Authors:
Thomas Müller
;
Pujan Ziaie
and
Alois Knoll
Affiliation:
Robotics and Embedded Systems Group, Technische Universtät München, Germany
Keyword(s):
Robot Vision, Multithreaded Realtime System, Asynchronous Data Management, Interpretation-Based Preselection, Optimal-Backoff Scheduling.
Related
Ontology
Subjects/Areas/Topics:
Image Processing
;
Informatics in Control, Automation and Robotics
;
Robotics and Automation
;
Vision, Recognition and Reconstruction
Abstract:
As multicore PCs begin to get the standard, it becomes increasingly important to utilize these resources. Thus we present a multithreaded realtime vision system, which distributes tasks to given resources on a single off-the-shelf multicore PC, applying an optimal-backoff scheduling strategy. Making use of an asynchronous data management mechanism, the system also shows non-blocking and wait-free behaviour, while data access itself is randomized, but weighted. Furthermore, we introduce the top-down concept of Interpretation-Based Preselection in order to enhance data retrieval and a tracking based data storage optimization.
On the performance side we prove that functional decomposition and discrete data partitioning result in an almost linear speed-up due to excellent load balancing with concurrent function- and data-domain parallelization.