Object state optimization algorithm based on Bayesian random sampling for visual object tracking

TL;DR

This paper introduces a new Bayesian random sampling-based algorithm to improve object tracking by better estimating object states and avoiding local optima.

Contribution

A novel hybrid model combining Bayesian random sampling and gradient ascent is proposed to enhance tracking performance.

Findings

The proposed algorithm improves tracking performance on multiple datasets.

The hybrid model successfully alleviates convergence instability in object localization.

Abstract

From the perspective of object state modeling, visual object tracking can be regarded as a unified process that combines object state estimation and object localization. In this framework, state estimation refers to predicting the complete state vector of the object–such as its position, scale, and motion dynamics–while localization specifically denotes identifying the object’s spatial position within the image, typically in the form of bounding box coordinates. Traditional optimization-based methods for state estimation often suffer from getting trapped in local optima, primarily due to the non-convexity of the objective function and the algorithm’s sensitivity to initialization. To address these issues, this research proposes an object state optimization algorithm based on Bayesian random sampling for visual object tracking. Firstly, a dense sampling method is introduced to mitigate the problem of local optima. Secondly, a hybrid model that merges Bayesian random sampling and gradient ascent is proposed to refine the bounding box, successfully alleviating convergence instability. Finally, our experimental results show that the proposed algorithm significantly improves tracking performance on multiple datasets, validating its efficiency and applicability in object state estimation tasks.

Figures