Fast and Reliable Object Classification in Video Based on a 3d Generic Model

We propose a new object classification approach for monocular video sequences, which allows to classify objects modelled independently from the camera position and object orientation. To achieve this independence, a simple 3D object model that represents an object as a parallelepiped is proposed. The approach is able to give good estimates of object dimensions and proposes visual reliability measures for the object dimensions. These measures give a representation of the visibility of the estimated dimension and are principally proposed to aid posterior phases of the video understanding process, as object tracking and event detection. The method obtains the 3D parallelepiped model estimation using a set of 2D moving regions (obtained in a segmentation phase), the perspective matrix transform (obtained from camera calibration using the pin-hole camera model) and predefined 3D models of expected objects in the scene. After classification, a merging step is performed to improve the classification performance by assembling 2D moving regions with better 3D model probability when together. This approach shows promising results on object classification, obtaining very high detection rates for complex situations and performing at video frame rate.