# Thresholding for Movement Classification

I have an application which creates a Gaussian model of the background based on consecutive frame difference and auto-generates a threshold for separating no-changes to movements. The application works perfectly for artificial lighting. But when daylight is present alongside artificial light, I get something like ths

The red dots corresponds to some scalar quantity under steady lighting conditions, and the green ones correspond to the sudden changes in illumination levels. As a result of the green spikes, selecting a threshold at the red-level would mean a lot of false positives where as selecting a threshold at the green level essentially means that detection probability falls because the small movements are not detected. What I tried was selecting separate thresholds for the green region and the red region but still the detector performance barely improved. Is there any known approach for tackling this problem.

False corresponds to the False Alarms. My model was if $RedMean-maxR\times RedVariance < val < RedMean + maxR\times RedVariance$ then no movement where $maxR = max(\frac{RedElement_{i} -RedMean}{RedVariance})$ and likewise for the green region and anything lying between the red and green would be classified as small movement.