Are the authors claiming motion helps them, because Figure 8 suggests the static and dynamic detection perform about the same? Also, it would be interesting if they tried it with pedestrian motion and other motion, like cars or bikes, etc. It seems like they didn't do that.

Very interesting. In Figure 1E, there's a very elborate rectangle filter which appears to be designed to catch pedestrians viewed under large perspective distortion. Do you think they used a symmetric version of it going along the other diagonal?

They use an exhaustive number of filters for various possible aspect ratios and positions. But a remedy should be suggested for this extra calculations. For example, knowing the distance of the camera to ground (assuming aerial view), we can restrict the size to and upper bound. This will also help eliminate other moving objects of different sizes, like cars.

Interesting. This idea could easily be extended to more than two frames. One reason you might want to do that is to separate out swaying motions (wind blowing trees around) from pedestrians and cars that follow a more progressive trajectory.

I agree with Stephen's comment. Since the authors claim that the novelty of their method is in the use of motion, the experimental results need to be less ambiguous: one ROC curve concludes that motion helps (fig 9) while the other says motion does not have an effect (fig 8). Thus, as it stands, the paper is inconclusive. It may have helped to determine in more detail the reason(s), if any, for the different results.