vehicle, the lane-changing condition data is identified and extracted from the vehicle's driving data in real time based on the online driving style identification model; Whenever a new lane-changing condition data is identified and extracted, the driving style represented by the new lane-changing condition data is identified based on the online driving style identification model; As the driver's driving process continues, the collected multiple style recognition results are statistically analyzed to dynamically update the driver's driving style.

3. The driving style online identification method based on LSTM network as claimed in claim 2, characterized in that: The method of dynamically updating the driver's driving style by statistically analyzing the collected multiple style recognition results includes: Dynamic update is performed through the following formula: Among them, d s Dynamically updated driver driving style s 、n n and n a are the times that the identification results are conservative, general and aggressive in multiple style identification results, n is the total number of lane change condition data extracted; the three values ​​of -1, 0, and 1 are the quantitative representations of conservative, general and aggressive driving styles respectively.

4. The driving style online identification method based on LSTM network as claimed in claim 2, characterized in that: Based on the training set, when training the LSTM network, the steering wheel angle δ in the training set is sw , vehicle speed V, steering wheel speed Longitudinal acceleration a x , lateral acceleration a y The time series of the heading angle ψ calculated is used as the input gate of the LSTM network.

5. The driving style online identification method based on LSTM network as claimed in claim 1, characterized in that: The free driving data includes at least: steering wheel angle, steering wheel speed, lateral and longitudinal acceleration, yaw angular velocity, vehicle speed, brake master cylinder pressure, accelerator pedal opening, longitude and latitude, and the position, speed, and acceleration of surrounding obstacles.

6. The driving style online identification method based on LSTM network as claimed in claim 1, characterized in that: Free driving data is collected through the driving data collection platform.

7. The driving style online identification method based on LSTM network as claimed in claim 6, characterized in that: The driving data collection platform includes: Inertial navigation, camera, angular millimeter-wave radar and CAN acquisition equipment are connected to each of the collected vehicles.

8. The driving style online identification method based on LSTM network as claimed in claim 2, characterized in that: Before training the LSTM network, the driver characteristic parameters are analyzed on the free driving data.

9. The driving style online identification method based on LSTM network as claimed in claim 8, characterized in that: The driver characteristic parameter analysis of the free driving data comprises: Lane changing conditions and lane keeping conditions are extracted from free driving data.

10. A driving style online identification system based on LSTM network, characterized in that: include: An online identification module, used for online identification of the driver's driving style based on a pre-trained, verified and tested online driving style identification model; The steps of pre-training, verifying and testing the online driving style identification model are as follows: Divide a large amount of free driving data into training set, validation set and test set; Based on the training set, train the LSTM network; Based on the validation set and test set, the trained LSTM network was verified and tested respectively to determine the online driving style identification model.

Images