FPGA Implementation of Feature Matching in ORB-SLAM2

Sammanfattning: Simultaneous Localization And Mapping (SLAM) is an important component in solving the problem of autonomous navigation — allowing machines such as selfdriving cars and mobile robots to find their way in the world without human instruction. Though there is a steadily growing body of literature in the field of SLAM, far fewer works currently address using hardware acceleration to speed up this computationally heavy task. That is precisely the concern of this thesis project, in which one of the largest bottlenecks in feature based visual SLAM — feature matching — is investigated for hardware acceleration. After comparing several state of the art methods, the Hamming Distance Embedding Binary Search Tree (HBST) is identified as the best candidate for a hardware-based feature matching system; and the specifics of such a system design are presented in detail. As a means of reducing memory requirements by up to 50%, thus enabling a component of the system to reside in on-chip memory, a new way of storing binary trees was invented: the Heterogeneous Binary Tree Array (HBTA). This method enables binary trees with different sizes of data in their internal and leaf nodes to be stored in an array-based layout with significantly less overhead than a traditional approach; thereby enhancing cache performance, prefetching capabilities, and minimizing storage space.