Citation

Dhatbale R, Chilukuri BR. J. Big Data Anal. Transp. 2021; 3(2): 141-157.

Copyright

(Copyright © 2021, Holtzbrinck Springer Nature Publishing Group)

DOI

10.1007/s42421-021-00042-3

PMID

unavailable

Abstract

Vehicle trajectories provide very useful empirical data for studying traffic phenomena such as vehicle following behavior, lane changing behavior, traffic oscillations, capacity drop, safety analysis, etc. However, there are a very limited number of studies on extracting trajectory data from mixed traffic and for congested conditions. This paper presents a deep learning-based framework to extract vehicle trajectories in mixed traffic under both free-flow and congested conditions. The popular YOLOv3 deep learning architecture is used and trained on a hybrid dataset generated from two different sets of frames with different scales and orientations. The anchor boxes for vehicle detection and classification are customized to improve accuracy and efficiency. The SORT algorithm is used to track the identified vehicles and the extracted trajectory data are benchmarked with a popular trajectory extraction portal that showed that the proposed model performs well for trajectory extraction. The paper also presents a methodology based on numerical integration techniques to impute missing trajectory data. Finally, the trajectory data obtained from the adjacent road sections are aligned and scaled to the real-world coordinates using coordination transformation and error correction methods to make it useful for research purposes. The extracted trajectories show remarkable accuracy with approximately 0.25-0.35 m of precision. It is expected that these trajectories capture traffic and driving behavior phenomena for a better understanding of mixed traffic conditions.

Language: en

Keywords

Deep learning; Image processing; Mixed traffic conditions; Vehicle trajectory extraction; YOLOv3