Robust Dynamic Vision From A Moving Platform

Robust Dynamic Vision from a Moving Platform

Robust dynamic vision from a moving platform is a difficult task. To be "robust" a vision algorithm must work under many different lighting conditions and environments. "Dynamic vision" is the interpretation of the 3D distance and movement from a scene. This is substantially more difficult when the scenes are captured from a "moving platform" as there are also independent moving objects along with the movement from the scene. This thesis provides a solution to the issue of robustness by introducing the concept of residual images. These images are particularly robust to illumination changes, which make them ideal for correspondence matching. The minimum requirement to solve "dynamic vision from a moving platform" are binocular stereo cameras. Performing correspondence matching between sequential stereo images over time enables motion and distance estimation. There are two approaches to achieve this: the first is to integrate the information over time using pixel-wise Kalman filters, the second is to loosely couple the distance and motion estimation using two sequential stereo image pairs to obtain the scene flow. The first solution provides noise-reduced partial motion information, while the second produces 3D motion estimates. All algorithms in this thesis are detailed and supported with extensive experimental evidence. The appendices of this thesis include computational improvements to achieve real-time performance, making them usable in real-life applications.

Dynamic Vision for Perception and Control of Motion

The application of machine vision to autonomous vehicles is an increasingly important area of research with exciting applications in industry, defense, and transportation likely in coming decades. Dynamic Vision for Perception and Control of Motion has been written by the world's leading expert on autonomous road-following vehicles and brings together twenty years of innovation in the field by Professor Dickmanns and his colleagues at the German Armed Forces university in Munich. The book uniquely details an approach to real-time machine vision for the understanding of dynamic scenes, viewed from a moving platform that begins with spatio-temporal representations of motion for hypothesized objects whose parameters are adjusted by well-known prediction error feedback and recursive estimation techniques. A coherent and up-to-date coverage of the subject matter is presented, with the machine vision and control aspects detailed, along with reports on the mission performance the first vehicles using these innovative techniques built at Munich. Pointers to the future development and likely applications of this hugely important field of research are presented. Dynamic Vision for Perception and Control of Motion will be a key reference for technologist working in autonomous vehicles and mobile robotics in general who wish to access the leading research in this field, as well as researchers and students working in machine vision and dynamic control interested in one of the most interesting and promising applications of these techniques.

Robust Artificial Intelligence for Neurorobotics