Monocular 3D Object Detection And Tracking in Industrial Settings

Abstract

3D Object detection and tracking are computer vision methods used in many applications. It is necessary for autonomous vehicles and robots to be able to reliably extract 3D localization information about objects in their environment to operate safely. Currently most 3D object detection and tracking algorithms use high quality LiDAR-sensors which are very expensive. This is why research into methods that use cheap monocular camera images as inputs is an active field in computer vision research.

Most current research into monocular 3D object detection and tracking is focused in autonomous driving. This thesis investigates how well current monocular methods are suited for use in industrial settings where the environment and especially the camera perspective can be very different compared to what it is in an automobile. This thesis introduces some of the most used 3D object detection and tracking methods and techniques and tests one detection method on a dataset where the environment and the point of view differs from what it would be in autonomous driving.

This thesis also analyzes the technical requirements for a detection and tracking system that could be be used for autonomous robots in an industrial setting and what future research would be necessary to develop such a system.