Researchers for this study developed improved video processing algorithms that can identify more complex behavior and actions of individuals (i.e., pedestrians, drivers, bicyclists) in various traffic situations. The research team wanted to build on existing video-processing algorithms that can identify basic behavior, “primitives” (such as when a subject’s hand is near their head), but cannot always identify a “high-level” behavior (such as talking on a cellphone). Specifically, the researchers aimed to:

• Use a bottom-up machine-learning approach to train an algorithm to recognize high-level behaviors using object detection, human-pose estimation, and behavior classification.
• Use a top-down approach to catalog primitive and high-level behaviors and develop a statistical prediction model to link them.

Recommended citation: Federal Highway Administration, Exploratory Advanced Research (EAR) Program Compendium of Papers from Funded Research Projects (Washington, DC: 2023) https://doi.org/10.21949/1521965.