A top-down perception approach for vehicle pose estimation

In this article, we present the application of our Generic Focusing Algorithm (GFA) to the fine pose estimation of a vehicle by monocular vision. This is a top-down algorithm that has already been used for object recognition [3] and localization purpose [2]. We assume we have a 3D model of the vehicle to detect and an estimation of the intrinsic parameters of the camera from where the image vehicle is taken. Furthermore, the vehicule is assumed decomposed in several parts for which we have a characterized detector (an estimation of the reliabiity and the accuracy of the detector is known). The recognition (and fine pose estimation) is made by detecting the differents parts. The initial values and the relationships between each part are grouped in a state vector with its a priori knowledge mean value and its covariance matrix. The selection of the best part to detect is achieved by a Bayesian network. The state update is done thanks to a Kalman filter. This allows to focus in an optimal way the detection of each part and, in case of detection failure, to dynamically decide which hypothesis has to be tried. Our GFA can give the final estimation pose in around 0.40s having all the detections already done. Having a rough angle and a first estimation of the translation, our GFA is able to well refine the pose.