This master thesis describes a system for detecting
abnormalities in video surveillance by
using motion, size, texture and direction features.
The method is based on an existing
solution, but includes improvements by using
a different optical flow algorithm.
The method is tested on the publicly available
UCSD anomaly detection dataset with good
results within the different categories compared
to other methods. Tests have been concluded
to view the results of motion, size and
texture features independently where the latter
have shown to be ineffective.
Two datasets were created for direction based
abnormalities. The results on these are very
good with an error rate of 0.3% which with
small alterations could be used in surveillance
systems.
Further research should be put on applicating
the methods for video surveillance systems and
improving the size and texture feature.