Resource Optimized Stereo Matching in Reconfigurable Hardware for Autonomous Systems

There is a need for compact, high-speed, and low-power vision systems for enabling real-timemobile autonomous applications. One approach to achieve this is to implement the low- to intermediate-level applications in hardware. Reconfigurable hardware have all these qualities without the limitation of fixed functionality that accompanies application-specific circuits. Resource constraints in reconfigurable hardware calls for resource optimized implementations with maintained performance. The research group in Robotics at Mälardalens University is moving toward the completion of a reconfigurable hardware-platform for stereo vision, coupled with a compact embedded computer. This system will incorporate hardware-based preprocessing components enabling visual perception for autonomous machines. This thesis covers the reconfigurable hardware section of the vision system concerning the realization of scene depth extraction. It shows the advantages of image preprocessing in hardware and propose a resource optimized approach to stereo matching. The work quantifies the impact of reduced resource utilization and a desire for increased accuracy in disparity estimation. The implemented stereo matching approach performs on par with recent similar implementations in terms of accuracy, but excels in terms of resource utilization and resource sharing, as the external memory requirement is removed for larger images. Future work aims to further include processes for navigation, and structure and object recognition. Furthermore, the system will be adapted to real world scenarios, both indoors and outdoors.