Learning Local-Global Multiple Correlation Filters for Robust Visual Tracking with Kalman Filter Redetection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Visual object tracking is crucial for camera-based sensor networks.
• Correlation filter (CF)-based trackers using convolutional features show promise but struggle with appearance variations and irrational filter updates, leading to tracking failures.

Purpose of the Study:

• To propose a novel local-global multiple correlation filters (LGCF) tracking algorithm for edge computing systems.
• To enhance the robustness and accuracy of visual object tracking, particularly for moving targets like vehicles and pedestrians.

Main Methods:

• Constructed a global correlation filter with deep convolutional features and two local filters using hand-crafted features, divided based on aspect ratio.
• Implemented a local-global collaborative strategy for information exchange between filters to prevent erroneous object appearance model learning.
• Introduced a time-space peak to sidelobe ratio (TSPSR) for stability evaluation and a Kalman filter redetection (KFR) model for object recapture when tracking is unreliable.

Main Results:

• The proposed LGCF algorithm demonstrated superior performance on the OTB-2013 and OTB-2015 datasets compared to 12 other state-of-the-art tracking algorithms.
• The algorithm effectively handled various challenging situations in object tracking, indicating improved robustness.

Conclusions:

• The LGCF tracking algorithm offers a significant advancement in visual object tracking for edge computing applications.
• The combination of local-global filters, collaborative strategy, and stability evaluation enhances tracking reliability and accuracy, especially in complex environments.