Naeem, Asad (2010) Single and multiple target tracking via hybrid mean shift/particle filter algorithms. PhD thesis, University of Nottingham.
This thesis is concerned with single and multiple target visual tracking algorithms and their application in the real world. While they are both powerful and general, one of the main challenges of tracking using particle filter-based algorithms is to manage the particle spread. Too wide a spread leads to dispersal of particles onto clutter, but limited spread may lead to difficulty when fast-moving objects and/or high-speed camera motion throw trackers away from their target(s). This thesis addresses the particle spread management problem. Three novel tracking algorithms are presented, each of which combines particle filtering and Kernel Mean Shift methods to produce more robust and accurate tracking.
The first single target tracking algorithm, the Structured Octal Kernel Filter (SOK), combines Mean Shift (Comaniciu et al 2003) and Condensation (Isard and Blake 1998a). The spread of the particle set is handled by structurally placing the particles around the object, using eight particles arranged to cover the maximum area. Mean Shift is then applied to each particle to seek the global maxima. In effect, SOK uses intelligent switching between Mean Shift and particle filtering based on a confidence level. Though effective, it requires a threshold to be set and performs a somewhat inflexible search.
The second single target tracking algorithm, the Kernel Annealed Mean Shift tracker (KAMS), uses an annealed particle filter (Deutscher et al 2000), but introduces a Mean Shift step to control particle spread. As a result, higher accuracy and robustness are achieved using fewer particles and annealing levels. Finally, KAMS is extended to create a multi-object tracking algorithm (MKAMS) by introducing an interaction filter to handle object collisions and occlusions.
All three algorithms are compared experimentally with existing single/multiple object tracking algorithms. The evaluation procedure compares competing algorithms' robustness, accuracy and computational cost using both numerical measures and a novel application of McNemar's statistic. Results are presented on a wide variety of artificial and real image sequences.
|Item Type:||Thesis (PhD)|
|Faculties/Schools:||UK Campuses > Faculty of Science > School of Computer Science|
|Deposited By:||Mrs Olga Lashkova|
|Deposited On:||21 Jun 2012 14:27|
|Last Modified:||21 Jun 2012 14:29|
Archive Staff Only: item control page