Severity analysis of abnormal traffic events at intersections

In this work, a novel real-time approach is designed to detect abnormal events in real-world traffic videos and evaluate the severity characteristics of the events by classifying into two severity classes. In order to detect the abnormal events, trajectory of normal vehicle motions are clustered and common route models are learned by using Continuous Hidden Markov Model. In the second part, partial vehicle motions are observed by Maximum Likelihood method in order to detect abnormal vehicle events. In the third part, the severity definition and classification is done for abnormal events using k-Nearest Neighborhood and Support Vector Machines methods. A training set is formed by collided vehicle motion videos, feature vectors are extracted from these videos and these vectors are labeled into low and high severity classes and classification module is trained by these event samples. The abnormality detection tests using real-world videos indicate that the proposed system reaches ip to 89% correct detection rate. The correct severity classification rate of real world events goes up to 75%. The results indicate that abnormal events can be detected and represented by likelihood probabilities, and depending on these probabilities, severity analysis can be done successfully.