Trajectory reconstruction is essential for localizing and tracking maritime vehicles, but non-Gaussian process and measurement noises, as well as time-varying measurement loss, are present in the marine environment. To enhance TR accuracy in such challenging scenarios, this article proposes a state sequence estimation method based on the Rauch-Tung-Striebel smoother, which fully leverages the distributional characteristics of outliers. This approach models the likelihood density of non- Gaussian noise using two multivariate Laplace distributions. Unknown noise covariances are then modeled via an inverse Wishart distribution, while unknown measurement losses and their occurrence probabilities are addressed using Bernoulli and Beta distributions, respectively. On this basis, Variational Bayes inference is utilized to decouple fused anomalous processes, and infer states within a hierarchical model, yielding the proposed robust smoother. Experiments with an Unmanned Underwater Vehicle and numerical simulations demonstrate the effectiveness and robustness of this proposed method for TR tasks.