Hardware Acceleration for Pedestrian Dead Reckoning in Embedded Systems
Type of DegreeMaster's Thesis
Electrical and Computer Engineering
MetadataShow full item record
Hardware acceleration within embedded systems can potentially allow algorithms to meet real time requirements in devices where it was previously impossible. Though many algorithms have been developed targeting embedded systems, their ability to meet the target environment's real time demands is often unclear. The primary focus of this thesis is the effects of GPU acceleration on a pedestrian dead reckoning system (PDR) targeting processing environments representative of the current wearable market. A software PDR system was developed from components used popularly within research. The system was validated by evaluating accuracy of results from each component against ground truth data. The validated system was simulated on several processors to generate a set of baseline execution times relative to the host processor. The most computationally intense component was selected based on profiling results and accelerated using a GPU. The computational speedup of the system was then used to determine expected execution time relative to each processor baseline execution time. This research was the first to apply hardware acceleration techniques to an embedded PDR system. The results showed a 86\% speedup of the accelerated system's CPU execution but no decrease in overall system computation time. Additionally, the benefits of this change were shown to allow certain particularly low performance processors to meet real time requirements. This shows that GPU acceleration can be applied to accelerate embedded algorithms to allow for smaller and cheaper processing systems to be used compared to those used previously.