A Novel Hybrid Level Set Model for Non-Rigid Object Contour Tracking.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Current object trackers often use bounding boxes, which include background noise, reducing accuracy, especially for non-rigid objects.
• Challenges in non-rigid object tracking include topology changes, occlusions, and abrupt motion, which degrade performance.

Purpose of the Study:

• To propose a novel hybrid level set model for robust non-rigid object tracking.
• To improve tracking accuracy by focusing on object contours rather than bounding boxes.

Main Methods:

• Developed a hybrid level set model integrating an appearance model (trained with competing one-class SVMs), an edge detector, and image spatial information.
• Incorporated a motion model to predict object location during occlusions and abrupt movements.
• Implemented a feedback mechanism to update the appearance model iteratively using the tracked object contour.

Main Results:

• The proposed model accurately tracks object contours, effectively handling topology changes, occlusions, and abrupt motion.
• Experimental results on benchmarks demonstrate superior accuracy and robustness compared to existing tracking methods.
• The method shows significant improvements in non-rigid object tracking performance.

Conclusions:

• The novel hybrid level set model offers a robust and accurate solution for non-rigid object tracking.
• Accurate contour tracking and adaptive appearance/motion modeling are key to overcoming challenges in non-rigid object tracking.