Distracted driving detection: video analytics pipeline with adaptive driver feature extraction

Deng, Yuzhong

Permalink

https://hdl.handle.net/2142/115949 Copy

Description

Title

Distracted driving detection: video analytics pipeline with adaptive driver feature extraction

Author(s)

Deng, Yuzhong

Issue Date

2022-07-18

Director of Research (if dissertation) or Advisor (if thesis)

Nahrstedt, Klara

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• distracted driving detection
• video analytics pipeline
• driver feature extraction

Language

eng

Abstract

Distracted driving has long been regarded as the main culprit of road accidents. Every two hours in the United States, a person’s life is taken away by car crashes because of a distracted driver. Given the severity of distracted driving, there has been an increasing academic interest in finding a suitable method to detect and prevent distracted driving behaviors in real-time before tragedy hits. In this thesis, we provide a thorough review of the existing literature on distracted driving detection. Among the existing methods, there are two popular approaches: (1) use of specialized sensors to collect and monitor relevant driver information that may indicate a distracted driver; and (2) use of a smart camera to capture driver images and an on-camera video analytics pipeline to detect distracting behaviors in these driver images. Inspired by these two approaches, we present a novel neural network architecture that can be used as a video analytics pipeline to detect distracting driving behaviors in real- time on the smart camera. Our model architecture incorporates a primary CNN neural network with a selected set of driver features (e.g., driver head pose, eye movement, and hand position). A novel contrastive training mechanism is introduced for the model to learn a compact representation of these external driver features. We have found that the learned driver embedding could then be used in downstream tasks such as distracted driving classification and is shown to be more effective compared to existing neural network methods. Most notably, our model is shown to be more robust against unseen drivers and distracted driving behaviors during test time, which makes our model a suitable candidate for real world application. In summary, our thesis contributes to the current knowledge of distracted driving detection by providing an in-depth evaluation of the existing solutions. We also contribute to the existing study by identifying areas of improvements for the existing video-based analytics pipelines and presenting a new model architecture that could serve as an incremental improvement to the existing pipelines.

Graduation Semester

2022-08

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/115949

Copyright and License Information

Copyright 2022 Yuzhong Deng

Owning Collections