The Effects of Downsampling on Computation Speed and Accuracy in Event-based Hough Transform

Abstract: Event cameras can register many million events per second, corresponding to a temporal resolution of thousands of frames per second in a traditional camera. Consequentially, it is often necessary to reduce the number of events in the event stream to achieve fast computation times. In this thesis, three methods of downsampling (temporal downsampling, spatial downsampling, and "event cooldown") were evaluated in terms of their impact on computation speed and accuracy when performing a event-based circle Hough Transform on the event stream. This algorithm was used to track a chaotic pendulum with different downsampling factors applied, and the computation time and accuracy were measured. The results showed that despite downsampling the event streams heavily and reducing the computation time to a few percent of the initial time, little to no decrease in accuracy was observed. Further, a slight accuracy increase was observed initially spatial downsampling. We conclude that both temporal and spatial downsampling can be used to lower computation times without affecting accuracy. However, more research is needed to determine the effects of event cooldown on computation time and accuracy of the Hough Transform.