Applying Object Recognition To Reduce False Alarms Triggered By Motion Detection In Video Surveillance
Emir Husic, Aleksandar Simeunovic · Lund University Publications Student Papers (Lund University) · 2018
Using motion detection in surveillance cameras is one way of detecting actions in environments.However, motion detection alone is incapable of determining the causing source, such as animals, flying objects, or humans.This incapability tends to trigger alarms where, more often then not, a human is not present.In this thesis, we study the effect of adding another evaluation layer before triggering an alarm -an object detection layer identifying humans explicitly.Video alarms triggered by motion detection, and simultaneously containing properties of a human, are weighed higher as they are more likely a real alarm.We present methods, and choices of data used while applying object detection, these manage to filter up to 85% of the false alarms without losing true ones.The benefit of our approach is the reduction of human hours spent on evaluating false alarms.It is also possible to train the detector to specific environments, which increases the accuracy of the detector by using neural networks.