3D Scene Segmentation and Object Tracking in Multiocular Image Sequences

In this contribution we describe a vision-based system for the 3D detection and tracking of moving persons and objects in complex scenes. A 3D point cloud of the scene is extracted by a combined stereo technique consisting of a correlation-based block-matching approach and a spacetime stereo approach based on spatio-temporally local intensity modelling. Hence, the result of stereo analysis is a 3D point cloud attributed with motion information. For localising persons and objects in the scene the point cloud is segmented into meaningful clusters by applying a hierarchical clustering algorithm, using velocity information as an additional discrimination criterion. Initial object hypotheses are obtained by partitioning the observed scene with cylinders, including the tracking results of the previous frame. Multidimensional unconstrained nonlinear minimisation is then applied to refine the position, velocity and size of the initial cylinder in the scene, such that neighbouring clusters with similar velocity vectors are grouped to form a compact object. A particle filter is applied to select hypotheses which generate consistent trajectories. The described system is evaluated based on a tabletop sequence and several real-world sequences acquired in an industrial production environment, based on manually obtained ground truth data. We find that even in the presence of moving objects closely neighbouring the person, all objects are detected and tracked in a robust and stable manner. The average tracking accuracy is of the order of several percent of the distance to the scene