Machine learning for industrial safety culture in the developing world

Anonymous

TL;DR: Active and continual learning for adaption to safety compliance

Abstract:

A disproportionate number of deaths and injuries in the workplace are concentrated
in the developing world due to inadequate enforcement capabilities in such areas,
a lack of capital equipment and a paucity of expertise. This requires solutions
that enable automated detection of harmful violations of safety standards and
sending appropriate warnings at real-time. We propose an automated method for
highlighting such violations of safety measures by foregoing the usage of safety
helmets, hard hats and similar safety equipment in the factory setting. A pipeline is
proposed wherein camera feed at real-time is processed to detect the safety critical
objects and their placement using a set of deep learning based architectures tailored
for detection and localization under compute-constrained environments, with the
helmet compliance sub-task treated as an exemplar of our approach.