Paper Draft Comments: vs Frames

We want to show the advantages and disadvantages of the event-driven particle filter and the frame-based particle filter. In addition we want a more thorough conclusion to the tracking conference papers.

Previous papers

Hough transform - tracking could be done at a high frequency, but a fixed temporal window wasn't good with high variation in velocity between camera and target. No "negative" input as it assumes events are only on edges which isn't the case if the temporal window is too high.

Particle filter - state in u, v, and t. optimise observation by searching in time. Includes "negative" space. Improved robustness to velocity changes. Only tested on the two small datasets in Hough. The logical extension is to show the particle filter works in more controlled conditions. Different target speeds and different robot movement speeds. Robot tracking the target in 3D coordinates would be nice. We didn't show any results in regards to update rate, latency, or cpu usage.

Difference with frame-based

• likelihood: template method vs. parameterised shape. using a fiducial removes this difference. A discussion point would then have to be: can we use a template with events too? or does the camera fundamentally have less information
• prediction model: constant acceleration vs. constant position. moving the target in a circle means the acceleration is constantly changing. This would give us an advantage as the frame-based prediction doesn't help as much -precision: I hope we are equal/more precision.
• update rate, latency, cpu usage. we can easily measure update rate, now we can measure latency up to millisecond resolution, we can measure CPU usage [the CPU usage I calculate is somewhat/sometimes different to what windows reports...]

Experiments

• stationary camera, moving object
• moving eyes, stationary object
• gaze following the target