Distracted driving behavior is a major concern and claims the lives of many every year. To address the issue, AI City has presented the Track 3 Challenge, which aims to encourage research and development towards solutions that recognize and localize when distracted driving behavior occurs in time. We present a novel online, multi-view architecture that aggregates temporal context and uniquely samples past and present events to label and pinpoint the start and end temporal boundaries of distracted driving actions. Our algorithm is a two-stage Temporal Action Localization (TAL) method, which does not require a boundary detection network or any localization training. It consists of several stages to predict and refine temporal boundaries: aggregation, prediction, consolidation, post-localization processing, and assessment. Furthermore, our method achieves top results in the AI City 2023 Track 3 Challenge and performs highly in run-time efficiency. Our code is available at  https://github.com/CarrotPeeler/WPI-Naturalistic-Driving-Action-Recognition-MQP.

• This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.

Creator

• Chan, Jared

Publisher

• Worcester Polytechnic Institute

Identifier

• E-project-082423-173508
• 113164

Keyword

• Video Understanding
• Computer Vision
• Temporal Action Localization
• Online
• AI City Challenge
• Naturalistic Driving Action Recognition

Advisor

• Zhang, Ziming
• Gujar, Sanket R.

Year

• 2023

Date created

• 2023-08-24

Resource type

• Major Qualifying Project

Major

• Robotics Engineering

Source

• E-project-082423-173508

Rights statement

• In Copyright

Last modified

• 2023-09-15