Mixed 2D-3D Object Recognition for Autonomous Inspection in Industrial Environments

The outstanding results of Machine Learning models and Artificial Intelligence in general have promoted the development of industrial solutions based on autonomous agents equipped with advanced sensing capabilities. This turns out to be particularly important in the case of automated inspection in industrial environments, where the risks faced by human inspectors can be strongly reduced by the use of fully or partially autonomous agents. Advances in sensing technologies (e.g., RGB and thermal cameras, LiDAR) and robotics have enabled the design of agents that combine mobility, perception, and computation, using methods ranging from classical algorithms up to Machine Learning techniques. In this paper, we review a number of methods for automating safety inspection using autonomous agents, focusing on two common challenges that emerge when employing different types of sensors: (i) the problem of 2D-to-3D environment mapping, which is typical when robotic agents are equipped with 2D sensors (RGB cameras, thermal cameras, etc.), and (ii) 3D object detection/recognition tasks from mixed point-clouds composed of LiDAR data and 2D-to-3D mapped data. We critically evaluate the most suitable methods for both problems, highlighting their key properties, benefits, and open challenges. While much of the existing literature has addressed these problems in the context of autonomous driving or generic indoor/outdoor environments, our work aims to extend this knowledge to industrial scenarios. Based on this analysis, we present a practical and computationally efficient pipeline that approaches both (i) and (ii), integrating well-established and state-of-the-art algorithms optimized for real-world deployment in industrial safety inspection tasks. This pipeline is under evaluation in an industrial plant explored by robotic agents.