Video Signal Processing: Compression Segmentation and Tracking

Doctoral thesis, 2010

This thesis considers three separate research problems within the field of video processing.The first is concerned with segmenting an image, the second with tracking an object and the third with the communication of video across an unreliable network. Segmentation is an application-specific problem. An image should be segmented to distinguish interesting regions, but it should not be segmented further. This thesis proposes an algorithm to optimally segment an image based on a maximum-likelihood criterion. The proposed algorithm is a modification of the popular mean-shift algorithm. However, unlike mean-shift, the proposed algorithm uses a model to compute the most likely image segmentation. It achieves this while preserving both the simplicity and speed of mean-shift. In this thesis two distinct methods are proposed to track objects. The first builds upon the joint anisotropic mean-shift and particle-filter framework. This framework consists of a gradient-ascent algorithm which is seeded by a particle filter to find all likely positions, orientations and scalings of a target. We have improved this algorithm by including spatial information in the description of the target. This makes the algorithm more robust against partial occlusions and background clutter. The second object-tracking algorithm is an improvement to the eigenface-based tracking algorithms. These algorithms represent a single appearance of a target by an interpolated NxN image and the multitude of appearances which a target can take by a linear subspace of all NxN images. It is shown that the subspace can be tracked using a Kalman filter. This is a better framework for tracking as it allows motion models and appearance models to be more accurately described. The research contributions related to video communication focus on specifically on packet-based computer networks. We propose that by monitoring the variations in transmission speeds of data packets,it is possible to predict the amount of congestion in the network. This allows us to predict the probability of packet loss in such a way that adaptive compression algorithms can be designed to efficiently deal with the expected packet loss. A modified quantized frame-expansion is then proposed in this thesis for this purpose. Using a gradient-descent algorithm, optimal transforms are found for the error-resilient transmission of data. This transform has been incorporated into a novel video-compression algorithm.