Computational geometric-based visual shape tracker
Loading...
Date
2021-07
Authors
Bandara, Pasan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The main intent of multiple shape tracking is to allocate distinct track
identities for all the salient shapes in a video sequence. The approach used
in this problem is ‘Track by Detection’ scheme, where it needs to first, find
shapes in a frame, map shapes to appropriate tracking tracks, and maintain
the data continuity model throughout the video sequence. This thesis proposes
a process to address shape tracking by maintaining a data continuity model
based on shape relative proximity, shape context descriptor, and spatial color
variation of a shape. Shape intersection is used to calculate the shape proximity.
A spatial color analyzer is used to compare shape’s color variations between
two consecutive frames, and shape similarity comparison is done by using the
shape context descriptor. These values are then used to build and maintain
the data continuity model throughout the video sequence. Besides, this also
introduces a methodology to reestablish tracking identity when the detection is
failed or occluded temporarily between small numbers of frames. The proposed
method can be implemented using both online and offline approaches. To assess
the proposed method’s robustness, we evaluated the online method in both
VOT-ST2020 (Visual Object Tracking Challenge - Short Term) and MOT20
(Multiple Object Tracking) benchmark datasets. It was observed that in both
benchmarks, the proposed approach can deliver state-of-the-art results.
Description
Keywords
Computer vision, Shape tracking