Learning Dynamic Shape Models for Bayesian Tracking
J. Giebel · MADOC (University of Mannheim) · 2005
This Thesis addresses appearance-based techniques for modelling complex deformable objects in order to support a detection and tracking process. As example application it considers pedestrian tracking from a moving vehicle. The a-priory knowledge about the object class is captured by a novel probabilistic spatio-temporal shape representation, which consists of a set of distinct linear subspace models and handles continuous as well as discontinuous shape changes. An un-supervised learning algorithm is presented, which fully automatically derives the proposed representation from training sequences of closed contours. In contrast to previous work, no prior feature correspondences are required. The 2D representation is applied to improve the performance of matching systems that correlate using shape templates. The basic idea involves extending an existing set of training shapes with generated "virtual" shapes, in order to improve the representational capability. In the experiments ROC curves demonstrate that generating "virtual" shapes by sampling the linear subspace models of the proposed representation increases the detection rate while simultaneously reducing the false alarm rate of a correlation-based object detection system. A framework for multi-cue 3d object tracking is introduced, which approximates optimal Bayesian tracking by a particle filter and combines the three cues shape, texture and depth, in its observation density function. The framework integrates an independently operating object detection system by means of importance sampling. In the experiments the approach is evaluated on the challenging topic of pedestrian tracking from a moving vehicle in synthetic, rural, and complex urban environments. It is shown on thousands of images, that the proposed approach improves upon a cutting edge pedestrian protection system, that was designed for the European community funded project PROTECTOR.