In order to support safe and efficient driving, it is important to classify the behaviors of vehicles and to understand what is going on with the traffic situations. For that purpose, the system employed a vision sensor rather than spot sensors because of its rich information. A dedicated vehicle-tracking algorithm based on the Spatio-Temporal MRF model, which is robust against heavy occlusions even in low-angle images, was developed. This algorithm outputs maps that represent distributions of vehicle regions and motions by every image frames. By analyzing time-series variations of these distribution maps, the system extracts observation sequences for traffic events classification. A classification network based on the Bayesian Network that consists of a fusion of deterministic sub-networks and probabilistic sub-networks was constructed. Since vehicle behaviors in the traffic should be restricted by traffic rules, such behaviors would be classified by deterministic network based on explicitly defined traffic rules. However, it is difficult to classify and to understand behaviors such as accidents, near misses, and some other ambiguous behaviors by deterministic classification networks. In order to classify such ambiguous behavior sequences, it is effective to employ probabilistic sub-networks such as Hidden Markov Model with its learning methods. Consequently, by integrating those deterministic networks and probabilistic networks into the Bayesian Network, the system was able to successfully classify behaviors of rule violations, accidents, near-miss situations distinguishing from ordinary behaviors in traffic images.
Abstract