A method for identifying assisted driving risk avoidance behavior patterns based on time series clustering

Abstract

This invention discloses a method for identifying assisted driving (ADAS) risk avoidance behavior patterns based on time series clustering. The method includes: collecting driver risk avoidance behavior data under the combined action of FCW and AEB through driving simulator experiments; performing time series preprocessing on the risk avoidance behavior data, using multivariate time series as the standard input for cluster analysis; calculating the temporal similarity between different risk avoidance samples using dynamic time warping; clustering the multivariate time series samples using the KMeans method based on DTW distance metric, and classifying patterns according to the clustering results; and further calculating posterior indices based on the obtained cluster labels for each sample to interpret the behavioral meaning of different risk avoidance patterns. This invention enables the identification and quantification of human-machine collaborative patterns throughout the entire risk avoidance process under ADAS conditions, providing a basis for driver assistance system evaluation, driver risk behavior analysis, and human-machine interaction optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and driving behavior analysis, and in particular to a method for identifying assisted driving risk avoidance behavior patterns based on time series clustering. Background Technology

[0002] Most existing methods for assessing driver avoidance behavior rely on outcome-oriented indicators such as time to collision (TTC), whether a collision was avoided, and maximum deceleration, or evaluate driver avoidance responses solely through questionnaires or discrete operational events collected by driving simulators. In vehicles equipped with Forward Collision Warning (FCW) and Automatic Emergency Braking (AEB), current technologies typically use the FCW trigger time, AEB intervention time, and whether a collision ultimately occurs as key evaluation criteria. Based on system logs or onboard sensor data, they determine parameters such as the driver's reaction time and braking intensity after a risk event. Using outcome- or single-event-oriented indicators such as reaction time exceeding a threshold after warning, driver failure to take over after AEB intervention, and failure to avoid a collision, they conduct static evaluations according to preset judgment rules to ultimately determine the safety level of the driver or the avoidance process.

[0003] The above evaluation method is a results-oriented assessment method. This assessment method mainly focuses on the final result of risk avoidance, ignoring the dynamic operation evolution characteristics of the driver throughout the entire risk evolution process, such as the rhythm of accelerator and manual braking, the timing coordination between driver and system braking, and the dynamic changes in the human-machine dominant relationship. It cannot make a quantitative and patterned evaluation of the continuous time-varying risk avoidance process, and lacks a comprehensive identification and analysis of the driver and system collaborative risk avoidance behavior under ADAS intervention conditions. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method for identifying assisted driving risk avoidance behavior patterns based on time series clustering, which can realize the identification and difference quantification of human-machine collaborative patterns in the entire process of driver risk avoidance under ADAS conditions, and provide a basis for driving assistance system evaluation, driver risk behavior analysis and human-machine interaction optimization.

[0005] The present invention solves the technical problem by adopting the following technical solution:

[0006] A method for identifying assisted driving risk avoidance behavior patterns based on time series clustering includes the following steps:

[0007] Step S1, Data Acquisition: Data on the driver's hazard avoidance behavior under the combined action of FCW and AEB is collected through driving simulator experiments;

[0008] Step S2, Time Series Preprocessing: Perform time series preprocessing on the risk avoidance behavior data, and use the preprocessed multivariate time series as the standard input for cluster analysis;

[0009] Step S3, DTW distance calculation: The temporal similarity between different risk avoidance samples is calculated using the dynamic time warping method;

[0010] Step S4, KMeans clustering and pattern segmentation: After obtaining the DTW distance between samples, the KMeans method based on the DTW distance metric is used to cluster the multivariate time series samples, and the pattern is segmented according to the clustering results.

[0011] Step S5, Posterior index analysis and pattern interpretation: After clustering is completed, the clustering input variables are no longer used to reclassify the categories. Instead, posterior indices are calculated based on the cluster labels obtained for each sample to explain the behavioral meaning of different risk avoidance patterns.

[0012] Furthermore, in step S1, the method for experimentally collecting driver avoidance behavior data under the combined action of FCW and AEB includes:

[0013] First, the trigger point of the risk event is determined based on the moment when the vehicle in front begins to urgently cut out and expose the stationary obstacle in the experimental scenario, and this moment is uniformly recorded as time zero. Then, starting from this time zero, a fixed-duration emergency avoidance window is extracted, covering the key processes of FCW warning triggering, driver throttle release, manual braking establishment, AEB intervention, and risk clearance or collision occurrence. Within this window, continuous data of vehicle speed, longitudinal acceleration, throttle pedal depth, manual braking force value, and mechanical braking force value, as well as discrete variables of FCW state, are extracted simultaneously.

[0014] Furthermore, in step S2, the method for time-series preprocessing of the risk-avoidance behavior data includes:

[0015] Using the risk event trigger point as a unified time zero point, the time axes of different samples are aligned, and emergency avoidance windows of the same length are extracted. Subsequently, the original high-frequency data is resampled at fixed time intervals. Continuous variables are obtained with unified sampling points using window mean or linear interpolation, while state variables are assigned values ​​based on the main state value or trigger state value within the window, thus ensuring that each sample has a consistent time step and sequence length. After downsampling, continuous variables such as vehicle speed, longitudinal acceleration, accelerator pedal depth, manual braking force, and mechanical braking force are standardized. Discrete variables of FCW state are encoded with 0 and 1. Finally, the processed variables are concatenated in chronological order.

[0016] Furthermore, in step S3, the method for calculating the temporal similarity between different risk-averse samples using dynamic time warping includes:

[0017] Each hazard avoidance sample is represented as a multivariate time series consisting of velocity, longitudinal acceleration, accelerator pedal depth, manual braking force value, machine braking force value, and FCW state. For any two hazard avoidance samples, the local distance between their multivariate feature vectors at different time points is first calculated. Then, using a dynamic programming method, under the conditions that the time order is not reversed, the matching path is continuous, and the start and end points correspond, the matching path with the minimum cumulative distance between the two sequences is found, and this minimum cumulative distance is taken as the DTW distance between the two samples.

[0018] Furthermore, in step S4, the method of clustering multivariate time series samples using the KMeans method based on DTW distance metric includes:

[0019] First, a range of candidate cluster numbers is set, and the clustering model is run under different cluster number conditions. The model first initializes several cluster centers, then calculates the DTW distance between each sample and each cluster center, and assigns the sample to the nearest cluster. After the sample assignment is completed, the cluster centers are recalculated using the DTW-based time series averaging method so that the cluster centers can represent the typical risk avoidance evolution process of this type of sample. The sample assignment-cluster center update process is iterated repeatedly until the cluster assignment result is stable or the preset number of iterations is reached.

[0020] Furthermore, in step S4, the method for pattern segmentation based on the clustering results includes:

[0021] To determine a reasonable number of patterns, the intra-cluster error and silhouette coefficient were calculated for different numbers of clusters. The elbow method was used to determine the inflection point of the decreasing trend of intra-cluster error, and the silhouette coefficient was used to evaluate the intra-cluster compactness and inter-cluster separation, and finally the optimal number of clusters was determined. After clustering, each valid sample was assigned a corresponding category label, and the samples were divided into several typical risk avoidance behavior patterns based on the cluster center sequence, the characteristics of input variable changes, and the differences in posterior indices.

[0022] Furthermore, in step S5, the method for posterior index analysis includes:

[0023] First, system intervention indicators are calculated based on FCW status, AEB status, or machine braking force value, including FCW duration, AEB duration, and AEB Level 2 duration, to characterize the intervention intensity of the warning system and automatic braking system during the hazard avoidance process. Then, human-machine interaction indicators are calculated based on the relationship between manual braking force value and machine braking force value, including machine dominance ratio, manual dominance ratio, and driver takeover time difference, to determine whether the hazard avoidance process is dominated by the system, the driver, or a combination of both. Finally, interaction result indicators indicating successful takeover are determined based on whether the driver applies effective manual braking, takes over in a timely manner, and completes risk avoidance.

[0024] Furthermore, in step S5, the method for pattern interpretation includes:

[0025] After obtaining the above-mentioned interaction result indicators, the mean, standard deviation, and distribution characteristics of each indicator were statistically analyzed according to the cluster category, and the differences between different clusters were compared using significance tests. For continuous posterior indicators, analysis of variance or Kruskal-Wallis test was used to test the overall differences, and pairwise comparisons were further performed when the overall differences were significant. For categorical indicators such as takeover success signs, chi-square test or Fisher's exact test was used for comparison. Finally, combining the temporal characteristics of the input variables and the differences in posterior indicators of each cluster, the patterns of different clusters were interpreted and named, thereby completing the quantitative analysis of driver risk avoidance behavior patterns.

[0026] The present invention discloses a method for identifying assisted driving risk avoidance behavior patterns based on time series clustering, which has the following beneficial effects:

[0027] 1. Dynamic identification of risk avoidance process based on time series clustering

[0028] This invention uses DTW as a similarity metric to perform nonlinear time axis alignment on the continuous risk avoidance process after the risk event is triggered. It overcomes the shortcomings of traditional Euclidean distance clustering, which requires point-by-point synchronization and is prone to misclassifying samples with similar shapes but misaligned time sequences as different categories. This makes the clustering results truly reflect the morphological similarity of risk avoidance behavior and improves the accuracy of pattern recognition.

[0029] 2. A quantitative evaluation system for human-machine collaborative risk avoidance mode under ADAS conditions was constructed.

[0030] This invention extracts multivariate time-series data and uses DTW-KMeans clustering to classify typical patterns of driver risk avoidance behavior. Posterior indices are introduced to quantitatively interpret different patterns from multiple dimensions, including system intervention intensity, human-machine dominance, and interaction outcomes. This evaluation system comprehensively characterizes the collaborative features between the driver and the system under ADAS intervention conditions, overcoming the shortcomings of existing methods that only focus on collision outcomes or discrete events.

[0031] 3. Automated identification and classification of driving risk avoidance behavior patterns have been achieved.

[0032] This invention, based on time-series data, establishes a complete process for identifying risk-avoidance behavior patterns through preprocessing, DTW distance calculation, KMeans clustering, and posterior index analysis. This method can complete the entire process from raw data input to typical pattern output without manual annotation or pre-set behavior templates, exhibiting strong operability and promotional value. The identified pattern classification results can provide quantitative basis for optimizing driver assistance system parameters, providing early warning of driver risk behavior, and designing human-machine interfaces. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method of the present invention;

[0034] Figure 2 This is a schematic diagram of the results of elbow-based clustering of driving behavior patterns;

[0035] Figure 3 It is a time-series distribution diagram of velocity;

[0036] Figure 4 It is a time-series distribution diagram of acceleration;

[0037] Figure 5 It is a time-series distribution diagram of the accelerator pedal depth;

[0038] Figure 6 It is a time-series distribution diagram of the artificial braking force values;

[0039] Figure 7 This is the timing distribution diagram of the FCW state;

[0040] Figure 8 It is a time-series distribution diagram of the machine's braking force values. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] This invention, based on driving avoidance behavior data under the synergistic effect of FCW and AEB, extracts speed, longitudinal acceleration, accelerator pedal depth, manual braking force value, machine braking force value, and FCW status within the emergency avoidance window after a risk event is triggered as multivariate time series inputs. It employs Dynamic Time Warping (DTW) combined with KMeans clustering to perform pattern recognition on the driver's continuous avoidance process, classifying the samples into typical avoidance behavior patterns. Posterior indices are used to quantitatively analyze the system intervention intensity, human-machine dominance relationship, and takeover effect of different patterns.

[0043] In urban road following scenarios, when a vehicle ahead suddenly cuts out, revealing a stationary obstacle, the driver faces a collision risk, and both FCW (Forward Collision Warning) and AEB (Automatic Emergency Braking) systems will be triggered successively. This invention collects emergency avoidance process data involving ADAS (Advanced Driver Assistance Systems) in driving simulator experiments, extracts multivariate time series within the key avoidance window after the risk event is triggered, and uses dynamic time warping combined with KMeans clustering to perform pattern recognition on the driver's continuous avoidance process, ultimately identifying typical avoidance behavior patterns.

[0044] refer to Figure 1 This invention provides a method for identifying assisted driving risk avoidance behavior patterns based on time series clustering, comprising the following steps:

[0045] Step S1, Data Acquisition: Data on the driver's hazard avoidance behavior under the combined action of FCW and AEB is collected through driving simulator experiments. The methods include:

[0046] First, the trigger point of the risk event is determined based on the moment when the vehicle in front begins to urgently cut out and expose the stationary obstacle in the experimental scenario, and this moment is uniformly recorded as time zero. Then, starting from this time zero, an emergency avoidance window of fixed duration is extracted, covering key processes such as FCW warning triggering, driver throttle release, manual braking establishment, AEB intervention, and risk clearance or collision occurrence. Within this window, continuous data such as vehicle speed, longitudinal acceleration, throttle pedal depth, manual braking force value, and mechanical braking force value, as well as discrete variables of FCW state, are extracted simultaneously.

[0047] Step S2, Time Series Preprocessing: Perform time series preprocessing on the risk avoidance behavior data, and use the preprocessed multivariate time series as the standard input for cluster analysis;

[0048] The clustering input variables selected in this invention cover three levels: vehicle operating status, driver control, and vehicle control, totaling six variables: speed (representing the longitudinal speed change of the main vehicle), longitudinal acceleration (representing the deceleration / acceleration dynamics of the main vehicle), accelerator pedal depth (representing the driver's deceleration behavior), manual braking force value (representing the intensity of the driver's active braking), mechanical braking force value (representing the degree of braking intervention of the AEB system), and FCW status (representing the FCW warning triggering status). All variables are time-series data, with a fixed-length emergency avoidance window extracted from the moment the risk event is triggered. Specific clustering input variables are shown in Table 1.

[0049]

[0050] Table 1

[0051] Methods for time series preprocessing of risk aversion behavior data include:

[0052] Using the risk event trigger point as a unified time zero point, the time axes of different samples are aligned, and emergency avoidance windows of the same length are extracted. Subsequently, the original high-frequency data is resampled at fixed time intervals. Continuous variables are obtained with unified sampling points using window mean or linear interpolation, while state variables are assigned values ​​based on the main state value or trigger state value within the window, thus ensuring that each sample has a consistent time step and sequence length. After downsampling, continuous variables such as vehicle speed, longitudinal acceleration, accelerator pedal depth, manual braking force, and machine braking force are standardized, typically using Z-score normalization to eliminate differences in dimensions and numerical ranges. Discrete variables such as FCW state are encoded with 0 and 1. Finally, the processed variables are concatenated in chronological order.

[0053] Step S3, DTW distance calculation: The dynamic time warping method is used to calculate the temporal similarity between different risk avoidance samples. The method includes:

[0054] Each hazard avoidance sample is represented as a multivariate time series consisting of velocity, longitudinal acceleration, accelerator pedal depth, manual braking force value, machine braking force value, and FCW state. For any two hazard avoidance samples, the local distance between their multivariate feature vectors at different time points is first calculated. Then, using a dynamic programming method, under the conditions that the time order is not reversed, the matching path is continuous, and the start and end points correspond, the matching path with the minimum cumulative distance between the two sequences is found, and this minimum cumulative distance is taken as the DTW distance between the two samples.

[0055] Step S4, KMeans clustering and pattern segmentation: After obtaining the DTW distance between samples, the KMeans method based on the DTW distance metric is used to cluster the multivariate time series samples, and the pattern is segmented according to the clustering results.

[0056] Methods for clustering multivariate time series samples using the KMeans method based on DTW distance metric include:

[0057] First, a range of candidate cluster numbers is set, and the clustering model is run under different cluster number conditions. The model first initializes several cluster centers, then calculates the DTW distance between each sample and each cluster center, and assigns the sample to the nearest cluster. After the sample assignment is completed, the cluster centers are recalculated using the DTW-based time series averaging method so that the cluster centers can represent the typical risk avoidance evolution process of this type of sample. The sample assignment-cluster center update process is iterated repeatedly until the cluster assignment result is stable or the preset number of iterations is reached.

[0058] Methods for pattern segmentation based on clustering results include:

[0059] To determine a reasonable number of patterns, the intra-cluster error and silhouette coefficient were calculated for different numbers of clusters. The elbow method was used to determine the inflection point of the decreasing trend of intra-cluster error, and the silhouette coefficient was used to evaluate the intra-cluster compactness and inter-cluster separation, and finally the optimal number of clusters was determined. After clustering, each valid sample was assigned a corresponding category label, and the samples were divided into several typical risk avoidance behavior patterns based on the cluster center sequence, the characteristics of input variable changes, and the differences in posterior indices.

[0060] Step S5, Posterior index analysis and pattern interpretation: After clustering is completed, the clustering input variables are no longer used to reclassify the categories. Instead, posterior indices are calculated based on the cluster labels obtained for each sample to explain the behavioral meaning of different risk avoidance patterns.

[0061] Methods for posterior index analysis include:

[0062] First, system intervention indicators are calculated based on FCW status, AEB status, or machine braking force value, including FCW duration, AEB duration, and AEB Level 2 duration, to characterize the intervention intensity of the warning system and automatic braking system during the hazard avoidance process. Then, human-machine interaction indicators are calculated based on the relationship between manual braking force value and machine braking force value, including machine dominance ratio, manual dominance ratio, and driver takeover time difference, to determine whether the hazard avoidance process is dominated by the system, the driver, or a combination of both. Finally, interaction result indicators such as whether the driver applies effective manual braking, takes over in a timely manner, and completes risk avoidance are determined based on whether the takeover success indicator is established.

[0063] Methods for pattern interpretation include:

[0064] After obtaining the aforementioned interaction result indicators, the mean, standard deviation, and distribution characteristics of each indicator were statistically analyzed according to the cluster category, and significance tests were used to compare the differences between different clusters. For continuous posterior indicators, analysis of variance or the Kruskal-Wallis test was used to test the overall differences, and pairwise comparisons were further performed when the overall differences were significant. For categorical indicators such as takeover success signs, chi-square tests or Fisher's exact tests were used for comparison. Finally, combining the temporal characteristics of the input variables and the differences in posterior indicators for each cluster, pattern interpretation and naming were performed for different clusters, thereby completing the quantitative analysis of driver risk avoidance behavior patterns. The posterior indicators are shown in Table 2.

[0065]

[0066] Table 2

[0067] This invention employs Time-Depth Wave (DTW) as a similarity metric to perform nonlinear time-axis alignment of continuous avoidance processes after a risk event is triggered. This overcomes the shortcomings of traditional Euclidean distance clustering, which requires point-by-point synchronization and is prone to misclassifying morphologically similar but temporally misaligned samples as different categories. This ensures that the clustering results truly reflect the morphological similarity of avoidance behaviors, improving the accuracy of pattern recognition. This invention extracts multivariate time-series data and uses DTW-KMeans clustering to classify typical patterns of driver avoidance behavior. Posterior indices are introduced to quantify different patterns from multiple dimensions, including system intervention intensity, human-machine dominance, and interaction results. This evaluation system comprehensively characterizes the collaborative features of the driver and system under ADAS intervention conditions, overcoming the shortcomings of existing methods that only focus on collision results or discrete events. Based on time-series data, this invention forms a complete avoidance behavior pattern recognition process through preprocessing, DTW distance calculation, KMeans clustering, and posterior index analysis. This method can complete the entire process from raw data input to typical pattern output without manual annotation or pre-set behavior templates, possessing strong operability and promotional value. The pattern classification results obtained can provide quantitative basis for optimizing driving assistance system parameters, warning of driver risk behavior, and design of human-machine interface.

[0068] Example

[0069] Taking an emergency lane-keeping scenario in urban roads as an example, the dataset contains 165 valid ADAS (Advanced Driver Assistance Systems) risk avoidance samples. Samples were extracted based on the emergency avoidance window following the triggering of the risk event, removing those with abnormal data or incomplete data collection. The effective time window ranges from 0 to 5 seconds. The main input variables for this risk avoidance process are vehicle speed, longitudinal acceleration, accelerator pedal depth, manual braking force, mechanical braking force, and FCW (Forward Collision Warning) status, as detailed below:

[0070] 1) Data Acquisition and Preprocessing

[0071] Data on driver avoidance behavior under the combined effects of FCW and AEB were collected through driving simulator experiments at a sampling frequency of 50Hz. Starting from the moment the risk event was triggered, an emergency avoidance window of 0-5 seconds was captured, resulting in 165 valid samples. Six time-series variables were extracted from each sample. The original high-frequency sequences were downsampled to 10Hz to standardize the time length of each sample, and Z-score normalization was performed on each variable to eliminate the influence of dimensions.

[0072] 2) DTW distance calculation and KMeans clustering

[0073] Using Time-Divergence Written (DTW) as a similarity measure between samples, pairwise distances were calculated for the multivariate time series data of 165 samples. After obtaining the DTW distance matrix, the KMeans clustering algorithm was used to partition the samples. Combining the elbow method and silhouette coefficient, the optimal number of clusters was determined to be 3, dividing all samples into 3 typical risk-avoidance behavior patterns. Figure 2 .

[0074] 3) Time series characteristics and significance analysis of input variables

[0075] Descriptive statistics and significance tests were performed on the six input variables of the three clusters. The results showed significant differences (p<0.05) among the three clusters in speed, longitudinal acceleration, accelerator pedal depth, manual braking force, mechanical braking force, and FCW status. Specifically, cluster 2 had the highest mean speed (approximately 16.8 m / s), the highest mean accelerator pedal depth (approximately 18.2%), the lowest mean manual braking force (approximately 18.5 N), and the highest mean mechanical braking force (approximately 32.1 N). Cluster 3 had the lowest mean speed (approximately 12.5 m / s), the lowest mean accelerator pedal depth (approximately 5.6%), the highest mean manual braking force (approximately 58.3 N), and the lowest mean mechanical braking force (approximately 15.2 N). Cluster 1 had all indicators at intermediate levels. Figures 3 to 8 On the timing curves, cluster 3 has the fastest throttle decrease and the earliest manual braking is established, cluster 2 has the slowest throttle decrease and the manual braking is significantly delayed, while cluster 1 shows a stable human-machine joint participation characteristic.

[0076] 4) Posterior index analysis

[0077] The system intervention indicators, human-computer interaction indicators, and interaction outcome indicators of the three clusters were compared, and the descriptive statistics of the posterior indicators are shown in Table 3.

[0078]

[0079] Table 3

[0080] The results show that cluster 2 had higher FCW duration, AEB duration, and AEB Level 2 duration than clusters 1 and 3, indicating the strongest system intervention intensity. Cluster 2 had a significantly higher machine-dominated ratio than the other two clusters, while its human-dominated ratio was significantly lower, suggesting that its hazard avoidance process was primarily system-driven. Cluster 3 had a negative driver takeover time difference, indicating that the driver initiated braking earlier than the system, demonstrating the strongest tendency for active takeover. Clusters 1 and 2 had positive takeover time differences, indicating that driver braking lagged behind the system. Regarding takeover success rate, clusters 1 and 3 were significantly higher than cluster 2, suggesting that although cluster 2 had a high system intervention intensity, its effective driver takeover was insufficient.

[0081] 5) Naming each cluster

[0082] Based on the descriptive statistics, temporal evolution characteristics, and posterior index analysis results of the input variables, three clusters were named: Cluster 1 exhibits a cooperative response characteristic with both driver and system participation and relatively obvious manual and machine braking, and is named "Human-Machine Cooperative Braking Type"; Cluster 2 exhibits a human-machine game characteristic with insufficient driver active braking, continuous intervention of system warnings and automatic braking, and low takeover success rate, and is named "Human-Machine Game Weak Takeover Type"; Cluster 3 exhibits a characteristic with the driver initiating braking earlier than the system, high intensity of manual braking, and strong tendency to take over actively, and is named "Active Takeover Forced Action Type".

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A time series clustering-based auxiliary driving risk-avoiding behavior pattern recognition method, characterized in that, Includes the following steps: Step S1, Data Acquisition: Data on the driver's hazard avoidance behavior under the combined action of FCW and AEB is collected through driving simulator experiments; Step S2, Time Series Preprocessing: Perform time series preprocessing on the risk avoidance behavior data, and use the preprocessed multivariate time series as the standard input for cluster analysis; Step S3, DTW distance calculation: The temporal similarity between different risk avoidance samples is calculated using the dynamic time warping method; Step S4, KMeans clustering and pattern segmentation: After obtaining the DTW distance between samples, the KMeans method based on the DTW distance metric is used to cluster the multivariate time series samples, and the pattern is segmented according to the clustering results. Step S5, Posterior index analysis and pattern interpretation: After clustering is completed, the clustering input variables are no longer used to reclassify the categories. Instead, posterior indices are calculated based on the cluster labels obtained for each sample to explain the behavioral meaning of different risk avoidance patterns.

2. The time series clustering-based auxiliary driving behavior pattern recognition method of claim 1, wherein, In step S1, the method for experimentally collecting driver avoidance behavior data under the combined action of FCW and AEB includes: First, the trigger point of the risk event is determined based on the moment when the vehicle in front begins to urgently cut out and expose the stationary obstacle in the experimental scenario, and this moment is uniformly recorded as time zero. Then, starting from this time zero, a fixed-duration emergency avoidance window is extracted, covering the key processes of FCW warning triggering, driver throttle release, manual braking establishment, AEB intervention, and risk clearance or collision occurrence. Within this window, continuous data of vehicle speed, longitudinal acceleration, throttle pedal depth, manual braking force value, and mechanical braking force value, as well as discrete variables of FCW state, are extracted simultaneously.

3. The time series clustering-based auxiliary driving behavior pattern recognition method of claim 2, wherein, In step S2, the method for time series preprocessing of risk avoidance behavior data includes: Using the risk event trigger point as a unified time zero point, the time axes of different samples are aligned, and emergency avoidance windows of the same length are extracted. Subsequently, the original high-frequency data is resampled at fixed time intervals. Continuous variables are obtained with unified sampling points using window mean or linear interpolation, while state variables are assigned values ​​based on the main state value or trigger state value within the window, thus ensuring that each sample has a consistent time step and sequence length. After downsampling, continuous variables such as vehicle speed, longitudinal acceleration, accelerator pedal depth, manual braking force, and mechanical braking force are standardized. Discrete variables of FCW state are encoded with 0 and 1. Finally, the processed variables are concatenated in chronological order.

4. The time series clustering-based auxiliary driving behavior pattern recognition method of claim 3, wherein, In step S3, the method for calculating the temporal similarity between different risk aversion samples using dynamic time warping includes: Each hazard avoidance sample is represented as a multivariate time series consisting of velocity, longitudinal acceleration, accelerator pedal depth, manual braking force value, machine braking force value, and FCW state. For any two hazard avoidance samples, the local distance between their multivariate feature vectors at different time points is first calculated. Then, using a dynamic programming method, under the conditions that the time order is not reversed, the matching path is continuous, and the start and end points correspond, the matching path with the minimum cumulative distance between the two sequences is found, and this minimum cumulative distance is taken as the DTW distance between the two samples.

5. The time series clustering-based auxiliary driving behavior pattern recognition method of claim 4, wherein, In step S4, the method for clustering multivariate time series samples using the KMeans method based on DTW distance metric includes: First, a range of candidate cluster numbers is set, and the clustering model is run under different cluster number conditions. The model first initializes several cluster centers, then calculates the DTW distance between each sample and each cluster center, and assigns the sample to the nearest cluster. After the sample assignment is completed, the cluster centers are recalculated using the DTW-based time series averaging method so that the cluster centers can represent the typical risk avoidance evolution process of this type of sample. The sample assignment-cluster center update process is iterated repeatedly until the cluster assignment result is stable or the preset number of iterations is reached.

6. The time series clustering-based auxiliary driving behavior pattern recognition method of claim 5, wherein, In step S4, the method for pattern segmentation based on the clustering results includes: To determine a reasonable number of patterns, the intra-cluster error and silhouette coefficient were calculated for different numbers of clusters. The elbow method was used to determine the inflection point of the decreasing trend of intra-cluster error, and the silhouette coefficient was used to evaluate the intra-cluster compactness and inter-cluster separation, and finally the optimal number of clusters was determined. After clustering, each valid sample was assigned a corresponding category label, and the samples were divided into several typical risk avoidance behavior patterns based on the cluster center sequence, the characteristics of input variable changes, and the differences in posterior indices.

7. The time series clustering-based auxiliary driving behavior pattern recognition method of claim 6, wherein, In step S5, the method for posterior index analysis includes: First, system intervention indicators are calculated based on FCW status, AEB status, or machine braking force value, including FCW duration, AEB duration, and AEB Level 2 duration, to characterize the intervention intensity of the warning system and automatic braking system during the hazard avoidance process. Then, human-machine interaction indicators are calculated based on the relationship between manual braking force value and machine braking force value, including machine dominance ratio, manual dominance ratio, and driver takeover time difference, to determine whether the hazard avoidance process is dominated by the system, the driver, or a combination of both. Finally, interaction result indicators indicating successful takeover are determined based on whether the driver applies effective manual braking, takes over in a timely manner, and completes risk avoidance.

8. The time series clustering-based auxiliary driving behavior pattern recognition method of claim 7, wherein, In step S5, the method for pattern interpretation includes: After obtaining the above-mentioned interaction result indicators, the mean, standard deviation, and distribution characteristics of each indicator were statistically analyzed according to the cluster category, and the differences between different clusters were compared using significance tests. For continuous posterior indicators, analysis of variance or Kruskal-Wallis test was used to test the overall differences, and pairwise comparisons were further performed when the overall differences were significant. For categorical indicators such as takeover success signs, chi-square test or Fisher's exact test was used for comparison. Finally, combining the temporal characteristics of the input variables and the differences in posterior indicators of each cluster, the patterns of different clusters were interpreted and named, thereby completing the quantitative analysis of driver risk avoidance behavior patterns.

Images