Data-Driven Modeling of Lane Changing on Freeways Application for Automated Vehicles in Mixed-Traffic

Abstract

Lane changing is known to have a substantial influence on traffic flow characteristics due to its interfering effect on the surrounding traffic, and most importantly, increases drivers’ workload and stress level. In recent years, the technological innovation in the automotive industry accelerated the emergence of automated vehicles allowing technology to alleviate some of the demanding driving tasks including lane changing maneuvers. One of the significant challenges continues to be the penetration of automated vehicles into the current transportation system and the associated mixed-traffic. Operating automated vehicles in conjunction with human drivers introduces mixed traffic complexities that present unique challenges to effectively predict the driving and mimic overall social behaviors of drivers.
A commonly used tool for developing and evaluating the impact of automated vehicles in various scenarios including mixed-traffic situations is microscopic traffic simulation. Unfortunately, most of the available microscopic simulation tools are based on rule-based lane changing models. These models do not account for all factors including driving behavior norms and heterogeneous driving behavior, which – as the research suggests – ultimately results in unrealistic and biased findings due to the fact that the simulation does not account for realizable trade-off among the factors such as speed, acceleration, jerk, and safety when performing the task of driving.
The goal of this dissertation is to provide a generic data-driven approach for modeling the lane changing behavior via a series of data gathering techniques based entirely on real world observations. The proposed approach allows the lane changing model to dynamically assess and learn socially acceptable driving behavior based on observations of the surrounding driving environment including other human drivers facing the same real world conditions. This method provides a realistic, non rule based model that accounts for both the tactical and operational behavior of lane changing. The proposed approach also accounts for stochasticity among human drives, which enables more accurate investigations and outcome projections using traditional microscopic traffic simulation.
This modeling technique requires real behavior data from driving of human drivers, and once collected, uses it to develop data driven models for the tactical and operational behavior of drivers for the application in microscopic traffic simulations. The data was collected from traffic video recordings at an on-ramp to an urban freeway in Munich. After post processing 13 hours of recorded videos, individual vehicle trajectories were extracted and processed using clustering methods to categorize human driving behaviors into three clusters: “timid”, ”moderate” and “aggressive”. Driving behaviors categorized as “aggressive” were then purposely excluded from the data to eliminate the undesirable characteristics and unsafe maneuvers from the dataset which is used subsequently for modeling the lane changing behavior for automated vehicles.
The decision to change lanes is recognized as a tactical decision and is executed in varying units of time ranging from second to minutes based entirely on immediate driving, traffic and environmental circumstances. Modeling the tactical lane changing behavior was done by deploying a supervised machine learning algorithm which simultaneously takes into account the status of the “ego vehicle”, surrounding dynamic objects such as other road users, and environmental components such as road infrastructure. Defining these parameters in the feature vector enabled the algorithm to establish the classifier as a prediction – with a high degree of certainty – on whether the driver will ultimately decide to perform a lane change to the right, to the left or remain in the current lane.
At a subconscious level, a driver executes action patterns in milliseconds to respond to the immediate traffic situation. Operational behavior modeling during lane changing was developed using ‘Inverse Reinforcement Learning’ (IRL) method. In this method, the agent’s policy or a behavioral history is an established variable, and the goal of machine learning is to determine the reward function that explains the given behavior, resulting in the automatic extraction of the respective rewards from the respective observation. This model uses the ‘Markov Decision Process’ (MDP) which is a mathematical framework for modeling decision making situations where outcomes are partly random and partly controlled by the decision maker.
Both models were integrated into the SUMO (Simulation of Urban MObility) software to evaluate the overall ability of the approach to realistically simulate traffic and compared its performance and transferability when faced with unseen events in the training dataset. The integrated model was evaluated by comparing the network throughput, the distribution of Time To Collisions (TTC) and the number of lane changes when compared with the default lane changing model in SUMO. The simulation results indicate that the proposed integrated modeling approach is capable of realistically simulating the lane changing behavior of both normal drivers and automated vehicles and it outperforms the underlying models in SUMO. Moreover, the model successfully executes plausible lane changing maneuvers in unseen traffic situations.

Full text

M.Sc. Nassim Motamedidehkordi
Technical University of Munich
Department of Civil, Geo and Environmental Engineering
Chair of Traffic Engineering and Control
Data-Driven Modeling of Lane Changing on
Freeways
Application for Automated Vehicles in Mixed-Traffic

•Introduction
•Research questions
•Traffic observations
•Clustering driving behavior
•Modeling tactical lane changing behavior
•Modeling operational lane changing behavior
•Discussion & limitations
2
Agenda

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
•Lane changing is a demanding driving task.
•Accident statistics from January-May 2019 in Germany showed the following:
•Recent developments in automotive industry has encouraged handing over some driving
tasks such as lane changing to automated vehicles.
•The goal is to design safer, more efficient, more comfortable and interactive vehicles while
maintaining the persona of an ideal human driver, without making the mistakes that a
human driver makes.
3
Topic Background
Killed
Seriously Injured
Slightly Injured
39 1201 7915

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
Challenge:
In development of automated vehicles, one of the biggest challenges remains to be mixed-traffic.
•The ways people move through the environment are already culturally and socially
encoded.
•Developing different driving behaviors for an automated vehicle by varying the model
parameters of its motion planning algorithm is not a plausible solution.
Solution:
Learning from demonstration with the goal of developing a more adaptive driving behavior for the
automated vehicles and thus enhancing social acceptance of them.
Result:
Automated vehicle can smoothly integrate into the flow of traffic and handle traffic interactions
without disrupting other road users.
4
Topic Background

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
The purpose of this study is use a data-driven approach to learn a socially acceptable tactical
and operational lane changing model from vast amount of observation of human’s driving
behavior on a freeway.
5
Study Goal

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
The purpose of this study is use a data-driven approach to learn a socially acceptable tactical
and operational lane changing model from vast amount of observation of human’s driving
behavior on a freeway.
6
Study Goal
Physical model vs. data-driven model
Forecast

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
Strategic Level
Maneuvering (Tactical)
Level
Control (Operational)
Level
Levels of Driving Behavior [Michon,1985]
The purpose of this study is use a data-driven approach to learn a socially acceptable tactical
and operational lane changing model from vast amount of observation of human’s driving
behavior on a freeway.
7
Study Goal

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
1. How can the socially acceptable driving behavior be sampled from the dataset of individual
vehicle trajectories?
2. Can the tactical and operational lane changing behavior for automated vehicles be learned
from the observations of human drivers’ lane changing behavior?
3. How can this methodology be generalized for modeling of the automated vehicle’ driving
behavior in other driving situations?
8
Research Questions

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
9
Methodology
Research question 1
Input
Traffic Observation
Tactical
Operational
Automated car
Modeling
Tactical
Human driver
Operational
Output
Model 1 Model 2
Clustering
Research question 2
Research question 3
Driving behavior
task Microsocpic lane
changing models
Motion planning &
control
Literature review
Microsocopic
simulation tools
Implementation in simulation tool & evaluation
Preprocessing

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
10
Traffic Observation

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Simulation Conclusions
Operational
Behavior
11
13 hours of video recording
47,000 road users
9300 kilometers of trajectories

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
Symbol
Notation
Description

Average
velocity

Average
relative speed to the vehicles in the scene

Average
longitudinal jerk

Average
relative speed to the preceding vehicle

Average
distance to preceding vehicle

Average
relative speed to the rear vehicle

Average
distance to the rear vehicle
12
Clustering Driving Behavior
List of extracted Features
Goal: extract a set of trajectory features that can be mapped properly to the driving
behaviors and categorize the driving behavior styles into separate driving behavior clusters.

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
-mean clustering algorithm developed by MacQueen [1967]
  



 

 the number of clusters
 the number of cases
: the centroid of each cluster

case 
13
Clustering Driving Behavior
determining the optimal number of clusters with the elbow method:

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
Silhouette coefficient:

 
: average intra-cluster distance for sample
i
: minimum average distance of sample
i
to points in different cluster
14
Validation of Results
Cluster #
Mean Silhouette Score
1
18490
0.71
2
18326
0.69
33953 0.79
Clustering result

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
15
Interpretation of the Results
Histogram of features of each cluster

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
Goal: approximating a mapping function from input variables to discrete output
variables so that the mapping function predicts a class for new observations.
16
Modeling Tactical Lane Changing Behavior
Output
Feature vector
Input
Modeling
  󰇟󰇠

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
17
Feature Vector
Core features of ego vehicle

Speed


Speed
in lane



Longitudinal
acceleration in lane



Lateral
acceleration in lane
Driving environment features

Index of the closest lane



The
mean velocity of vehicles in the scene
Vehicle relative features


Distance
between the ego vehicle and vehicle 


Relative
speed between ego vehicle and vehicle



Longitudinal
acceleration of vehicle 


Lateral
acceleration of vehicle 


Class
of vehicle 

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
18
Modeling Tactical Lane-changing Behavior
Feature vector
Input
Modeling
  󰇟󰇠
Output
Tactical behavior
No lane change Lane change to right Lane change to left
Classifier
[Motamedidehkordi et al., 2017]

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
Model
Class
Recall
Precision
F1
Human drivers
Lane change to left
0.94
0.93
0.93
No lane change
0.93
0.95
0.94
Lane change to right
0.94
0.93
0.93
Overall performance
0.94
0.94
0.94
Automated vehicles
Lane change to left
0.97
0.97
0.96
No lane change
0.97
0.97
0.97
Lane change to right
0.96
0.96
0.97
Overall performance
0.97
0.97
0.97
19
Classification Results
  

  
    

Classification performance of Random Forest classifier
 


Result of 10-fold cross validation

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
20
Modeling Operational Lane Changing Behavior
Inverse reinforcement learning (IRL) defined by Ng & Russel [2000]
Goal: learn reward function that explains the given behavior.
    

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
21
IRL Results
True vs. Predicted Trajectory Real Reward Recovered Reward
Reward
-1 +1
0

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
22
IRL Results
Indicators:
•Total displacement errors
•Average displacement errors
•Fréchet distance

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
23
Simulation Experiment
Simulation workflow

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
Simulation scenarios:
•Scenario 1: Base scenario
•Scenario 2: Accident
•Scenario 3: High pressure from on-ramp traffic
•Scenario 4: Higher speed limit
Evaluation measures:
•Distribution of Time-To-Collisions (TTC)
•Throughput of the network
•Number of lane changes
24
Simulation Experiment

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior
Operational
Behavior Simulation Conclusions
25
Simulation Results
Scenario 2: Accident
Measure Unit SUMO IRL-based for human
drivers IRL-
based for automated
vehicles
Throughput
Veh
/
hr
2465 2732 (+11%) 2654 (+8%)
Number of lane changes
-1781 1853 (+4%) 1821 (+2%)

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
1. How can the socially acceptable driving behavior be sampled from the dataset of individual
vehicle trajectories?
2. Can the tactical and operational lane changing behavior for automated vehicles be learned
from the observations of human drivers’ lane changing behavior?
3. How can this methodology be generalized for modeling of the automated vehicle’ driving
behavior in other driving situations?
26
Discussion

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
•Nature of data:
- Limited infrastructure variability; and
-Lack of information about strategic driving behavior and drivers’ motivation.
•A strong assumption in IRL method is that the expert demonstrations are locally optimal.
•Sensors were not specifically considered in this modeling study.
•Transferability was tested with microscopic traffic simulation.
27
Limitations

Introduction Methodology Traffic
Observations Clustering Tactical
Behavior Operational
Behavior Simulation Conclusions
•Integrated modeling approach that is able to overcome the deficiencies and combine the
merits of both physical and data-driven modeling approaches.
•Implement the modeling approach in more complex situations such as urban areas.
•Future enhancements for proposed data-driven approach:
- Problem of sub-optimality in the agent’s actions.
- Reward functions that are not necessarily the linear combination of features.
28
Recommendations for Future Study

•Destatistics, 2019, Verkehr:Verkehrsunfälle, Fachserie 8 Reihe 7
https://www.destatis.de/DE/Themen/Gesellschaft-Umwelt/Verkehrsunfaelle/Publikationen/Downloads-
Verkehrsunfaelle/verkehrsunfaelle-jahr-2080700187004.pdf?__blob=publicationFile
•J. A., Michon, 1985, A Critical View of Driver Behavior Models: What Do We Know, What Should We
Do?, Human Behavior and Traffic Safety, Springer US.
•K. Pearson, 1901, LIII. On lines and planes of closest fit to systems of points in space, The London,
Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 2:11, 559-572.
•J. MacQueen, 1967, Some methods for classification and analysis of multivariate observations.
Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Volume 1:
Statistics, 281-297, University of California Press, Berkeley, Calif..
•N. Motamedidehkordi, S. Amini, S. Hoffmann, F. Busch and M. R. Fitriyanti, 2017, Modeling tactical lane-
change behavior for automated vehicles: A supervised machine learning approach, 2017 5th IEEE
International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS),
Naples, 2017, pp. 268-273.
•A.Y. Ng , S. Russell, 2000, Algorithms for Inverse Reinforcement Learning, in Proc. 17th International
Conf. on Machine Learning, 663-670,Morgan Kaufmann.
•R. Bellman,1954, The theory of dynamic programming. Bull. Amer. Math. Soc. 60, no. 6, 503-515.
29
References

Thank you for you attention!
30