clustering human behaviors with dynamic time warping and hidden markov models for a video...

67
Introduction Human Behavior Pattern Clustering Experimental Results Conclusion Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System Kan Ouivirach and Matthew N. Dailey Computer Science and Information Management Asian Institute of Technology ECTI-CON May 19-21, 2010 1 / 29 Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Upload: kan-ouivirach

Post on 22-Nov-2014

369 views

Category:

Education


1 download

DESCRIPTION

 

TRANSCRIPT

Page 1: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Clustering Human Behaviors with Dynamic Time Warpingand Hidden Markov Models for a Video Surveillance System

Kan Ouivirach and Matthew N. Dailey

Computer Science and Information ManagementAsian Institute of Technology

ECTI-CONMay 19-21, 2010

1 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 2: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Outline

1 Introduction

2 Human Behavior Pattern Clustering

3 Experimental Results

4 Conclusion

2 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 3: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction

Human behavior understanding is important for intelligent systems.

Difficult due to the wide range of activities possible in any givencontext

Figure: Reprinted from http://www.sourcesecurity.com/

A classic work by Yamato et al. who model tennis actions usinghidden Markov models (HMMs)

3 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 4: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction

Human behavior understanding is important for intelligent systems.

Difficult due to the wide range of activities possible in any givencontext

Figure: Reprinted from http://www.sourcesecurity.com/

A classic work by Yamato et al. who model tennis actions usinghidden Markov models (HMMs)

3 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 5: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction

Human behavior understanding is important for intelligent systems.

Difficult due to the wide range of activities possible in any givencontext

Figure: Reprinted from http://www.sourcesecurity.com/

A classic work by Yamato et al. who model tennis actions usinghidden Markov models (HMMs)

3 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 6: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

To help security personnel work reliably and efficiently,

filter out typical events;automatically present anomalous events to human operator.

Figure: Reprinted from http://sikafutu.com/

4 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 7: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

To help security personnel work reliably and efficiently,

filter out typical events;automatically present anomalous events to human operator.

Figure: Reprinted from http://sikafutu.com/

4 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 8: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

To help security personnel work reliably and efficiently,

filter out typical events;automatically present anomalous events to human operator.

Figure: Reprinted from http://sikafutu.com/

4 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 9: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

One limitation of most of the work

The number of “normal” behavior patterns need to be knownbeforehand.Nair and Clark (2002) use HMMs to model a common, predefinedactivity in a scene.

Unsupervised analysis and clustering of behaviors for a variety ofpurposes has started to draw attentions.

Li et al. (2006) cluster human gestures by constructing an affinitymatrix using dynamic time warping (DTW), and apply thenormalized-cut approach to cluster.

5 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 10: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

One limitation of most of the work

The number of “normal” behavior patterns need to be knownbeforehand.Nair and Clark (2002) use HMMs to model a common, predefinedactivity in a scene.

Unsupervised analysis and clustering of behaviors for a variety ofpurposes has started to draw attentions.

Li et al. (2006) cluster human gestures by constructing an affinitymatrix using dynamic time warping (DTW), and apply thenormalized-cut approach to cluster.

5 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 11: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

One limitation of most of the work

The number of “normal” behavior patterns need to be knownbeforehand.Nair and Clark (2002) use HMMs to model a common, predefinedactivity in a scene.

Unsupervised analysis and clustering of behaviors for a variety ofpurposes has started to draw attentions.

Li et al. (2006) cluster human gestures by constructing an affinitymatrix using dynamic time warping (DTW), and apply thenormalized-cut approach to cluster.

5 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 12: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

One limitation of most of the work

The number of “normal” behavior patterns need to be knownbeforehand.Nair and Clark (2002) use HMMs to model a common, predefinedactivity in a scene.

Unsupervised analysis and clustering of behaviors for a variety ofpurposes has started to draw attentions.

Li et al. (2006) cluster human gestures by constructing an affinitymatrix using dynamic time warping (DTW), and apply thenormalized-cut approach to cluster.

5 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 13: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

One limitation of most of the work

The number of “normal” behavior patterns need to be knownbeforehand.Nair and Clark (2002) use HMMs to model a common, predefinedactivity in a scene.

Unsupervised analysis and clustering of behaviors for a variety ofpurposes has started to draw attentions.

Li et al. (2006) cluster human gestures by constructing an affinitymatrix using dynamic time warping (DTW), and apply thenormalized-cut approach to cluster.

5 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 14: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

Some recent related works using HMMs to cluster behavior patterns

Swears et al. (2008) propose hierarchical HMM-based clustering tofind motion trajectories and velocities in a highway interchangescene.

6 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 15: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

Some recent related works using HMMs to cluster behavior patterns

Swears et al. (2008) propose hierarchical HMM-based clustering tofind motion trajectories and velocities in a highway interchangescene.

6 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 16: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

New method for clustering human behaviors in the context of videosurveillance

Combination of clustering and HMMs to group human behaviors

Summary flow

1 After extracting sequences, we use DTW to measure the pairwisesimilarity between sequences.

2 Construct an agglomerative hierarchical clustering dendrogrambased on the DTW similarity measure.

3 Recursively, find the optimal set of behavior clusters using HMMs.

Oates et al. (2001) first proposed the idea of using the DTW withHMMs to cluster time series.

7 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 17: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

New method for clustering human behaviors in the context of videosurveillance

Combination of clustering and HMMs to group human behaviors

Summary flow

1 After extracting sequences, we use DTW to measure the pairwisesimilarity between sequences.

2 Construct an agglomerative hierarchical clustering dendrogrambased on the DTW similarity measure.

3 Recursively, find the optimal set of behavior clusters using HMMs.

Oates et al. (2001) first proposed the idea of using the DTW withHMMs to cluster time series.

7 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 18: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

New method for clustering human behaviors in the context of videosurveillance

Combination of clustering and HMMs to group human behaviors

Summary flow

1 After extracting sequences, we use DTW to measure the pairwisesimilarity between sequences.

2 Construct an agglomerative hierarchical clustering dendrogrambased on the DTW similarity measure.

3 Recursively, find the optimal set of behavior clusters using HMMs.

Oates et al. (2001) first proposed the idea of using the DTW withHMMs to cluster time series.

7 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 19: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

New method for clustering human behaviors in the context of videosurveillance

Combination of clustering and HMMs to group human behaviors

Summary flow

1 After extracting sequences, we use DTW to measure the pairwisesimilarity between sequences.

2 Construct an agglomerative hierarchical clustering dendrogrambased on the DTW similarity measure.

3 Recursively, find the optimal set of behavior clusters using HMMs.

Oates et al. (2001) first proposed the idea of using the DTW withHMMs to cluster time series.

7 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 20: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

New method for clustering human behaviors in the context of videosurveillance

Combination of clustering and HMMs to group human behaviors

Summary flow

1 After extracting sequences, we use DTW to measure the pairwisesimilarity between sequences.

2 Construct an agglomerative hierarchical clustering dendrogrambased on the DTW similarity measure.

3 Recursively, find the optimal set of behavior clusters using HMMs.

Oates et al. (2001) first proposed the idea of using the DTW withHMMs to cluster time series.

7 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 21: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

New method for clustering human behaviors in the context of videosurveillance

Combination of clustering and HMMs to group human behaviors

Summary flow

1 After extracting sequences, we use DTW to measure the pairwisesimilarity between sequences.

2 Construct an agglomerative hierarchical clustering dendrogrambased on the DTW similarity measure.

3 Recursively, find the optimal set of behavior clusters using HMMs.

Oates et al. (2001) first proposed the idea of using the DTW withHMMs to cluster time series.

7 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 22: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Introduction (cont.)

New method for clustering human behaviors in the context of videosurveillance

Combination of clustering and HMMs to group human behaviors

Summary flow

1 After extracting sequences, we use DTW to measure the pairwisesimilarity between sequences.

2 Construct an agglomerative hierarchical clustering dendrogrambased on the DTW similarity measure.

3 Recursively, find the optimal set of behavior clusters using HMMs.

Oates et al. (2001) first proposed the idea of using the DTW withHMMs to cluster time series.

7 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 23: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Compared to the related works

Potential to improve upon the state of the art in intelligent videosurveillance applications by

Bootstrapping human behavior classification and anomaly detectionmodulesSupporting incremental HMM learning (performing statistical teststo select which cluster should be incrementally updated)

8 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 24: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Compared to the related works

Potential to improve upon the state of the art in intelligent videosurveillance applications by

Bootstrapping human behavior classification and anomaly detectionmodulesSupporting incremental HMM learning (performing statistical teststo select which cluster should be incrementally updated)

8 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 25: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Compared to the related works

Potential to improve upon the state of the art in intelligent videosurveillance applications by

Bootstrapping human behavior classification and anomaly detectionmodulesSupporting incremental HMM learning (performing statistical teststo select which cluster should be incrementally updated)

8 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 26: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Outline

1 Introduction

2 Human Behavior Pattern Clustering

3 Experimental Results

4 Conclusion

9 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 27: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview

We divide the proposed method into 2 phases.

1 Blob extraction

2 Behavior clustering

10 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 28: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Extraction

Video

Foreground

Extraction

Background

Modeling

Single�Blob

Tracking

List�of�blobs

Vector

Quantization

Blob�features

CCTV�camera

Discrete�symbolsequences

Backgroundmodel

Sequence

Aggregation

Observationsymbols

Figure: Block Diagram of Blob Extraction

11 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 29: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Extraction (cont.)

We represent a blob at time t by the feature vector

~ft =[xt yt st rt dxt dyt vt

],

where

(xt , yt) is the centroid of the blob.

st is the size of the blob in pixels.

rt is the aspect ratio of the blob’s bounding box.

(dxt , dyt) is the unit-normalized motion vector for the blobcompared to the previous frame.

vt is the blob’s speed compared to the previous frame.

12 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 30: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Behavior Clustering

Similarity

Measurement

Discrete�symbol

sequences

Agglomerative

Hierarchical�Clustering

Distancematrix

Dendrogram

HMM-based

Hierarchical�Clustering

Set�of�HMMs

Figure: Block Diagram of Behavior Clustering

13 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 31: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Behavior Clustering (cont.)

Is�the�HMMa�sufficient�modelof�the�sequences�

in�c?

Is�C�empty?

c����cluster�at�rootof�dendrogram

Train�a�HMMon�the�sequences

in�c

No

Yes

Yes

C���{��}

Add�the�trained�HMMto�model�list�M

Replace�c�in�Cwith�the�children

of�c�from�DTW�dendrogram

0

0

c����any�clusterin�C

Remove�c�from�C

No

c

Figure: Processing flow of the use of HMM clustering method

14 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 32: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Behavior Clustering (cont.)

How the processing flow works

Root

15 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 33: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Behavior Clustering (cont.)

How the processing flow works

Root

The HMM is sufficient?

15 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 34: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Behavior Clustering (cont.)

How the processing flow works

Root

Not sufficient

15 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 35: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Behavior Clustering (cont.)

How the processing flow works

Root

Child Child

15 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 36: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Behavior Clustering (cont.)

How the processing flow works

Root

Child Child

The HMM is sufficient?

The HMM is sufficient?

15 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 37: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Clustering (cont.)

The HMM is not sufficient to model the sequences

When there are more than N sequences in a cluster whoseper-observation log-likelihood is less than a threshold.

16 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 38: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Clustering (cont.)

The HMM is not sufficient to model the sequences

When there are more than N sequences in a cluster whoseper-observation log-likelihood is less than a threshold.

16 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 39: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Clustering (cont.)

To determine the optimal rejection threshold

Use an approach similar to that of Oates et al. (2001).Generate random sequences from the HMM.Calculate µc and σc of the per-observation log-likelihood over theset of generated sequences.Let a threshold be pc = µc − zσc , where z is experimentally tuned.

17 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 40: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Clustering (cont.)

To determine the optimal rejection threshold

Use an approach similar to that of Oates et al. (2001).Generate random sequences from the HMM.Calculate µc and σc of the per-observation log-likelihood over theset of generated sequences.Let a threshold be pc = µc − zσc , where z is experimentally tuned.

17 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 41: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Clustering (cont.)

To determine the optimal rejection threshold

Use an approach similar to that of Oates et al. (2001).Generate random sequences from the HMM.Calculate µc and σc of the per-observation log-likelihood over theset of generated sequences.Let a threshold be pc = µc − zσc , where z is experimentally tuned.

17 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 42: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Clustering (cont.)

To determine the optimal rejection threshold

Use an approach similar to that of Oates et al. (2001).Generate random sequences from the HMM.Calculate µc and σc of the per-observation log-likelihood over theset of generated sequences.Let a threshold be pc = µc − zσc , where z is experimentally tuned.

17 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 43: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Blob Clustering (cont.)

To determine the optimal rejection threshold

Use an approach similar to that of Oates et al. (2001).Generate random sequences from the HMM.Calculate µc and σc of the per-observation log-likelihood over theset of generated sequences.Let a threshold be pc = µc − zσc , where z is experimentally tuned.

17 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 44: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Outline

1 Introduction

2 Human Behavior Pattern Clustering

3 Experimental Results

4 Conclusion

18 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 45: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview

Recorded videos at a resolution of 320× 240 and 25 fps over 1 week.

Used a motion detection to save disk space.

Obtained videos corresponding to over 500 motion events, butselected the 298 videos containing only a single motion.

19 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 46: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Found that at least 4 common behaviors:

Walking into the building (Walk-in)Walking out of the building (Walk-out)Parking a bicycle (Cycle-in)Riding a bicycle out (Cycle-out)

Other less common activities:

Walking while telephoning, etc. (Other)

20 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 47: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Figure: Example of common human activities in our testbed scene. (a)Walking in. (b) Walking out. (c) Cycling in. (d) Cycling out.

21 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 48: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Our main hypothesis

Using DTW as a pre-process prior to HMM-based clustering shouldimprove the quality of the clusters in term of separating anomalousfrom typical behaviors.

Compared to

Using only HMMsSupervised classification with HMMs

22 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 49: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Our main hypothesis

Using DTW as a pre-process prior to HMM-based clustering shouldimprove the quality of the clusters in term of separating anomalousfrom typical behaviors.

Compared to

Using only HMMsSupervised classification with HMMs

22 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 50: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Our main hypothesis

Using DTW as a pre-process prior to HMM-based clustering shouldimprove the quality of the clusters in term of separating anomalousfrom typical behaviors.

Compared to

Using only HMMsSupervised classification with HMMs

22 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 51: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Our main hypothesis

Using DTW as a pre-process prior to HMM-based clustering shouldimprove the quality of the clusters in term of separating anomalousfrom typical behaviors.

Compared to

Using only HMMsSupervised classification with HMMs

22 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 52: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Performed 3 experiments.

Evaluated the results according to how well the induced categoriesseparate the anomalous sequences (hand-labeled with the category“Other”) from typical sequences (Walk-in, Walk-out, Cycle-in,Cycle-out).

1 Using our proposed method2 Using only HMMs3 Using HMMs with supervised learning

In all 3 experiments, we chose linear HMMs with 4 states based onour previous empirical experience.

23 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 53: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Performed 3 experiments.

Evaluated the results according to how well the induced categoriesseparate the anomalous sequences (hand-labeled with the category“Other”) from typical sequences (Walk-in, Walk-out, Cycle-in,Cycle-out).

1 Using our proposed method2 Using only HMMs3 Using HMMs with supervised learning

In all 3 experiments, we chose linear HMMs with 4 states based onour previous empirical experience.

23 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 54: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Performed 3 experiments.

Evaluated the results according to how well the induced categoriesseparate the anomalous sequences (hand-labeled with the category“Other”) from typical sequences (Walk-in, Walk-out, Cycle-in,Cycle-out).

1 Using our proposed method2 Using only HMMs3 Using HMMs with supervised learning

In all 3 experiments, we chose linear HMMs with 4 states based onour previous empirical experience.

23 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 55: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Performed 3 experiments.

Evaluated the results according to how well the induced categoriesseparate the anomalous sequences (hand-labeled with the category“Other”) from typical sequences (Walk-in, Walk-out, Cycle-in,Cycle-out).

1 Using our proposed method2 Using only HMMs3 Using HMMs with supervised learning

In all 3 experiments, we chose linear HMMs with 4 states based onour previous empirical experience.

23 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 56: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Performed 3 experiments.

Evaluated the results according to how well the induced categoriesseparate the anomalous sequences (hand-labeled with the category“Other”) from typical sequences (Walk-in, Walk-out, Cycle-in,Cycle-out).

1 Using our proposed method2 Using only HMMs3 Using HMMs with supervised learning

In all 3 experiments, we chose linear HMMs with 4 states based onour previous empirical experience.

23 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 57: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Performed 3 experiments.

Evaluated the results according to how well the induced categoriesseparate the anomalous sequences (hand-labeled with the category“Other”) from typical sequences (Walk-in, Walk-out, Cycle-in,Cycle-out).

1 Using our proposed method2 Using only HMMs3 Using HMMs with supervised learning

In all 3 experiments, we chose linear HMMs with 4 states based onour previous empirical experience.

23 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 58: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Overview (cont.)

Our configuration

the number of deviant patterns allowed in a cluster N = 10

z = 2.0 for a threshold pc = µc − zσc

24 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 59: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Results for Experiment I

Clustering results for Experiment I (DTW+HMMs).

Cluster # Walk-in Walk-out Cycle-in Cycle-out Other

1 96 0 18 0 0

2 0 54 0 5 0

3 0 3 0 8 0

4 0 2 0 0 0

5 0 1 0 2 0

...

14 0 0 0 0 4

15 0 0 0 0 4

16 0 0 0 0 2

17 0 0 0 0 2

One-seqclusters 4 17 34 21 4

25 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 60: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Results for Experiment II

Begin by training a single HMM on all sequences.

Assign every sequence with a per-observation log-likelihood above athreshold pc to a cluster.

Repeat the process by training a new HMM on the remainingsequences.

Clustering results for Experiment II (HMMs only).

Cluster # Walk-in Walk-out Cycle-in Cycle-out Other

1 15 77 49 43 16

2 80 0 11 2 0

3 5 0 0 0 0

26 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 61: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Results for Experiment III

Trained 4 HMMs on each of the four typical beahviors.

Maximize the F1 value to determine the best per-observationlog-likelihood threshold for each HMM.

For the best separation between the positive and negative testpatterns

Results for Experiment III (Supervised classification with HMMs).

Anomaly detection rate (%) False alarm rate (%)

50 24.6

27 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 62: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Outline

1 Introduction

2 Human Behavior Pattern Clustering

3 Experimental Results

4 Conclusion

28 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 63: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Conclusion

We have proposed and evaluated a new method for clusteringhuman behaviors.

Our method

provides an initial partitioning of a set of behavior sequences, thenautomatically identifies where to cut off the hierarchical clusteringdendrogram.could be used to bootstrap an anomaly detection module forintelligent video surveillance applications.shows a perfect separation between typical and anomalous behaviorson real-world surveillance data without any information about thelabels.

29 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 64: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Conclusion

We have proposed and evaluated a new method for clusteringhuman behaviors.

Our method

provides an initial partitioning of a set of behavior sequences, thenautomatically identifies where to cut off the hierarchical clusteringdendrogram.could be used to bootstrap an anomaly detection module forintelligent video surveillance applications.shows a perfect separation between typical and anomalous behaviorson real-world surveillance data without any information about thelabels.

29 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 65: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Conclusion

We have proposed and evaluated a new method for clusteringhuman behaviors.

Our method

provides an initial partitioning of a set of behavior sequences, thenautomatically identifies where to cut off the hierarchical clusteringdendrogram.could be used to bootstrap an anomaly detection module forintelligent video surveillance applications.shows a perfect separation between typical and anomalous behaviorson real-world surveillance data without any information about thelabels.

29 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 66: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Conclusion

We have proposed and evaluated a new method for clusteringhuman behaviors.

Our method

provides an initial partitioning of a set of behavior sequences, thenautomatically identifies where to cut off the hierarchical clusteringdendrogram.could be used to bootstrap an anomaly detection module forintelligent video surveillance applications.shows a perfect separation between typical and anomalous behaviorson real-world surveillance data without any information about thelabels.

29 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Page 67: Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System

Introduction Human Behavior Pattern Clustering Experimental Results Conclusion

Conclusion

We have proposed and evaluated a new method for clusteringhuman behaviors.

Our method

provides an initial partitioning of a set of behavior sequences, thenautomatically identifies where to cut off the hierarchical clusteringdendrogram.could be used to bootstrap an anomaly detection module forintelligent video surveillance applications.shows a perfect separation between typical and anomalous behaviorson real-world surveillance data without any information about thelabels.

29 / 29

Clustering Human Behaviors with Dynamic Time Warping and Hidden Markov Models for a Video Surveillance System