This paper presents a novel approach to estimating the 3D velocity of vehicles from video. Here we propose using a Bayesian Network to classify objects into pedestrians and different types of vehicles, using 2D features extracted from the video taken from a stationary camera. The classification allows us to estimate an approximate 3D model for the different classes. The height information is then used with the image co-ordinates of the object and the camera's perspective projection matrix to estimate the objects 3D world co-ordinates and hence its 3D velocity. Accurate velocity and acceleration estimates are both very useful parameters in traffic monitoring systems. We show results of highly accurate classification and measurement of vehicle's motion from real life traffic video streams.

This paper considers video surveillance applied to traffic video streams. We present a framework for analyzing and recognizing different traffic behaviors from image sequences acquired from a fixed camera. Two types of interactions have been mainly considered. In one there is interaction between two or more mobile objects in the Field of View (FOV) of the camera. The other is interaction between mobile objects and static objects in the environment. The framework is based on two types of a priori information: (1) the contextual information of the camera's FOV, in terms of the description of the different static objects of the scene and (2) sets of predefined behavior scenarios, which need to be analyzed in different contexts. At present the system is designed to recognize behavior from stored videos and retrieve the frames in which the specific behaviors took place. We demonstrate successful behavior recognition results for pedestrian-vehicle interaction and vehicle-checkpost interactions.