A variational approach to simultaneous tracking and classification of multiple objects

This paper presents a method for multi-object tracking which provides estimates of the dynamic state of the objects along with class identities. The estimated identities provide information about the objects' behaviour, improving high level reasoning tasks. However, jointly estimating class assignments, dynamic states and data associations results in a computationally intractable problem. This paper proposes a probabilistic model for the multi-object tracking and classification problem, and an inference procedure that renders the problem tractable through a variational approximation. Our framework integrates the efficient Kalman filtering and smoothing recursions into a system that considers the dynamics of the environment to leverage both tracking and classification. The method is evaluated and compared to state-of-the-art techniques using stereo-vision data collected from a moving platform in urban scenarios.