A method for predicting early lane-changing intention of vehicles based on an improved KAN network

Abstract

The present invention discloses a method for predicting early lane-changing intention of a vehicle based on an improved KAN network. The method includes collecting the driving data of a target vehicle, dividing positive and negative samples according to the driving data of the target vehicle, obtaining a lane-changing data set according to the divided positive and negative samples, constructing a vehicle early lane-changing intention prediction model by using the lane-changing data set and training it, inputting the continuously collected driving data of the target vehicle to be measured into the trained vehicle early lane-changing intention prediction model for processing to obtain the lane-changing probability of the target vehicle to be measured, and finally comparing the lane-changing probability with a preset lane-changing probability threshold to obtain whether the target vehicle to be measured has changed lanes. The method of the present invention uses the KAN network based on a penalty lane-changing weight loss function to process the lane-changing data set, effectively improving the perception ability of the driverless vehicle in a mixed traffic environment, giving the driverless vehicle sufficient reaction time, and thus improving its driving safety.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation, and particularly to a method for predicting early lane-changing intention of vehicles based on an improved KAN network. Background Art

[0002] With the rapid progress of technology, driverless technology is booming with each passing day. However, for a long time in the foreseeable future, a complex situation will be formed on traffic roads where manned vehicles and driverless vehicles share road rights together. Against this background, accurately identifying the lane-changing behavior of manned vehicles or driverless vehicles will undoubtedly become a crucial means to ensure road traffic safety.

[0003] Currently, in-vehicle sensor technology has gradually become perfect. New energy vehicles at home and abroad, whether high-end luxury models or popular household cars, are mostly equipped with advanced sensor devices such as vision and radar. These sensors are like the "eyes" and "ears" of the vehicle, and can obtain the environmental information around the driverless vehicle in real time and accurately. Whether it is the driving speed and distance of the vehicle ahead, or the traffic flow in the surrounding lanes, they can be clearly captured. At the same time, the deep integration of big data and the transportation field has also played a huge role. The driving behavior data of historical vehicles on the road has been comprehensively collected and sorted out, and driverless vehicles can easily obtain these valuable data resources. These rich road traffic information provide a solid data support for in-depth data analysis and accurate vehicle behavior prediction.

[0004] However, the current existing technologies have obvious deficiencies. Most research and applications mainly focus on pursuing the accuracy of predicting vehicle lane changes, but seriously ignore the key factor of lane-changing time. In the actual road traffic scenario, the lead of lane-changing time plays a crucial role. When the prediction lead is too small, the reaction time of driverless vehicles will be greatly compressed. On a high-speed road, when a driverless vehicle is suddenly informed that the vehicle ahead is about to change lanes, but there is not enough time to make a reasonable reaction, this is very likely to lead to serious traffic accidents such as rear-end collisions and collisions, greatly affecting the driving safety of driverless vehicles. The data related to lane-changing time is of different qualities, with inconsistent formats and standards, insufficient representativeness, untimely updates, difficult to reflect new traffic changes, and there are also privacy and security risks, with the risk of data leakage. Summary of the Invention

[0005] In view of the deficiencies in the background art, the present invention proposes a method for predicting early lane-changing intention of vehicles based on an improved KAN network, which solves the technical problem that the prior art ignores vehicle early lane-changing prediction based on lane-changing time.

[0006] I. A Method for Predicting Early Lane - Changing Intent of Vehicles Based on Improved KAN Network

[0007] S1. Collect the driving data of several target vehicles through the on - vehicle sensors of the driverless vehicle, and divide the positive samples and negative samples according to all the target vehicle driving data.

[0008] S2. Pre - process the driving data according to the divided positive samples and negative samples to obtain the time - series data of several samples, and all the time - series data are aggregated to obtain a lane - changing data set.

[0009] S3. Construct a vehicle early lane - changing intent prediction model, input the lane - changing data set into the vehicle early lane - changing intent prediction model that adopts a penalty lane - changing weight loss function for training, and obtain a trained vehicle early lane - changing intent prediction model.

[0010] S4. Real - time collect the continuous driving data of several target vehicles to be measured from several frames before the current moment to the current moment frame, and input them into the trained vehicle early lane - changing intent prediction model for processing to obtain the lane - changing probability of each target vehicle to be measured in real - time. Compare the lane - changing probability of each target vehicle to be measured with a preset lane - changing probability threshold to obtain whether each target vehicle to be measured has changed lanes in real - time.

[0011] According to whether all the target vehicles to be measured have changed lanes obtained in real - time, the driverless vehicle makes corresponding driving adjustments, such as making lane - changing and decelerating operations.

[0012] The specific steps of step S1 are as follows:

[0013] S11. Collect the driving data of several target vehicles per frame through the on - vehicle sensors of the driverless vehicle.

[0014] S12. Divide the target vehicles into lane - changing vehicles and non - lane - changing vehicles according to the driving data.

[0015] S13. For lane - changing vehicles, the continuous driving data composed of the lane - changing completion moment of the target vehicle and the N - 1 frames before the lane - changing completion moment are used as positive samples, and the continuous driving data from the 2N - 1 frames before to the N frames before the lane - changing completion moment are used as negative samples; for non - lane - changing vehicles, any continuous N - frame driving data are used as negative samples.

[0016] The target vehicle is the vehicle ahead in the adjacent lane on the left or right side of the driverless vehicle. The driving data includes speed, left longitudinal distance, right longitudinal distance, and lateral distance; the left longitudinal distance is the longitudinal distance between the target vehicle and the vehicle ahead in the adjacent lane on the left side of the target vehicle; the right longitudinal distance is the longitudinal distance between the target vehicle and the vehicle ahead in the adjacent lane on the right side of the target vehicle; the lateral distance is the distance between the center point of the target vehicle and the center line of the lane where it is located. The lane change completion moment is the moment when the center point of the target vehicle coincides with the lane line on one side of the target vehicle.

[0017] Specifically, step S2 is as follows:

[0018] S21. Divide the samples into time series windows in the way that the sliding window size is M frames and the sliding step is one frame, obtaining N - M + 1 time series, and the N - M + 1 time series are combined to obtain the time series data of one sample.

[0019] S22. Process in the same way as in step S21 to obtain the time series data of several samples, and the time series data of all samples are aggregated to obtain the lane change dataset.

[0020] The vehicle early lane change intention prediction model in step S3 is the KAN network.

[0021] The penalty lane change weight loss function in step S3 is set according to the following formula:

[0022] L i + (p t + ) = -∑ x=1 N-M+1 αlog(p ix + )

[0023] L j - (p t - ) = -∑ y=1 N-M+1 log(1 - p jy - )

[0024] α = e (-max(0，d-Φ(epoch-1)-γ))

[0025] Φ(epoch - 1) = 1 / k∑ f=1 k G f epoch-1

[0026] Among them, both i and x represent indices; Li + () is the loss function for the i-th positive sample; p ix + represents the lane-changing probability predicted for the x-th time series in the i-th positive sample of the vehicle early lane-changing intention prediction model; α represents the weight; j and y both represent indices; L j - () is the loss function for the j-th negative sample; p jy - represents the lane-changing probability obtained for the y-th time series in the j-th negative sample of the vehicle early lane-changing intention prediction model; e represents a constant; max( ) represents taking the maximum value; d represents the difference between the index of the last frame of the x-th time series and the index of the lane-changing completion moment; epoch represents the index of the training round and also represents the epoch-th training; Φ(epoch - 1) represents the mean of the lane-changing time differences of all positive samples in the (epoch - 1)-th training round; γ represents a hyperparameter; k represents the total number of positive samples; f represents an index; G f epoch-1 represents the lane-changing time difference of the f-th positive sample in the (epoch - 1)-th training round.

[0027] The lane-changing time difference is specifically the difference between the index of the time series when the first lane-changing probability in each positive sample is not less than the preset lane-changing probability threshold and the index of the last time series.

[0028] When comparing the lane-changing probability of the target vehicle to be measured in step S4 with the preset lane-changing probability threshold, it is set according to the following formula to obtain whether the target vehicle to be measured has changed lanes:

[0029] f(p q ) = 1, p q ≥P th

[0030] f(p q ) = 0, p q <P th

[0031] where f(p q ) represents whether the q-th target vehicle to be measured has changed lanes; p q represents the lane-changing probability predicted by inputting the continuous driving data of the q-th target vehicle to be measured into the trained vehicle early lane-changing intention prediction model; 1 represents changing lanes; 0 represents not changing lanes; P th represents the preset lane-changing probability threshold.

[0032] II. A computer device, comprising a memory and a processor, where the memory stores a computer program, and the processor, when executing the computer program, implements the steps of the above method.

[0033] III. A computer-readable storage medium, on which a computer program is stored, and the computer program, when executed by a processor, implements the steps of the above method.

[0034] The innovation of the present invention lies in using the KAN network with a penalty lane-changing weight loss function to process the driving data of the target vehicle, achieving the improvement of the prediction time lead of the target vehicle's lane change while ensuring the accuracy rate of predicting the lane-changing behavior of the vehicle, and obtaining an advantage with stronger practical significance compared to the traditional method of predicting lane-changing behavior.

[0035] The beneficial effects of the present invention are as follows:

[0036] The method of the present invention processes the collected driving data of the target vehicle through the KAN network based on the penalty lane-changing weight loss function. While ensuring the accuracy rate of predicting the lane-changing behavior of the vehicle, it improves the prediction time lead of the target vehicle's lane change, increases the reaction time to changes in the surrounding environment, has stronger practical significance compared to the traditional method of predicting lane-changing behavior, and can better ensure traffic safety in the mixed driving environment of manned vehicles and driverless vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flowchart of the method of the present invention.

[0038] Figure 2 It is a schematic flowchart of the vehicle early lane change intention prediction model of the present invention using a penalty lane-changing weight loss function.

[0039] Figure 3 It is a schematic diagram of the driverless vehicle observing the lane-changing behavior of the target vehicle in Embodiment 1.

[0040] Figure 4 It is a schematic diagram of the target vehicle and surrounding vehicles in Embodiment 1.

[0041] Figure 5 It is a lane change probability map predicted by the vehicle early lane change intention prediction model in Embodiment 1.

[0042] Figure 6 It is a lane change probability map predicted by the vehicle early lane change intention prediction model in Embodiment 2 and Comparative Examples 1 - 3. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] The present invention will be described in more detail below with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those of ordinary skill in the art in the technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as within the protection scope of the present invention. The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art.

[0044] Embodiment 1

[0045] As Figure 1 shown, the method for predicting the early lane-changing intention of a vehicle in this embodiment includes the following steps:

[0046] S1. Collect the driving data of a number of target vehicles through the on-vehicle sensors of the driverless vehicle, and divide the positive samples and negative samples according to all the driving data of the target vehicles.

[0047] S11. Collect the driving data of each frame of a number of target vehicles through the on-vehicle sensors of the driverless vehicle;

[0048] In this embodiment, the driverless vehicle collects the driving data of the target vehicles on a 427m section of the westbound 10-lane to the eastbound 3-lane section of a certain urban expressway. The total collection time is 5 minutes and 33 seconds, and the collection frame rate is 24 frames per second.

[0049] S12. Divide the target vehicles into lane-changing vehicles and non-lane-changing vehicles according to the driving data.

[0050] S13. For the lane-changing vehicles, the continuous driving data composed of the lane-changing completion moment of the target vehicle and the N-1 frames before the lane-changing completion moment is used as the positive sample, and the continuous driving data from the 2N-1 frames before the lane-changing completion moment to the N frames before is used as the negative sample.

[0051] For the non-lane-changing vehicles, any continuous N frames of driving data are used as the negative sample. In specific implementation, the label of the positive sample is 1, and the label of the negative sample is 0. In this embodiment, N takes 192 frames, which is 8 seconds of driving data.

[0052] As Figure 3 and Figure 4 shown, the target vehicle is the vehicle in front on the adjacent lane to the left or right of the driverless vehicle; in specific implementation, only the nearest vehicle in front on each adjacent lane is considered, and the target vehicles of the driverless vehicle include the nearest vehicle in front on the adjacent lane to the left and the nearest vehicle in front on the adjacent lane to the right. As Figure 2 shown, at the current moment T0, one of the target vehicles of the driverless vehicle is in the left lane. When it reaches the moment T f the target lane has completely merged with the lane of the driverless vehicle.

[0053] The driving data includes speed, left longitudinal distance, right longitudinal distance, and lateral distance; the left longitudinal distance is the longitudinal distance between the target vehicle and the vehicle in front on the adjacent lane on the left side of the target vehicle; the right longitudinal distance is the longitudinal distance between the target vehicle and the vehicle in front on the adjacent lane on the right side of the target vehicle; the lateral distance is the perpendicular distance between the center point of the target vehicle and the center line of the lane where it is located. The moment when the lane change is completed is the moment when the center point of the target vehicle coincides with the lane line on one side of the target vehicle. In specific implementation, when the target vehicle is the vehicle in the leftmost lane, the left longitudinal distance is 0, indicating that the target vehicle cannot change lanes to the left; when the target vehicle is the vehicle in the rightmost lane, the right longitudinal distance is 0, indicating that the target vehicle cannot change lanes to the right. When the target vehicle merges from the ramp into the main road, it is regarded as being in the rightmost lane. In specific implementation, all distance measurements are made according to the center point in front of each vehicle.

[0054] In this embodiment, a total of 3,197 target vehicles are collected, including 782 positive samples (lane-changing vehicles) and 2,415 negative samples (non-lane-changing vehicles).

[0055] S2. Preprocess the driving data according to the divided positive and negative samples to obtain time series data of several samples, and all the time series data are aggregated to obtain a lane change data set.

[0056] S21. Divide the samples into time series windows in the way that the sliding window size is M frames and the sliding step is one frame, obtaining N - M + 1 time series, and the N - M + 1 obtained time series are combined to obtain the time series data of one sample.

[0057] In specific implementation, the time series window division is carried out in the order from the front to the back of the time frames. In this embodiment, the sliding window size of 72 frames (a total of 3 seconds) and the sliding step of one frame are used to divide the samples into time series windows, obtaining 121 time series, and the 121 obtained time series are combined to obtain the time series data of one sample.

[0058] S22. Process in the same way as in step S21 to obtain the time series data of several samples, and all the time series data of the samples are aggregated to obtain a lane change data set.

[0059] S3. Build a vehicle early lane change intention prediction model, input the lane change data set into the vehicle early lane change intention prediction model using a penalty lane change weight loss function for training, and obtain a trained vehicle early lane change intention prediction model.

[0060] The vehicle early lane change intention prediction model is a KAN (Kolmogorov - Arnold) network.

[0061] The design inspiration of the KAN network adopted in this embodiment comes from the Kolmogorov - Arnold theorem, which states that a multivariate continuous function can be represented as a composition of a finite number of univariate continuous functions and binary addition operations. Based on this theory, the KAN network improves the flexibility and expressive power of the model by introducing learnable activation functions on the network edges (i.e., weights).

[0062] The KAN network replaces the elements in the weight parameter matrix of the multi - layer perceptron (MLP) with non - linear activation functions containing variational parameters. These activation functions are usually learnable rather than fixed, which gives the KAN network higher flexibility in processing data.

[0063] The core of the KAN network is to use spline functions as activation functions, which replace the linear weight parameters in traditional neural networks. The introduction of spline functions enables the KAN network to simulate complex functions with fewer parameters and enhances the interpretability of the model.

[0064] The KAN network has a fully - connected structure, similar to the MLP, but its weights are represented as spline functions rather than real - valued weights. This design allows the KAN to adopt learnable activation functions at the network edges, providing extremely high flexibility.

[0065] The KAN network in this embodiment adopts the Kolmogorov - Arnold representation theorem, and the formula of the Kolmogorov - Arnold representation theorem is set as follows:

[0066] f(x1,…,…,x n )=∑ w=1 2n+1 ξ w (∑ r=1 n ψ w,r (x r ))

[0067] Where x i ∈[0,1], i = 1,2,…,n; f(x1,…,…,x n ) represents an n - ary continuous function composed of n basis functions; ξ w represents a univariate continuous function; ψ w,r also represents a univariate continuous function; x r represents the independent variable of each basic univariate continuous function ψ w,r ; w and r are both indices; n represents the arity of the continuous function; x1 and x n both represent the continuous function f(x1,…,…,x nindependent variable. The Kolmogorov-Arnold representation theorem states that a multivariate continuous function on a bounded domain can be represented as the sum of a series of univariate basis functions combined and added together.

[0068] In this embodiment, KAN uses two fully connected layers. The edges of the network use learnable activation functions to replace the weight parameters. Finally, the predicted lane-changing probability is output through the Sigmoid function. The lane-changing probability and the label are input into the improved loss function to calculate the loss function value of the current training round and the mean value of the lane-changing time differences of all positive samples in the current training round. And the mean value of the lane-changing time differences of all positive samples in the current training round is used as the weight for calculating the loss function of positive samples in the next training round.

[0069] As Figure 2 shown, the penalty lane-changing weight loss function is set according to the following formula:

[0070] L i + (p t + ) = -∑ x=1 N-M+1 α log(p ix + )

[0071] L j - (p t - ) = -∑ y=1 N-M+1 log(1 - p jy - )

[0072] α = e (-max(0，d-Φ(epoch-1)-γ))

[0073] Φ(epoch - 1) = 1 / k ∑ f=1 k G f epoch-1

[0074] Among them, both i and x represent indices; L i + () is the loss function of the i-th positive sample; p ix + represents the lane-changing probability predicted from the x-th time series in the i-th positive sample of the vehicle early lane-changing intention prediction model; α represents the weight; both j and y represent indices; L j - () is the loss function of the j-th negative sample; p jy -It represents the lane - changing probability predicted for the y - th time series in the j - th negative sample of the vehicle early lane - changing intention prediction model; e represents a constant; max( ) represents taking the maximum value; d represents the difference between the index of the last frame of the x - th time series and the index of the lane - changing completion moment; epoch represents the index of the training round and also represents the epoch - th training; Φ(epoch - 1) represents the mean of the lane - changing time differences of all positive samples in the (epoch - 1) - th training round; γ represents a hyperparameter; k represents the total number of positive samples; f represents an index; G f epoch-1 It represents the lane - changing time difference of the f - th positive sample in the (epoch - 1) - th training round.

[0075] Specifically, the lane - changing time difference is the difference between the index of the time series when the first lane - changing probability in each positive sample is not less than the preset lane - changing probability threshold and the index of the last time series (including the time series at the lane - changing completion moment). In this embodiment, the preset lane - changing probability threshold is taken as 0.7.

[0076] For each positive sample, the time - series window is divided from the front to the back according to time frames, and when the time - series data is input into the model for training, it is also divided from the front to the back according to time frames. The index of the last frame in the time series is used as the index of the time series; the index of the time series from 191 frames before the lane - changing completion moment to 120 frames before the lane - changing completion moment is 120, the index of the time series from 190 frames before the lane - changing completion moment to 119 frames before the lane - changing completion moment is 119, and so on. The index of the time series from 71 frames before the lane - changing completion moment to the frame at the lane - changing completion moment is 0. The time series of the time - series data is input into the model for training in the index order of 120, 119, 118, …, 0.

[0077] Therefore, each sample has 121 time series. When input into the model, 121 consecutive lane - changing probabilities will be obtained. When the probability value is not less than the preset lane - changing probability threshold, it means that a lane - change occurs. The difference is made between the index of the first time series among the 121 consecutive lane - changing probabilities that is not less than the lane - changing probability threshold and the index 0 of the last time series. For example, for a sample, when the lane - changing probability of the time series with index 83 is greater than the lane - changing probability threshold for the first time, the lane - changing time difference is 83 - 0 = 83.

[0078] In specific implementation, when training the vehicle early lane - changing intention prediction model, the closer the driving data is to the lane - changing completion moment, the easier it is for the model to learn the relationship between the driving data and the lane - changing behavior. While the driving data farther away from the lane - changing completion moment has less significant correlation, and it is more difficult for the model to learn the rules from it.

[0079] S4. Continuously collect the driving data of several target vehicles to be measured from several previous frames to the current frame at the current moment, and input it into the trained vehicle early lane-changing intention prediction model for processing to obtain the lane-changing probability of each target vehicle to be measured in real time. Compare the lane-changing probability of each target vehicle to be measured with a preset lane-changing probability threshold to obtain whether each target vehicle to be measured has changed lanes in real time.

[0080] In specific implementation, several target vehicles to be measured collected in real time are the target vehicles in front of the advancing direction of the driverless vehicle, and the specific number of target vehicles depends on the actual situation. The range of several previous frames is 47 frames (including the current frame, a total of 2 seconds) - 239 frames (including the current frame, a total of 10 seconds).

[0081] When comparing the lane-changing probability of the target vehicle to be measured with the preset lane-changing probability threshold, it is set according to the following formula to obtain whether the target vehicle to be measured has changed lanes:

[0082] f(p q ) = 1, p q ≥P th

[0083] f(p q ) = 0, p q <P th

[0084] Wherein, f(p q ) represents whether the qth target vehicle to be measured has changed lanes; p q represents the lane-changing probability predicted by inputting the continuous driving data of the qth target vehicle to be measured into the trained vehicle early lane-changing intention prediction model; 1 represents changing lanes; 0 represents not changing lanes; P th represents the preset lane-changing probability threshold.

[0085] According to whether all target vehicles to be measured have changed lanes obtained in real time, the driverless vehicle makes corresponding driving adjustments, such as making lane-changing and decelerating operations.

[0086] As Figure 5 shown, it is the result graph of the lane-changing probability predicted by the vehicle early lane-changing intention prediction model in this embodiment on the lane-changing data set collected and processed on the expressway in a certain city. It can be seen from the graph that the lane-changing of the target vehicle has been well predicted in the time series corresponding to the first 111 frames of the time series including the lane-changing completion moment.

[0087] Embodiment 2

[0088] In this embodiment, the same method as steps S1 - S3 in Embodiment 1 is used to conduct experiments on the existing domestic and foreign traffic dataset SQM1. In the SQM1 dataset, the sections where the road narrows are located at 90 meters to 150 meters and 210 meters to 270 meters in the east - west direction. These driving data are deleted to avoid the mis - learning of the vehicle early lane - changing intention prediction model. There are 1041 target vehicles in the SQM1 dataset, including 722 positive samples (lane - changing vehicles) and 319 negative samples (non - lane - changing vehicles). In this embodiment, the preset lane - changing probability threshold is taken as 0.9.

[0089] The conditions for dividing the lane - changing dataset are as follows: In this embodiment, N is taken as 120 frames, which is a total of 5 seconds of driving data. The time - series window is divided in the order from the front to the back of the time frames. The sliding window size M used is 72 frames (a total of 3 seconds), and the sliding step is one frame. The samples are divided into time - series windows in this way to obtain 49 time - series, and the 49 obtained time - series are combined to obtain the time - series data of a sample.

[0090] The lane - changing probability results of the vehicle early lane - changing intention prediction model in this embodiment on the SQM1 dataset are as Figure 6 shown by the KAN_punitive lane - changing weight loss function curve in. It can be seen that 2 seconds before the lane - changing completion moment, the vehicle early lane - changing intention prediction model in this embodiment can predict a very high lane - changing probability.

[0091] Comparative Example 1

[0092] In this comparative example, the same method, dataset, and lane - changing dataset division conditions as in Embodiment 2 are used to conduct experiments on the existing domestic and foreign traffic dataset SQM1. However, in this comparative example, the loss function is modified to the conventional cross - entropy loss function.

[0093] The lane - changing probability results of the vehicle early lane - changing intention prediction model in this comparative example on the SQM1 dataset are as Figure 6 shown by the KAN_cross - entropy loss function curve in. It can be seen from Figure 6 that the lane - changing probability predicted early by the vehicle early lane - changing intention prediction model in this comparative example before the lane - changing completion moment is much lower than that of the vehicle early lane - changing intention prediction model in Embodiment 2.

[0094] Comparative Example 2

[0095] In this comparative example, the same method, dataset, and lane - changing dataset division conditions as in Embodiment 2 are used to conduct experiments on the existing domestic and foreign traffic dataset SQM1. However, in this comparative example, the vehicle early lane - changing intention prediction model uses a long short - term memory network (LSTM) with a cross - entropy loss function.

[0096] The lane - changing probability results of the vehicle early lane - changing intention prediction model of this comparative example on the SQM1 dataset are as follows Figure 6 shown by the LSTM_punitive lane - changing weight loss function curve in Figure 6 It can be seen that the lane - changing probabilities predicted by the vehicle early lane - changing intention prediction model of this comparative example in the mid - term before the lane - changing completion moment and at the final lane - changing completion moment are both lower than those of the vehicle early lane - changing intention prediction model in Embodiment 2.

[0097] Comparative Example 3

[0098] This comparative example conducts experiments in the existing domestic and foreign traffic dataset SQM1 using the same method, dataset, and lane - changing dataset division conditions as in Embodiment 2. However, in this comparative example, the vehicle early lane - changing intention prediction model uses a long short - term memory network (LSTM) with a punitive lane - changing weight loss function.

[0099] The lane - changing probability results of the vehicle early lane - changing intention prediction model of this comparative example on the SQM1 dataset are as follows Figure 6 shown by the LSTM_cross - entropy loss function curve in Figure 6 It can be seen that the performance of the vehicle early lane - changing intention prediction model of this comparative example is far lower than that of the vehicle early lane - changing intention prediction model used in Embodiment 2.

[0100] From Figure 6 the comparison of the curve results of the four vehicle early lane - changing intention prediction models in, it can be seen that the method of the present invention can greatly increase the prediction time lead and output the prediction value to assist in the decision - making of driverless behavior, increase the prediction time lead and output its prediction value to assist in the decision - making of driverless behavior.

[0101] The method of the present invention processes the collected target vehicle driving data through a KAN network based on a punitive lane - changing weight loss function. While ensuring the correct rate of predicting the lane - changing behavior of the vehicle, it improves the prediction time lead of the target vehicle's lane - changing, increases the reaction time to changes in the surrounding environment, and has stronger practical significance compared with traditional methods for predicting lane - changing behavior, and can better ensure traffic safety in the mixed environment of manned and driverless vehicles.

[0102] The present invention is not limited to the embodiments described above. The above description of the specific embodiments is intended to describe and illustrate the technical solutions of the present invention. The above - mentioned specific embodiments are merely illustrative and not restrictive. Without departing from the spirit of the present invention and the scope protected by the claims, those of ordinary skill in the art can make many specific transformations in various forms under the inspiration of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the early lane-changing intention of a vehicle based on an improved KAN network, characterized in that, The steps include: S1. Collect driving data of several target vehicles through the onboard sensors of the unmanned vehicle, and divide the driving data of all target vehicles into positive samples and negative samples; The step S1 is specifically as follows: S11, collecting driving data of each frame of several target vehicles through the onboard sensor of the unmanned vehicle; S12. Classify the target vehicle into a lane-changing vehicle and a non-lane-changing vehicle based on the driving data; S13. For a lane-changing vehicle, the continuous driving data consisting of the target vehicle's lane-changing completion time and N-1 frames before the lane-changing completion time is used as a positive sample, and the continuous driving data from 2N-1 frames before the lane-changing completion time to the N frames before the lane-changing completion time is used as a negative sample. For a vehicle that has not changed lanes, any N consecutive frames of driving data are used as a negative sample. S2. Preprocess the driving data according to the divided positive samples and negative samples to obtain time series data of several samples. All time series data are aggregated to obtain a lane change dataset. S3. Constructing a vehicle early lane change intention prediction model, inputting the lane change dataset into the vehicle early lane change intention prediction model using a penalized lane change weight loss function for training, and obtaining a trained vehicle early lane change intention prediction model; The penalty lane-changing weight loss function in step S3 is set according to the following formula: L i + (p t + ) = -∑ x=1 N-M+1 α log(p ix + ) L j - (p t - )=-∑ y=1 N-M+1 log(1 - p jy - ) α=e (-max(0，d-Φ(epoch-1)-γ)) Φ(epoch -1)=1 / k∑ f=1 k G f epoch-1 where both \(i\) and \(x\) represent indices; \(L\) i + () is the loss function of the \(i\)-th positive sample; \(p\) ix + represents the lane-changing probability predicted for the \(x\)-th time series in the \(i\)-th positive sample; \(\alpha\) represents the weight; both \(j\) and \(y\) represent indices; \(L\) j - () is the loss function of the \(j\)-th negative sample; \(p\) jy - represents the lane-changing probability obtained for the \(y\)-th time series in the \(j\)-th negative sample; \(e\) represents a constant; \(\max(\ )\) represents taking the maximum value; \(d\) represents the difference between the index of the last frame of the \(x\)-th time series and the index of the lane-changing completion moment; \(epoch\) represents the index of the training round and also represents the \(epoch\)-th training; \(\varPhi(epoch - 1)\) represents the mean of the lane-changing time differences of all positive samples in the \((epoch - 1)\)-th training round; \(\gamma\) represents a hyperparameter; \(k\) represents the total number of positive samples; \(f\) represents an index; \(G\) f epoch-1 represents the lane-changing time difference of the \(f\)-th positive sample in the \((epoch - 1)\)-th training round; The lane-changing time difference is specifically the difference between the index of the time series when the first lane-changing probability in each positive sample is not less than the preset lane-changing probability threshold and the index of the last time series; S4. Continuous driving data from several frames before the current moment to the current moment frame of several target vehicles to be tested are collected in real time, and the data is input into a trained vehicle early lane change intention prediction model for processing to obtain the lane change probability of each target vehicle to be tested in real time. The lane change probability of each target vehicle to be tested is compared with a preset lane change probability threshold to obtain in real time whether each target vehicle to be tested has changed lanes.

2. The method for predicting vehicle lane change intention in the early stage based on the improved KAN network according to claim 1, characterized in that: The target vehicle is the vehicle in front of the unmanned vehicle in the adjacent lane on the left or right side; the driving data includes speed, left longitudinal distance, right longitudinal distance and lateral distance; the left longitudinal distance is the longitudinal distance between the target vehicle and the vehicle in front of the target vehicle in the left adjacent lane; the right longitudinal distance is the longitudinal distance between the target vehicle and the vehicle in front of the target vehicle in the right adjacent lane; the lateral distance is the distance between the center point of the target vehicle and the center line of the lane in which it is located; the lane change completion time is the time when the center point of the target vehicle coincides with the lane line on one side of the target vehicle.

3. The vehicle early lane change intention prediction method based on the improved KAN network according to claim 1, wherein The step S2 is specifically as follows: S21. Divide the sample into time series windows using a sliding window size of M frames and a sliding step size of one frame to obtain N-M+1 time series. The obtained N-M+1 time series are combined to obtain the time series data of one sample. S22: Processing is performed in the same manner as step S21 to obtain time series data of several samples, and the time series data of all samples are aggregated to obtain a lane change dataset.

4. A method for predicting early lane - changing intention of a vehicle based on an improved KAN network according to claim 1, characterized in that: The vehicle early lane - changing intention prediction model in step S3 is a KAN network.

5. A method for predicting early lane - changing intention of a vehicle based on an improved KAN network according to claim 1, characterized in that: When comparing the lane - changing probability of the vehicle to be measured in step S4 with a preset lane - changing probability threshold, it is set according to the following formula to obtain whether the vehicle to be measured changes lanes: f(p q ) = 1, p q ≥ P th f(p q ) = 0, p q <P th Among them, f(p q ) indicates whether the q-th target vehicle to be measured changes lanes; p q represents the lane-changing probability predicted by inputting the continuous driving data of the q-th target vehicle to be measured into the trained vehicle early lane-changing intention prediction model; 1 indicates a lane change; 0 indicates no lane change; P th represents a preset lane-changing probability threshold.

6. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 5.

Images