2021

Our world is moving towards the goal of fully autonomous driving at a fast pace.
While the latest automated vehicles (AVs) can handle most real-world scenarios
they encounter, a major bottleneck for turning fully autonomous driving into
reality is the lack of sufficient corner case data for training and testing AVs. Near crash data, as a widely used surrogate data for traffic safety research, can also
serve the purpose of AV testing if properly collected. To this end, this paper
proposes an Internet-of-Things (IoT) system for real-time near-crash data
collection. The system has several cool features. First, it is a low-cost and
standalone system that is backward-compatible with any existing vehicles.
People can fix the system to their dashboards for near-crash data collection and
collision warning without the approval or help of vehicle manufacturers. Second,
we propose a new near-crash detection method that models the target's size
changes and relative motions with the bounding boxes generated by deep learning-
based object detection and tracking. This near-crash detection method
is fast, accurate, and reliable; particularly, it is insensitive to camera parameters,
thereby having an excellent transferability to different dashboard cameras. We
have conducted comprehensive experiments with 100 videos locally processed
at Jetson, as well as real-world tests on cars and buses. Besides collecting
corner cases, it can also serve as a white-box platform for testing innovative
algorithms and evaluating other AV products. The system contributes to the real world testing of AVs and has great potential to be brought into large-scale
deployment.

The use of deep machine learning methods, such as convolutional neural
networks (CNN), has led to rapid advances in automatic understanding of
consumer photographs. Such approaches have the potential to dramatically
reduce the costs associated with large-scale infrastructure assessment. Here we
present two applications of deep machine learning for image-driven
transportation applications.

Application 1: Using panorama images to assess highway safety

Application 2: Fusing overhead images and airborne LiDAR data to predict freeflow
speed

The fast-evolving COVID-19 pandemic has dramatically reshaped urban travel
patterns. Through the pandemic, one term that nearly everyone throughout the
world has quickly become familiar with is “social distancing.” In this research, we
explore the relationship between social distancing and urban mobility during the
pandemic. Understanding the social distancing behavior will allow urban
planners and engineers to better understand the new norm of urban mobility
amid the pandemic, and what it might hold for individual mobility post-pandemic
or in the event of future pandemics.

There are still very limited efforts in obtaining precise information on social
distancing status for those (e.g., pedestrians) exposed in urban environments.
This is largely attributed to numerous burdens in safely implementing effective
field data collection approaches during the time of crisis. This paper aims to
shorten this deployment gap by developing a data-driven machine learning and
AI framework that leverages existing public video data sources and advanced
computer vision techniques to monitor the evolution of social distancing patterns
in urban areas. Specifically, the proposed framework develops a YOLOv3 based
deep-learning approach with a pre-trained convolutional neural network (using
COCO dataset) to mine the massive amount of public video data captured in
urban areas. Our approach for calculating the social distancing between
pedestrians is also novel because it does not require a site-specific reference for
the calibration, which would have made it impossible to tackle very large number
of remote cameras used in this research. Real-time traffic camera data collected
in New York City (NYC) from more than 100 traffic cameras was used as the
case study to demonstrate the feasibility and validity in using the proposed AI-based
approach to analyze pedestrian social distancing patterns.

One of the main innovations of this research is the additional local training used
to improve the accuracy of the pre-trained CNN model that does not work with
the New York City video streams perfectly. It was thus necessary to improve the
model by retraining it by some well-labeled NYC’s local training set. For this
purpose, we have built a GUI for labeling the images collected from traffic
cameras in NYC. Our results show that microscopic pedestrian social distancing
patterns can be quantified using a generalized real-distance approximation
method. The estimated social distances can be compared to social distancing
guidelines to evaluate the policy compliance/effectiveness during a pandemic.
The quantified social distancing status will aid decision-makers in having a
better understanding of the prevailing social contact challenges. It also provides
informative insights into the development of response strategies or plans for
different reopening phases and potential future scenarios.

Emerging connected and autonomous vehicle (CAV) technologies enable
accurate control of vehicle trajectories with real-time optimization algorithms.
Reinforcement-learning-based controllers have been adopted to control CAVs
optimally and dynamically. However, existing attempts on CAV trajectory
optimization with reinforcement learning may not be flexible enough to overcome
CAV technologies' heterogeneity in a mixed traffic environment (e.g., involving
surrounding human-driven vehicles as well). To fill the gap, we develop an
adaptive dynamic control architecture to optimally control the trajectory of an
individual CAV in mixed traffic based on reinforcement learning. The proposed
controller adapts to various input structures and sizes that may come from multi-sources, as opposed to existing learning-based controllers that only suit specific
input data types. The numerical experiments with field data show that the
proposed controller outperforms well-mannered human drivers and an optimized
Adaptive Cruise Control method.

Real-time forecasting of travel demand is essential to enhance the operational
efficiency of the micromobility system. Accurate spatial-temporal demand
prediction for micromobility services can guide the vehicle rebalancing and
charging to offer better services to travelers. In this study, we propose a context-aware recurrent multi-graph convolutional neural network (CA-RMGCNN)
approach to forecast the demand for micromobility services in real-time. We first
use multiple graphs to encode different relationships among the zones. These
relationships include neighborhood effect, accessibility, land-use effect,
socioeconomic effect, and demographic effect. Then we use Long Short-Term
Memory (LSTM) to aggregate observations at different times and use multigraph
convolution to capture different types of spatial dependency. Lastly, we
use the fully connected neural network to generate hourly predictions. The
proposed method shows better prediction performance than the state-of-the-art
baselines based on the e-scooter demand data in Washington, D.C. from June
2019 to December 2019.