01713218

Component

This study implements a video-based traffic queue monitoring system using Mask RCNN: A convolutional neural network (CNN) approach for predicting pixel-level segmentation masks on classified regions of interest. Taking advantage of a large database of annotated video surveillance data and recent advances in machine learning and high-performance computing, the authors train a deep-learning based model that is able to accurately extract traffic queue-related information from infrastructure mounted video cameras. Several experiments are conducted to fine-tune the system’s robustness in different traffic and environmental conditions. Overall, the system achieves 92.8% accuracy in daylight, night, and rainy conditions. Although extremely poor but rare conditions affects the system’s accuracy, it is able to learn and correct for false detections when re-trained with data captured under such conditions. A comparative analysis with YOLO (You Look Only Once), a classical single stage CNN method is also conducted. Although Mask RCNN underperformed YOLO by approximately 3% error margin in all categories, its ability to provide pixel level segmentation makes it superior for extracting traffic queue parameters. The outcome of this study could be seamlessly integrated into traffic system such as smart work zone management systems, signal control systems, etc.

19-00850

English

15p

2019

Transportation Research Board 98th Annual Meeting

Digital/other

Figures; Photos; References; Tables

Accuracy; Image analysis; Machine learning; Neural networks; Traffic queuing; Traffic surveillance

Data and Information Technology; Highways; Operations and Traffic Management

Transportation Research Board Annual Meeting 2019 Paper #19-00850

TRIS, TRB, ATRI

Aug 2 2019 11:07AM

• A Deep Learning Approach for Lane-Based Short-Term Traffic Volume Prediction at Signalized Intersections
• A New Multi-Source Traffic Data Integration Framework for Traffic Analysis and Performance Measurement
• Analysis of the Effects of an HOV Lane on a German Freeway - A Simulation Study with PTV Vissim
• Combining Attitudes and Travel Behavior - A Comparison of Urban Mobility Types Identified in Shanghai, Berlin and San Francisco
• Comparison of Superpave Gyratory Compactor and Impact Hammer for Compaction of Unbound Granular Materials
• Containership Routing and Scheduling Model with Multiple Time Windows, Split Loads and Berth Constraints
• Feasibility Investigation of Upgrading Gravel Road to Otta Seal Surface: an Economic Analysis Approach
• Identifying Usage Profiles of Station-Based Car-Sharing Members Using Cluster Analyses
• Machine Learning and Daily Activity Patterns
• Multiobjective Capacitated Location-Allocation Model for Urban Logistics Delivery Facilities
• Regulating Airborne Shipping Emissions: Current Status, Future Outlook and Insights on Industry Responses
• Simulation of Autonomous Vehicles Based on Wiedemann's Car Following Model in PTV Vissim
• The impact of movement speed on pedestrians’ obstacle avoidance