ANOMALY DETECTION IN VIDEO SURVEILLANCE SYSTEM USING AUTOENCODER
Anomaly in video can be defined as an unexpected event and it doesn’t conform to a behavior that can be predicted. Anomaly detection is a very challenging problem because of the unbounded nature of an anomaly. This paper proposed an unsupervised learning approach to learn the normal definition...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Project |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/55951 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | Anomaly in video can be defined as an unexpected event and it doesn’t conform to a
behavior that can be predicted. Anomaly detection is a very challenging problem because
of the unbounded nature of an anomaly. This paper proposed an unsupervised learning
approach to learn the normal definition contained in a video and use it to detect anomalies.
The model architecture is based on convolutional autoencoder with double decoder to
predict future frame appearance and optical flow. The unsupervised learning approach
was employed so that the model can be trained end-to-end with minimal tuning. Anomaly
can be detected by calculating reconstruction error of a future frame prediction based on
previous frames. On top of that, the model would also locate the anomalous region within
a frame based on the error patch. UCSD Anomaly Dataset was used for testing purpose
to measure model performance. The model achieved 77.3% AUC on Peds1 and 87.8%
AUC on Peds2.
|
|---|