Charles Explorer logo
🇬🇧

Smart extensions to regular cameras in the industrial environment

Publication at Faculty of Mathematics and Physics |
2022

Abstract

Data mining from unstructured data can be skillfully employed to improve the performance of manufacturing or industrial processes. The main goal of this paper is to create a fast emergency aid system for object detection in SME industrial premises.

The basic assumption is that SMEs do not have any IT-trained personnel, and the solution has to be unsupervised edge computing. We use several off-the-shelf models of deep neural networks pre-trained for smart city applications, ready for online object recognition and edge computing.

Our system works without any retraining, additional annotation, or human intervention. Specifically, we present heuristics for the automated creation of PGT (Pseudo-Ground Truth).

Based on PGT, we can automatically decide which model is the best in the specific environment. We present an application of fully automated enhancing image capture camera outputs to smarter ones.

We evaluate our system in a controlled experiment. Low-resolution cameras and large areas cause problems for our method.

We present a proof-of-concept for improving our system even in these challenging situations. The benefit is a knowledge extraction in a simple and inexpensive way to expand the organizations' databases with information from unstructured data from CCTV/IP cameras. (C) 2022 The Authors.

Published by Elsevier B.V.