This sessions presents various real-time applications of machine learning models and tools. Applications include short-term forecasting of traffic speeds, prediction of traffic status in spatiotemporal dimension, estimation of vehicle energy, and assessment of pavement condition.
Deep Learning and UAV Based Solution to Real Time Pavement Condition Assessment
MURAD AL QURISHEE, Tennessee Department of TransportationShow Abstract
WEIDONG WU, University of Tennessee, Chattanooga
Babatunde Atolagbe, Maryland State Highway Administration
Joseph Owino, UTC
Ignatius Fomunung, University of Tennessee, Chattanooga
Mbakisya Onyango, University of Tennessee, Chattanooga
We propose a real-time and low-cost solution to autonomous condition assessment of pavement using deep learning, Unmanned Aerial Vehicle (UAV) and Raspberry Pi microchip computer, which makes roads maintenance and renovation management more efficient and cost effective. We conducted experiments to compare the performance of various combinations of meta-architectures Faster R-CNN, R-FCN, SSD, YOLO with feature extractors Inception v2, NasNet, ResNet101, and MobileNet v1 for pavement distress classification. A low-cost raspberry Pi smart defect detector was configured using the trained SSD with MobileNet v1 deep neural network, which can be deployed with UAV for real-time and remote pavement condition assessment. Our preliminary results show that the detector camera achieves an accuracy of 60% at 1.2 frames per second at its early development stage. The performance of the smart detector will be improved with more powerful Raspberry Pi motherboard available.
A Real-time Spatiotemporal Prediction and Imputation of Traffic Status Based on LSTM and Graph Laplacian Regularized Matrix Factorization
Jin-Ming Yang (firstname.lastname@example.org), Shanghai Jiao Tong UniversityShow Abstract
Zhong-Ren Peng, University of Florida
Lei Lin, University of Rochester
Accurate prediction of traffic status in real time is critical for advanced traffic management and travel navigation guidance. There are many attempts to predict short-term traffic flows using various deep learning algorithms. Most existing prediction models are only tested on spatiotemporal data assuming no missing data entries. However, this ideal situation rarely exists in real world due to sensor or network transmission failure. Missing data is an unnegligible problem. Previous studies either remove time series with missing entries or impute missing data before building prediction models. The former may cause insufficient data for model training, while the latter adds extra computational burden and the imputation accuracy has direct impacts on the prediction performance. In this study, we propose an online framework that can make spatiotemporal predictions based on raw incomplete data and impute possible missing values at the same time. We design a novel spatial and temporal regularized matrix factorization model, namely LSTM-GL-ReMF, as the key component of the framework. The Long Short-term Memory (LSTM) model is chosen as the temporal regularizer to capture temporal dependency in time series data and the Graph Laplacian (GL) serves as the spatial regularizer to utilize spatial correlations among network sensors to enhance prediction and imputation performance. The proposed framework integrating with the LSTM-GL-ReMF model are tested and compared with other state-of-the-art matrix factorization models and deep learning models on two spatiotemporal datasets: traffic data and urban air pollution data. The experimental results show our approach has a robust and accurate performance in terms of prediction and imputation accuracy under various data missing scenarios.
A Meta-Learner Ensemble Framework for Real-Time Short-Term Traffic Speed Forecasting
Divyakant Tahlyan, Northwestern UniversityShow Abstract
Eunhye Kim, Northwestern University
Hani Mahmassani, Northwestern University
This study presents a novel ensemble learning approach called stacking for real-time short-term traffic state prediction. The approach consists of a level-1 meta-learner that combines predictions from several different level-0 models by making use of the past performance information of the level-0 models. The meta-learner consists of least square estimation procedure with non-negatively constraints on past predictions and observed traffic state information. The parameters obtained from the meta-learner are then used to assign weights to the predictions made by level-0 models to obtain the final prediction. The proposed approach is general enough and could be applied in a variety of different situations where longitudinal data feeds might be available, ranging from the demand for certain products and services to freight container demand. We apply the proposed approach to one-day traffic speed data from 13 different loop detectors located on interstate 435 in Kansas City in the state of Missouri. The loop detectors provide space mean speed information at 1-minute interval and we used this data to predict traffic space mean speed 15 minutes and 60 minutes in advance. Our results indicate that the proposed approach outperforms all level-0 models and provides significantly better results in both breakdown and non-breakdown states of traffic flow.
Real-Time Highly Resolved Spatial-Temporal Vehicle Energy Estimation Using Machine Learning and Probe Data
Joseph Severino, National Renewable Energy Laboratory (NREL)Show Abstract
Yi Hou, National Renewable Energy Laboratory (NREL)
Ambarish Nag, National Renewable Energy Laboratory (NREL)
Jacob Holden, National Renewable Energy Laboratory (NREL)
Lei Zhu, University of North Carolina, Charlotte
Juliette Ugirurmurera, National Renewable Energy Laboratory (NREL)
Stanley Young, National Renewable Energy Laboratory (NREL)
Wesley Jones, National Renewable Energy Laboratory (NREL)
Jibonananda Sanyal, Oak Ridge National Laboratory
Real-time highly resolved spatial-temporal vehicle energy consumption is a key missing dimension in transportation data. Most roadway link-level vehicle energy consumption data are estimated using average annual daily traffic (AADT) measures derived from the Highway Performance Monitoring System (HPMS). However, this method does not reflect day-to-day energy consumption fluctuations. As transportation planners and operators are becoming more environmentally attentive, they need accurate real-time link-level vehicle energy consumption data to assess energy and emissions, incentivize energy-efficient routing, and estimate energy impact due to congestion, major events, and severe weather. This paper presents a computational workflow to automate the estimation of time-resolved vehicle energy consumption for each link in a road network of interest, utilizing vehicle probe speed and count data in conjunction with machine learning methods in real-time. The real-time pipeline is capable of delivering energy estimates within a couple seconds upon query to its interface. The proposed method was evaluated on the transportation network of the metropolitan area of Chattanooga, TN. The model results were validated with ground truth traffic volume data collected in the field and from AADT. Energy consumption was estimated and compared for three scenarios, including a COVID-19 period, free flow condition, and peak hour, to demonstrate the effectiveness of the proposed method, estimate energy reduction due to mitigation policies to slow COVID-19 spread, and measure energy loss due to congestion.
DISCLAIMER: All information shared in the TRB Annual Meeting Online Program is subject to change without notice. Changes, if necessary, will be updated in the Online Program and this page is the final authority on schedule information.