Driving Behavior Group
Driving behavior group is studying data-oriented modeling approaches to the human driving behavior based on signal processing and machine learning of multi-modal sequential data. The topics of this sub group so far covered; driver recognition, driving behavior prediction, risky driving detection and so on. We are also actively building unique signal corpora such as NU-Drive and 1000-lane change.
Detecting Risky Lane Changes Using Ego and Surrounding Vehicle Information Integrated by a Multi-modal Variational Autoencoder
This project focuses on detecting risky lane changes by integrating ego and surrounding vehicle information using a multi-modal variational autoencoder (MVAE). The relationship between ego and surrounding vehicle behavior during safe (or normal) driving is represented using the MVAE, under the assumption that the reconstruction error, i.e., the difference between the input and output signals, will be relatively larger for risky (or unusual) driving situations compared to those for safe (or normal) driving situations. We conduct a risky lane change detection experiment using expressway driving data, which includes 432 lane changes to the right executed by ten drivers during passing maneuvers. The ground truth for the level of risk for each lane change is based on the subjective evaluations of ten human evaluators.
We plan to explore the use of a recurrent structure for the autoencoders, which would allow the use of temporal context within a time series of driving data for behavior detection.
Personalized Safety-focused Control by Minimizing Subjective Risk
Objective driving risk assessment has a long history and mature methodology background, however, research field of subjective risk assessment, causing reason understanding and risk factor identification are still remaining blank. We built subjective risk model (1) to analyze individual differences of subjective risk causing factors from ego vehicle driving signals and surrounding vehicle signals, and (2) to build a subjective risk assessment model to detect risky or non-risky lane change scenes.
In addition, we work on a data-driven control framework for autonomous driving which combines learning-based risk assessment with personalized, safety-focused, predictive control. Different control strategies are employed depending on driving situation (risky vs. non-risky).
- A data-driven framework to model subjective risk for individuals during lane change scenarios.
- Driving features including ego vehicle driving signals and surrounding vehicle directions, to understand individual differences on perceiving driving risk.
- Exploring effective factors that affect driving risk can quantify the safety of individual driving behavior and use it as an indicator of accident risk to achieve effective risk prediction for individuals.
- SMPC: Safety-focused Model Predictive Control for constraining distances to surrounding vehicles
- PMPC: Personalized Model Predictive Control for tracking personalized velocity
Driving Feature Extraction to Reproduce Driving Styles for Autonomous Vehicles
We work towards creating an AI driving agent by extracting behaviors from human expert drivers’ data. Deep learning techniques were used to extract latent features to create velocity profiles so that an autonomous driving agent could drive in a human-like manner. Using a method to compare trajectories, it was shown that the agent was able to use the extracted proactive behaviors to drive similarly to a skilled human driver.
- A data-driven approach with deep auto-encoders to extract driving features and cluster driving behaviors. The behaviors are then used in a driving agent which can drive similarly to a human driver.
- We clustered these latent features into behaviors and created velocity profiles. This allowed us to create an autonomous agent to correctly select the proper behavior to use depending on the environment it was in.
Future works stemming from this work will focus on dynamic obstacles. We would like to use our techniques to extract behaviors and create driving agents that are able to navigate while avoiding dynamic obstacles such as traffic and pedestrians.
The end-goal of this technique would be to generate required driving behaviors for an autonomous agent simply by giving it a path and a vector map. This can likely be done with extensive data to train more types of behaviors and a recurrent neural network (RNN) such as a long term-short term memory (LSTM) network to identify the required behaviors along the path.