Bayesian model for feature representation fusion in person re-identification

This paper addresses the problem of person re-identification in surveillance systems based on the analysis and integration of heterogeneous descriptive features. Unlike traditional approaches relying on a single biometric modality, a general-purpose probabilistic method is proposed to combine features derived from various sources. The method is based on Bayesian inference and models each feature as a random variable following a multivariate normal distribution. Parameters of the distributions are estimated from the training data. Decision making is performed by maximizing the posterior probability of identity given the available evidence. To extract features, transformer-based architectures with attention mechanisms are used, ensuring robustness to visual noise and viewpoint variation. The proposed model is compared with classical approaches based on linear combination of feature vectors. Experiments conducted on an open re-identification dataset demonstrate that the Bayesian scheme improves recognition accuracy and remains effective in cases of partial information loss. Moreover, the method provides a quantitative confidence score associated with each decision, making it particularly suitable for deployment in safety-critical environments such as automated video surveillance and access control systems.