A fast region-based active contour for non-rigid object tracking and its shape retrieval

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Traditional object tracking relies on bounding boxes, insufficient for non-rigid shapes needed in gait, gesture, and medical analyses.
• Active contour models (ACM) are crucial for shape retrieval but face computational challenges in video processing.
• Existing methods struggle with accurate shape separation and background differentiation for non-rigid objects.

Purpose of the Study:

• To develop an efficient active contour model (ACM) integrated with a mean-shift tracker for simultaneous video object tracking and shape retrieval.
• To address the computational time limitations of traditional contour tracing in video analysis.
• To improve the accuracy and robustness of non-rigid object tracking.

Main Methods:

• Proposed a region-based active contour model (ACM) formation utilizing a mean-shift tracker.
• Implemented contour re-initialization removal and fast deformation techniques for efficient shape retrieval.
• Modified the contour model to enable simultaneous tracking and shape retrieval capabilities.

Main Results:

• The proposed contour-based tracking achieved 71.86% accuracy in tracking and retrieving object shapes.
• The integrated mean-shift tracker effectively resolved scale-orientation selection issues in non-rigid tracking.
• The method mitigated the problem of erroneous object localization common in other tracking algorithms.

Conclusions:

• The developed contour-based mean-shift tracker offers an effective solution for non-rigid object tracking and shape retrieval in videos.
• This approach enhances accuracy and robustness compared to conventional bounding box methods.
• The technique shows promise for applications requiring precise non-rigid shape analysis.