Segmentation and object identification in warehouse environments using machine learning
|Title||Segmentation and object identification in warehouse environments using machine learning|
|Summary||Segmentation and object identification in warehouse environments using machine learning|
|Keywords||Vision, lidar, machine learning, segmentation, object identification|
|TimeFrame||Oct 2019 to June 2020, with possible extension to Sep 2020|
|Prerequisites||Programming (any of C++, Python, Matlab)|
|Supervisor||Björn Åstrand, Naveed Muhammad|
Context and overview: This master thesis falls in the context of autonomous mobile robots (such as forklift trucks, industrial cleaning robots etc.) that operate in warehouse or factory kind of environments. An autonomous robot needs to perceive its environment and detect and identify objects and agents around it, in order to achieve any given goal (e.g. transporting an object from point A to B). This project focuses on perceiving robot environment using vision and lidar sensing modalities, and then using vision modality (for instance acquired using a camera or Kinect module) for annotating range data (for instance acquired using a time-of-flight camera or a lidar), which will in turn be used for tasks of scene segmentation and object identification.
Data: Datasets, in the form of ROS bag files, acquired in a warehouse environment (including vision and range data) is available to be used in the project. There might be possibilities for more data collection.
Research questions: The following research questions will be addresses in the project.
(i) Can state-of-the-art machine learning algorithms for object identification and scene segmentation based on vision sensing, that work in outdoor environments, be employed for indoor warehouse-like environments using transfer learning. (ii) Can vision-based-algorithm-generated labels be used for annotation of objects lidar data in warehouse-like settings? (iii) How can machine learning be used for the tasks of segmentation and object identification in range data in warehouse-like environments.