Camera-Based Safety System for Collaborative Assembly

Abstract

Collaborative assembly represents one of the most prevalent practical applications of collaborative robots in intelligent manufacturing. Developing intelligent systems to ensure safety of collaborative assembly processes requires a special attention. In this work, we introduce a visual safety system designed to monitor hazardous situations that may occur during collaborative assembly, potentially resulting in operator injuries. Unlike many other vision-based systems, we solely rely on data from two RGB cameras, without acquiring additional depth information from other sensors. These cameras provide top and side projections of a collaborative workspace. The safety system assesses a current level of a risk by employing two neural network YOLOv8-cls models. These models are pretrained on the ImageNet dataset and subsequently fine-tuned on our dataset. Upon identifying a potential hazard, the system employs our proposed algorithm to determine whether to slow down or halt a robot’s motion. Additionally, the system integrates with a visual control system that utilizes an operator gesture control throughout an assembly process. We further conduct experiments to compare our system’s assessment with an assessment of human experts. An analysis of the experiments demonstrated a high level of correlation between the evaluations of the autonomous system and the human experts. Benefits of the proposed system encompass its relative cost-effectiveness and ease of setup.