A framework for object detection, tracking and classification in urban traffic scenarios using stereovision

Abstract

Driving assistance systems provide either safety or comfort functions. Such systems must evaluate the state of the world and take necessary actions. A preliminary step for evaluating the state of the world is to detect, track and classify scene objects. The classification step becomes especially important in complex urban traffic scenarios. In such scenarios the sensors of choice are vision based, as they provide detailed scene information. In this paper we present the architecture of a detection, tracking and classification framework based on stereovision. We detect objects using either a points grouping algorithm (for large objects) or a density map grouping algorithm (for small objects). We perform first a rough classification based on objects' dimensions and track objects according to the motion model of each class. We extract motion features and perform a refined classification. Class specific dichotomizers are subsequently used to filter the classified objects, rejecting incorrect classification. A large database of manually labeled objects is used for determining motion models, training the classifiers and measuring the performance of the system.