Deep learning and anthropometric plane based workflow monitoring by detecting and tracking workers

The worker productivity, a critical variable in project management, significantly affects the progress of a project. The key to measuring productivity is analysis of activities, which provides necessary information by identifying how workers spend their time at certain areas in the site. In this work, we propose a novel joint image-trajectory space for automatic detection and tracking of workers using a single fixed camera. A two-branch convolutional neural network detects workers and their body joints. Instead of tracking the body joints in the image space, we transform detected joints onto virtual parallel planes called “Anthropometric Planes”. The detected joints are, then, tracked using a Kalman Filter on these planes which are created based on anthropometric measures of an average American male. Finally, an uncertainty measure is introduced to reduce the number of identity changes and to handle missing joints. The experiments conducted on an image sequence captured in a nuclear plant shows promising detection and tracking results.