FROM CLASSICAL VISION TO DEEP LEARNING: A SURVEY OF COMPUTER VISION METHODS IN ROBOTICS

Abstract

Understanding the scene is one of the fundamental principles of robotics autonomy. Robots need to abstract, locate, and infer objects in the environment they should eventually navigate, handle, or interact with safely. Traditionally, such capabilities have been achieved through highly crafted pipelines based on geometric reasoning and hand-engineered features. Over the past decade, however, deep learning has revolutionized computer vision so that robots can process rich scenes without explicit feature retrieval by training on data. In this survey, we present a literature review of computer vision algorithms used for scene understanding in robotics, from traditional to modern deep learning methods. We introduce a novel taxonomy that differentiates between geometric mapping, object detection, semantic segmentation, and high-level scene understanding; we survey existing methods for each category; and evaluate the accuracy, robustness, and computational requirements of such techniques through extensive experimentation. We also overview key benchmark datasets which have spurred development and emphasize the open issues and avenues for future work.