Variance Reduction Techniques in

Particle-based Visual Contour Tracking

This web page contains additional material of the paper

Variance reduction techniques in particle-based visual contour tracking.
Daniel Ponsa and Antonio López.
ADAS Group, Computer Vision Center, Universitat Autònoma de Barcelona, Spain

List of Acronyms

Acronym	Definition
PF	Particle Filter. The amount of particles used in the experiments is 1000.
PS	Partitioned Sampling. The amount of particles used in the experiments is 333.
RBPF	Rao-Blackwellised Particle Filter. The amount of particles used in the experiments is 1000.
SNR	Signal-to-Noise ratio
UKF	Unscented Kalman Filter
UPF	Unscented Particle Filter. The amount of particles used in the experiments is 500.

Hand Tracking Experiments

A contour model of a hand with a pointing finger has been constructed from the principal component analysis of the hand silhouettes extracted from the following video sequence.

Training Sequence	Comments
	The segmentation of the hand in this sequence resulted in 468 training shapes, each one modelled by a cubic spline with 17 control points.

Experiments on synthetic sequences

The following videos exemplify the noisy situations considered in the quantitative performance analysis carried out in the paper.

Noise Free Sequence	Uncorrelated Noise	Correlated Noise

The following videos show an example of the performance of the algorithms compared in the paper, for each noise situation considered.

Uncorrelated Noise
SNRIN 17dB	SNRIN 10dB	SNRIN 7dB

Correlated Noise
SNRIN 17dB	SNRIN 13dB	SNRIN 10dB

Note that results compiled in the paper have been obtained by processing, for each one of these noisy situations, 100 randomly distorted sequences.

The variance of the ouput of each algorithm, for the different experiments considered for each noise situations, can be seen for each estimated parameters in the following files: StateVarianceUncorr.pdf StateVarianceCorr.pdf

Experiments on real sequences

The following videos show the performance of compared algorithms in real sequences, preprocessed by the skin segmentation procedure proposed in [1]

Sequence 1: Circular movement	Comments
	The performance of all algorithms is fair, until occlusions appear. However, the UKF delimits less precisely the hand contour. When the first occlusion occurs, in this example the best performance is given by the UPF. The PF, RBPF and PS show a coherent behaviour and in fact, deform the contour to shapes of higher likelihood according to contours in the image, although this edges correspond to the occluding object. The UKF performance is strongly penalized by the occluding object. With the second occlusion, the PF temporally misstracks the pointing finger, while the performance of other methods is fair.
Sequence 2: Quadrangular movement	Comments
	The performance of all algorithms is fair, until occlusions appear. However, the UKF and RBPF react slightly slower to the hand movement. All methods manage quite nicely the situations of occlusions, although the UKF and RBPF recover later from them.
Sequence 3: Sinusoidal movement	Comments
	The performance of all algorithms is fair, until occlusions appear. In the first occlusion, the PF is the best performing method. UPF and PS get confused for a while, but recover fast from their failure. UKF and RBPF also get confused, but for a longer period. When the second occlusion occurs, the UKF is the method showing more problems. Finally, in the third occlusion, all methods show a similar performace.

Pedestrian Tracking Experiment

A contour model of a pedestrian walking sideways has been constructed from the principal component analysis of the silhouettes extracted from the following video sequence. In our experiments, a flag is used to flip the learned model when pedestrians walking from left to right in the image have to be tracked.

Training Sequence	Comments
	The segmentation of the pedestrian in this sequence by diferent methods resulted in 383 training shapes, each one modelled by a cubic spline with 32 control points.

Experiments on real sequences

In our experiments, frames are first processed with a rough background subtraction method. The background model is constructed by estimating the Gaussian distribution of the RGB values of each pixel in the whole sequence. Then, for each image, pixels whose RGB values are out of the 98% confidence region of the learned Gaussian are labelled as foreground. These generates as result sequences with a significant presence of correlated artifacts (pedestrian reflections on the ground and shadows). In order to check the performance of algorithms also under more fair conditions, its performance is also studied in sequences where these correlated artifacts have been manually eliminated. However, note than the foreground blobs are still far from an ideal pedestrian segmentation, showing holes and an irregular outline.

Sequence 1 : Pedestrian walking from left to right.
Background subtraction	Comments
	At the first half of the sequence, UPF, PF and PS fit better to the silhouette. In the middle of the sequence, the UKF starts to degrade its performance, due to its bad estimation of scaling parameters. This does not happens in the RBPF, since these parameters are estimated by means of particles, which keep them in an a priori given range of feasible values. When shadows appear ( which are correlated artifacts attached to the pedestrian side), filters using some kind of Gaussian aproximations at some point (UKF, UPF and RBPF) start to have problems, while the performance of methods purelly based on particles remains robust (PF and PS) .
Background subtraction with reflections, shadows and small artifacts manually removed.	Comments
	The holes in the segmentation generated by the naked arms of the pedestrian challenge the UKF performance, and more slightly RBPF and UPF. All particle based filters show a remarkable robust performance.

Sequence 2 : Pedestrian walking from left to right, with uncorrelated artifacts.
Background subtraction.	Comments
	Particle-based filters fit more tightly to the silhouette than UKF, which at the middle of the video starts to misstrack the shape. RBPF overcomes the UKF performance, since by managing the scale parameters with particles, these parameters are constrained inside an a priori range of feasible values. When shadows appear, filters using some kind of Gaussian aproximations at some point (UKF, UPF and RBPF) start to have problems, while the performance of methods purelly based on particles remains robust (PF and PS) .
Background subtraction with shadows and small artifacts manually removed.	Comments
	Although all methods have an acceptable performance, the UPF fits better the tracked silhouette. In the middle of the sequence, the UKF misstracks temporally the shape, but later it recovers from this failure.

References

[1] L. Sigal, S. Sclaroff and V. Athitsos. Skin Color-Based Video Segmentation under Time-Varying Illumination. IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 26, no. 7, pp. 862-877, July 2004

Last modification: April 17th, 2009