ABSTRACT
Object tracking is an important topic in multimedia, particularly in applications such as
teleconferencing, surveillance and human-computer interface. The goal of object tracking is to
determine the position of the object in images continuously and reliably in dynamic scenes that
mean scenes that are prone to illumination changes, object to scene occlusions and cluttering.
The key steps involved in object tracking are foreground detection to detect moving objects,
clustering to enable representation of an object by its centroid and tracking the centroids to
determine the motion parameters. In this thesis, we implement an object tracking system on a
hardware accelerator that is optimized for low power.
Power optimization can be achieved at various levels: circuit level, logic level,
architecture level, memory level and algorithmic level. We optimize our implementation for
power by making algorithm-level modifications. First, we explore the different foreground
detection techniques to figure out the best technique that involves the least computations without
compromising on the performance. We find that the Gaussian Mixture Model proposed by
Zivkovic gives the best performance with respect to both accuracy and number of computations.
We further try to optimize this for implementation on a parallel hardware accelerator by adopting
a parallelization technique that exploits cache-line locality.
Clustering algorithms are inherently sequential and offer very less scope of
parallelization. We propose a new clustering algorithm that uses a combination of both
DBSCAN and K-Means algorithm along with a diagnostic algorithm on K-Means to estimate the
right number of centroids and show that this is faster than the DBSCAN algorithm with minimal
loss in accuracy.
We implement the 2D Kalman filter assuming constant acceleration model. Since the
computations involved in Kalman filter is just a set of recursive equations, the sequential model
in itself exhibits a good performance, thereby alleviating the need for parallelization. We
achieved a speedup of 5.8 for Gaussian Mixture model and a 2.5x improvement in the
performance of our clustering algorithm over the regular DBSCAN algorithm for 15 frames of a
video sequence of cars captured from a stationary camera that was taken from the PETS
database.