Arjun Sarup

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2021-188

August 13, 2021

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/EECS-2021-188.pdf

In this thesis, I investigate the potential that optical flow offers for automobile driver de-identification and behavior classification. I use Farneback’s algorithm for optical flow to extract the temporal movements that characterize facial behavior from a recently introduced driver-facing dataset. To enhance the behavioral variety offered within this driver-facing dataset, I introduce several data augmentation procedures that take advantage of publicly available content on media platforms like YouTube, Netflix. One of the most impactful contributions within this augmentation framework is likely our “shot-partitioner-algorithm". This shot-partitioner-algorithm utilizes human bounding boxes and changes in the cosine similarity of 2D body pose to partition videos according to the constituent camera field of views (FOVs) used in filming them. I train a classifier on the augmented data that, in its testing stage, is able to predict driver behavior for a small subset of driver-facing videos taken from the wild with a 70% testing accuracy. Finally, I carry out an experiment with 13 participants in order to qualitatively determine if visualizations of optical flow in place of the actual concerned driver-facing video can sufficiently mask driver identity.

Advisors: Gerald Friedland

---

BibTeX citation:

@mastersthesis{Sarup:EECS-2021-188,
    Author= {Sarup, Arjun},
    Title= {Optical Flow for De-Identification and Driver Behavior Classification},
    School= {EECS Department, University of California, Berkeley},
    Year= {2021},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/EECS-2021-188.html},
    Number= {UCB/EECS-2021-188},
    Abstract= {In this thesis, I investigate the potential that optical flow offers for automobile driver de-identification and behavior classification. I use Farneback’s algorithm for optical flow to extract the temporal movements that characterize facial behavior from a recently introduced driver-facing dataset. To enhance the behavioral variety offered within this driver-facing dataset, I introduce several data augmentation procedures that take advantage of publicly available content on media platforms like YouTube, Netflix. One of the most impactful contributions within this augmentation framework is likely our “shot-partitioner-algorithm". This shot-partitioner-algorithm utilizes human bounding boxes and changes in the cosine similarity of 2D body pose to partition videos according to the constituent camera field of views (FOVs) used in filming them. I train a classifier on the augmented data that, in its testing stage, is able to predict driver behavior for a small subset of driver-facing videos taken from the wild with a 70% testing accuracy. Finally, I carry out an experiment with 13 participants in order to qualitatively determine if visualizations of optical flow in place of the actual concerned driver-facing video can sufficiently mask driver identity.},
}

EndNote citation:

%0 Thesis
%A Sarup, Arjun 
%T Optical Flow for De-Identification and Driver Behavior Classification
%I EECS Department, University of California, Berkeley
%D 2021
%8 August 13
%@ UCB/EECS-2021-188
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2021/EECS-2021-188.html
%F Sarup:EECS-2021-188