Active Random Hypersurface Models: Simultaneous Shape and Pose Tracking of Extended Objects in Noisy Point Clouds

• Applicant: Professor Dr.-Ing. Uwe D. Hanebeck
• Subject Area: Automation, Mechatronics, Control Systems, Intelligent Technical Systems, Robotics
• Term: from 2013 to 2016
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 234520279

Tracking the pose of quickly moving extended 3D-objects based on noisy point cloud measurements from the surface of the object is an important problem in several applications. These include automobile safety, innovative control for entertainment devices, telepresence applications, and industrial production lines.Measurements are acquired by sensors such as laser scanners, depth cameras, multi-camera setups, or radar devices. In general, due to occlusion effects, only certain parts of the object are visible at a given time step. In addition, depending on where the object is located relative to the sensor, the number of measurements and their quality strongly varies. In order to accurately estimate the pose, several measurements of the object have to be sequentially collected over time while the object is in motion. Tracking the pose also requires the continuous determination of its shape, even if the shape of the target object is not the primary interest, in order to combine information from different perspectives. For shape modeling, the well-established concept of Random Hypersurface Models (RHMs) and Active Contours will be combined, resulting in the ARHMs. This will form the basis of a Bayesian tracking algorithm, which recursively estimates the shape and pose of an object simultaneously. Using RHMs, noisy measurements are related to the object parameters using explicit measurement equations with multiplicative noise. The key idea is to describe the object by applying transformations on a base shape according to a probability distribution. For example, a cylinder can be described as a circular base shape being transformed by an extrusion. This model allows information to be extracted from measurements of a point cloud without knowing where the exact source points were generated. The main research contribution of the proposal is the extension of RHMs to three-dimensional objects. This consists of the development of three new types of RHMs, which can be combined to describe more complex objects, as well as articulated structures. The application of symmetries and transformation invariances avoids redundancies in the representation, allowing even complex forms to be described using a small amount of parameters. The parametrized surface, together with regularization constraints, allows modeling of parts that have not been observed. This concept is based on ideas from Active Contours. The non-trivial combination of RHMs and Active Contours will yield an estimation method capable of reliably tracking the pose of extended 3D-objects using a parametrized form, resulting in an efficient model that allows the derivation of estimation procedures in closed form. In addition, we expect our proposed ideas and algorithms to result in a significant contribution to the field of 3D-object tracking, as our approach is based on solid mathematical fundaments, yet will still be intuitive.

• DFG Programme: Research Grants