NUST Institutions Library Catalogue

Visual object tracking is an active and challenging computer vision research domain having
wide range of civil and defence applications. Mean shift (MS) is a commonly used
target tracking technique due to its ease of implementation and real time response. However,
it has certain short-comings that limits its tracking performance. In this thesis short
comings of MS tracking like poor localization, complicated background distraction, partial/
full occlusion and distraction due to similar target resemblance are addressed using 2D
and 3D features.
To improve MS target localization problem due to the presence of complex/mingled
background features (in target representation), a novel 2D spatio-spectral technique is
proposed. True background weighted histogram features are identified in target model
representation using spectral and spatial weighting. A transformation is then applied to
minimize their effect in target model representation for localization improvement. Edge
based centroid re-positioning is applied to adjust/re-position the MS estimated target position
for further localization improvement. Occlusion avoidance method is developed for
MS tracking algorithm using adaptive window normalized cross correlation (NCC) based
template matching. The Bhattacharyya coefficient based similarity threshold is used to
detect partial/full occlusion and to initiate the NCC part in MS tracking. A target model
updation for background weighted histogram through online feature consistency data is
also proposed. The proposed 2D MS tracking techniques effectively solved the tracking
problems of clutter, similar target resemblance, complex/fast object movement and partial/
full occlusion.
The failure cases for proposed 2D tracking technique include guidewire tip tracking
for image guided cardiovascular interventions. The guidewire tip being thin, featureless
and deformable structure is easily distracted with its own and similar object like vane structures in neighborhood. Moreover, the tracking of guidewire under low contrast fluoroscopic
images and abrupt shape variations due to cardiac motion make the problem more
challenging. 3D visual tracking techniques are used to incorporate object depth information
to improve robustness. However, the existing 3D tracking techniques lack accuracy
and robustness mainly due to non availability of precise depth features.
In this thesis, depth features are acquired through shape from focus (SFF) technique and
integrated with spectral and spatial features for robust 3D target representation/tracking.
For 3D shape representation through SFF, a novel adaptive focus measure based on linear
combination of multiple morphological gradient operators is proposed. The morphological
edge gradient operators aided by multi-structuring elements are employed for sharpness
measurement. The robust focus measure is then computed by combining the weighted
response of gradient operators. The depth features acquired are integrated in joint histogram
with grey level intensity and texture features to develop a novel technique for real
time 3D representation and tracking of guidewire for image guided cardiovascular interventions.
The grey level intensity is represented through conventional histogram method
whereas the texture and depth features are represented through filtered local binary pattern
histogram and filtered local depth pattern histogram respectively.
The result shows the significant improvement in the accuracy, robustness and computational
efficiency through proposed 2D/3D MS tracking and depth estimation techniques.