Citation

Miao H, Zhang S, Flannagan CAC. J. Big Data Anal. Transp. 2022; 4(1): 41-55.

Copyright

(Copyright © 2022, Holtzbrinck Springer Nature Publishing Group)

DOI

10.1007/s42421-022-00053-8

PMID

unavailable

Abstract

Naturalistic driving data (NDD) is an important source of information to understand crash causation and human factors and to further develop crash avoidance countermeasures. Videos recorded while driving are often included in such datasets. While there is a large amount of video data in NDD, only a small portion of it can be annotated by human coders and used for research. In this paper, we explored a computer vision method to automatically annotate behaviors in videos. More specifically, we developed a 3D ConvNet algorithm to automatically extract cell phone behaviors from videos. The experiments show that our method can extract chunks from videos, most of which (∼ 85%) contain the automatically labeled cell phone behaviors. Importantly, we discuss and evaluate two use cases: (1) using algorithm labels without subsequent human review, and (2) using algorithm labels with subsequent human review. We find that even a 99% accurate algorithm will produce statistics that are appreciably biased towards the null, relative to ground truth, when labels are used without review. Thus, while the algorithm is not accurate enough to support the direct use of its labels in analysis, in conjunction with a human review of the extracted chunks, this approach can find cell phone behaviors much more efficiently than simply viewing a video.

Language: en

Keywords

3D ConvNets; Cell phone behaviors; Video processing