Detecting unusual activities at vehicular intersections

In this work a proposal to model the activity at vehicular intersections with the aim of detecting unusual events is presented. Using the particular constraints that this kind of scenarios provide, we develop methods to detect, track, and model the activity of moving objects. Our description of activity is based on the local definition, at each pixel, of a multimodal model for the direction of motion. During operation, a particular observation is compared with the learned model. Our experiments give clear indication that the proposed scheme has a good performance in detecting such unusual events as vehicles running on red light and making forbidden turns.