Single-Stage Related Object Detection for Intelligent Industrial Surveillance

Detecting the position and safe wearing of workers is an significant topic in industrial production. However, mainstream detectors aware object instances individually instead of exploring contextual information. In this article, a relation extraction module (REM) is proposed to introduce local and global contexts at the same time. It processes a set of anchors simultaneously through interaction between their appearance feature and location, thus allowing building local context and generating enhanced anchors. It can be plugged into most popular detectors without additional labeling. Experiments on public datasets and onsite surveillance video indicate that REM improves the accuracy of single-stage detectors especially small models while maintains real-time performance. A real-time intelligent surveillance system has already been established and applied in the factory, which makes great significance to the management of safety supervision departments.