Adaptive Kalman filtering for histogram-based appearance learning in infrared imagery

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• Infrared target tracking faces challenges due to appearance variations from target maneuvers, sensor motion, and background clutter.
• Maintaining reliable detection and track lock over time is difficult with conventional methods.
• Robust target representation and online appearance learning are critical for effective tracking.

Purpose of the Study:

• To develop a robust target representation and online appearance learning method for infrared target tracking.
• To investigate the application of autocovariance least-squares (ALS) for adaptive Kalman filtering in visual tracking.
• To evaluate the performance of the ALS method against covariance matching and histogram similarity-based methods.

Main Methods:

• Utilized a pixel intensity histogram and local standard deviation distribution model for target representation.
• Formulated appearance learning as an adaptive Kalman filtering problem with unknown noise variances.
• Applied both covariance matching and the autocovariance least-squares (ALS) method for appearance learning.

Main Results:

• Demonstrated the effectiveness of the ALS method under piecewise stationarity assumptions, beyond global stationarity.
• Simulated results show performance advantages of ALS over covariance matching for stationary and nonstationary systems.
• Real-world data validation shows ALS-based tracking outperforms covariance matching and histogram similarity methods, achieving sub-pixel accuracy.

Conclusions:

• The ALS method offers a significant advancement for infrared target tracking, providing robust performance against appearance variations.
• The proposed approach achieves high accuracy and reliability in challenging tracking scenarios.
• This work validates the practical applicability of ALS in visual tracking beyond theoretical assumptions.