www.monash.edu.au

Abnormal Behaviour Recognition

Mahfuzul Haque

www.monash.edu.au

2

Literature

• Supervised Approach

• Unsupervised Approach

• Semi-supervised Approach

www.monash.edu.au

3

Supervised Approach

• Based on the assumption that there exist well-defined and known a priori behaviour classes (both normal and abnormal)

• However, in reality, abnormal behaviour is both rare and far from being well defined, resulting in insufficient clearly labelled data required for supervised model building.

www.monash.edu.au

4

Unsupervised Approach

• Further categorized into two different

types:

– With explicit behaviour model

– With no behaviour explicit model

www.monash.edu.au

5

Unsupervised – No Explicit Model

• Either clustering on observed patterns and labelling those forming small clusters as being abnormal

• Or building a database of spatiotemporal patches using only regular / normal behaviour and detecting those patterns that cannot be composed from the database as being abnormal.

• Applicability on previously unseen (normal) behaviour?

www.monash.edu.au

6

Semi-supervised Approach

• Two-stage training process

• Stage one: normal behaviour model is learned using labelled normal patterns.

• Stage two: an abnormal behaviour model is then learned unsupervised using Bayesian adaptation.

• Still suffers from the laborious and inconsistent manual data labelling process

www.monash.edu.au

7

Abnormal Behaviour Detection in Literature

Abnormal Behaviour Detection

Supervised Unsupervised Semi-supervised

No Behaviour Model Explicit Behaviour Model

Clustering of observed patterns,

Database of spatiotemporal patches

Normal behaviour model using

manual labelling,

Abnormal behaviour model

unsupervised using Bayesian

adaptation

Manual Labelling,

Prior assumption of well

define behaviour classes

More Recent Approach

Mixture of Dynamic Bayesian Networks (DBNs)