Dangerous Action Recognition for Ship Sailing to Limited Resource Environment

Abstract

Action recognition is a comprehensive application prospect in the surveillance video source but most of the well-known models have significant computational cost feature. Take a bottom-up algorithm named OpenPose as an example, it consumes a lot of time to solve a task. If dangerous action monitoring is deployed on overseas ships, which can not afford such a large amount of computation and storage cost. Meanwhile, a oversea ship is mostly offline during ocean voyages and cannot upload and download data. This paper addresses this issue by improving on the PyTorch-OpenPose model, adjusting the resolution of the input images and removing some of the unneeded skeletal key-point information from the model. Experimental results demonstrate that the execution time is reduced by about 71%, while the memory footprint is reduced by about 62%.