ScholarWorks@중앙대학교

초록

In this paper, we reconceptualize visual tracking as a multivariate time-series forecasting (MTSF) problem. Specifically, the goal of visual tracking-predicting the target's state over time, including its center coordinates and scale-can be naturally framed as forecasting future states from a sequence of past observations. Viewed through this lens, visual tracking aligns with the challenges of MTSF, where the objective is to capture complex temporal dependencies among multiple variables. However, applying MTSF to visual tracking introduces new difficulties due to the inherently intricate nature of object motion, which often involves abrupt and nonlinear variations in direction, velocity, and behavioral patterns. To address these complexities, we propose a principled approach that models the dynamic nature of target motion using a regime-switching framework. This method employs an underlying Markov jump process (MJP) to govern transitions between multiple latent motion patterns, each characterized by its own stochastic differential equation (SDE). By doing so, our model adapts to diverse temporal dynamics in a data-driven manner, enabling robust and precise prediction of future target states. Experimental results demonstrate that our method outperforms state-of-the-art visual tracking approaches, particularly in scenarios where target objects exhibit diverse and dynamic motion patterns over time.

키워드