Abstract
Traffic Scenes Monitoring has been a topic of large research in the last decade. An important step is the recognition of cars. Indeed, recognizing 3D models of cars could allow efficient tracking and detection. In this work we propose to develop new flexible and powerful nonparametric frameworks for the problem of data modeling and 3D recognition. In particular, we propose a Bayesian inference method via scaled Dirichlet mixture models. The consideration of scaled Dirichlet mixture is encouraged by its flexibility recently obtained in several real-life applications. Moreover, the consideration of Bayesian learning is attractive in several ways. It makes it possible to take uncertainty into account by introducing prior information on the parameters, it permits to overcome learning issues regarding the under and/or over-fitting. and it permits simultaneous parameters estimation and model selection. We investigate in this work the integration of both Markov Chain Monte Carlo (MCMC) and reversible jump MCMC (RJMCMC) techniques for learning the resulting models. Detailed experiments have been conducted to demonstrate the advantages of our Bayesian frameworks.