Authors: Yue Lin*, The Ohio State University, Ningchuan Xiao, The Ohio State University
Topics: Transportation Geography, Applied Geography, Geographic Information Science and Systems
Keywords: Deep neural network, traffic flow, urban transit
Session Type: Paper
Presentation File: No File Uploaded
The growing number of real-time camera feeds in many urban settings has made it possible for traffic density measurement in order to provide data support for a wide range of applications such as effective traffic control and for the understanding of urban systems. However, reliable traffic density estimation at real-time or with a high temporal resolution has been a challenge due to highly variable environmental (e.g., air and light pollution, fog and storms) and camera (e.g., angles and scopes) conditions. In this study, we propose a deep learning approach to estimate traffic density using real-time camera images in the area of Columbus, Ohio. Our method is developed based on Convolutional Neural Networks (CNNs) that are widely adopted in computer vision. First, state-of-the-art domain adaptive object detection models are fine-tuned and employed to identify the motor vehicles present in images. By incorporating domain adaptive components in model architectures, these models are able to handle the scene variations in different camera images. Then, traffic density is calculated based on the outputs of vehicle detection and the lengths of roadways within camera views, where camera calibration is performed to determine both intrinsic and extrinsic camera parameters so as to assess roadway lengths using world coordinates under the projected coordinate system (PCS). We further employ the proposed approach to time series analysis of real-time traffic density during the Third Street Pop-Up Mobility Experiment, a practice for new traffic modes in Columbus, and to provide suggestions for transportation planning of the city.