Boosting Clusters of Samples for Sequence Matching in Camera Networks Valtteri Takala,Yinghao Cai and Matti Pietikäinen Machine Vision Group University of Oulu {vallu, yinghao, mkp}@ee.oulu.fi

1. Introduction •

•

3. Boosted Learning and Voting

A classification algorithm for learning and matching image sequences in view-independent object tracking and rerecognition.

•

Gentle AdaBoost based offline learning (one vs. others).

•

Classification and regression trees as weak learners.

Adaptive boosting (AdaBoost) and classification trees for learning.

•

Voting between clusters of samples.

•

Algorithm:

•

A wide collection of features (shape, pose, color, texture, etc.) form an object model.

•

Temporal dimension: k-mean clusters of sequence samples and temporal quality of features.

•

The experiments show that with a proper boosting approach it is feasible to do view-independent sequence matching in sparse camera networks.

2. Object Model •

4. Experiments

Image descriptors:

Feature Contour area Width-height ratio: latest and avg. Object ellipse axis angle Temporal difference: min., latest and avg. Bounding box: x, y (upper left corner), width, height Ellipse bounding box: x, y (center), width, height Color histogram 1 & 2 (for the upper and lower half of the bounding box) Color correlogram 1 & 2 (upper and lower) LBP 1 & 2 (upper and lower) color 1 & 2color (upper and lower) •Opponent Opponent presentation and w invariant: W invariant 1 & 2 (upper and lower)

•

Opponent color space and w invariant: R G     2   O1      R  G  2B   O2     6 O     3 R  G  B     3  

•

Size 1 2 1 3 4 4 432 256 512 1024 128

•

Dataset: 53 sample videos, five camera views, 16 different objects (persons ) appearing in 1-5 views.

•

Object descriptions were extracted using a background subtraction based object tracker.

Fig. 3. Tracking sequences. There might be great variance between the beginning and the end of a sequence as is shown in the pictures. Fig. 2. Object classes.

•

Results:

 O1 O2   O ,O   3   3

Temporal dimension through Gaussian weighting in histogram descriptors and k-mean clustering of all features.

Fig. 4. Different sampling and voting schemes.

Fig. 5. Features selected by the algorithm. Color histogram (28-459), correlogram (460715), LBP (716-1227), opponent color (12282251), w invariant (2252-2379)

5. Conclusions and Future Work

Fig. 1. Image clusters. The samples are often clustered into three distinctive visual clusters in a tracking sequence.

•

This paper presents a model-based approach for sequence matching in a camera network.

•

The results show the advantage of boosting clusters of samples.

•

Future: evaluation in a real-time system.

The 20th International Conference on Pattern Recognition, 23-26 August 2010 Istanbul Turkey