In Defense and Revival of Bayesian Filtering for Thermal Infrared Object Tracking

TL;DR

This work addresses the fragility of deep-learning–only TIR tracking by introducing Deep Bayesian Filtering (DBF), a dual-model framework that fuses motion priors from a $2$D Brownian-motion system model with infrared-based likelihoods from a deep observer. The system and observation models are integrated through the Bayesian update to produce a posterior for target position, enabling stable template updates in challenging scenarios. Experimental results on LSOTB-TIR and PTB-TIR show DBF achieving competitive or superior performance, with the Laplace-based motion model offering improvements over Gaussian. The approach demonstrates that incorporating motion data and probabilistic update mechanisms can significantly enhance robustness in TIR object tracking and points toward future integration with Transformer-based observations and super-resolution techniques for further gains.

Abstract

Deep learning-based methods monopolize the latest research in the field of thermal infrared (TIR) object tracking. However, relying solely on deep learning models to obtain better tracking results requires carefully selecting feature information that is beneficial to representing the target object and designing a reasonable template update strategy, which undoubtedly increases the difficulty of model design. Thus, recent TIR tracking methods face many challenges in complex scenarios. This paper introduces a novel Deep Bayesian Filtering (DBF) method to enhance TIR tracking in these challenging situations. DBF is distinctive in its dual-model structure: the system and observation models. The system model leverages motion data to estimate the potential positions of the target object based on two-dimensional Brownian motion, thus generating a prior probability. Following this, the observation model comes into play upon capturing the TIR image. It serves as a classifier and employs infrared information to ascertain the likelihood of these estimated positions, creating a likelihood probability. According to the guidance of the two models, the position of the target object can be determined, and the template can be dynamically updated. Experimental analysis across several benchmark datasets reveals that DBF achieves competitive performance, surpassing most existing TIR tracking methods in complex scenarios.

Figures (9)

• Figure 1: Overview of the state space of DBF. It consists of system model and observation model.
• Figure 2: Overview of the Observation Model. The observation $z_t$ is the $t$-th frame raw image. Candidates are cropped on $z_t$ according to the system state generated by the system model, and the feature maps of each candidate are obtained. $v_i$ is the response score output by the classifier to the system state $s_t^i$ based on the candidate region. The classifier is learned through the feature maps of the target object.
• Figure 3: Comparison on the LSOTB-TIR lsotb benchmark dataset.
• Figure 4: Evaluation of four scenario attribute subsets of the LSOTB-TIR lsotb benchmark dataset.
• Figure 5: Evaluation of 12 challenge attribute subsets of the LSOTB-TIR lsotb benchmark dataset.