Hardware-software Codesing of Feature Tracking Algorithms

Relative motion of camera and environment results in such visual cues, which may characterize 3D motion as well as the 3D structure of the a-priori unknown environment. 3D navigation includes both tasks: motion and 3D structure estimation. Earlier we showed that navigation can be based on the KLT algorithm where feature-type visual cues (e.g. corners) are selected and tracked throughout the video frame sequence. Reliability and real-time feature of 3D reconstruction mainly depends on the specificities of selected algorithms and design metrics of implementation technology. The paper first summarizes the theoretical background of the KLT feature tracking algorithm. Then hardware-software partition alternatives are evaluated in more detail resulting in FPGA implementations optimizing speed and cost. The pipeline structure of the hardware proved to be necessary due of the great amount of data which should be processed continuously as fast as possible. This work has been motivated by various real-time applications described in other papers. [8], [9].