Detection of Moving Objects by Spatio-Temporal Motion Analysis
Clemens Rabe · Multimedialen Archiv und Publikationsserver der Christian-Albrechts-Universität zu Kiel (Christian-Albrechts-Universität zu Kiel) · 2012
Driver assistance systems of the future, that will support the driver in complex driving situations, require a thorough understanding of the car's environment.This includes not only the comprehension of the infrastructure but also the precise detection and measurement of other moving traffic participants.In this thesis a novel principle is presented and investigated in detail, that allows the reconstruction of the 3D motion field from the image sequence obtained by a stereo camera system.Given correspondences of stereo measurements over time, this principle estimates the 3D position and the 3D motion vector of selected points using Kalman Filters, resulting in a realtime estimation of the observed motion field.Since the state vector of the Kalman Filter consists of six elements, this principle is called 6D-Vision.To estimate the absolute motion field, the ego-motion of the moving observer must be known precisely.Since cars are usually not equipped with high-end inertial sensors, a novel algorithm to estimate the ego-motion from the image sequence is presented.Based on a Kalman Filter, it is able to support even complex vehicle models, and takes advantage of all available data, namely the previously estimated motion field and eventually available inertial sensors.As the 6D-Vision principle is not restricted to particular algorithms to obtain the image measurements, various optical flow and stereo algorithms are evaluated.In particular, a novel dense stereo algorithm is presented, that gives excellent precision results and runs at real-time.In addition, two novel scene flow algorithms are introduced, that measure the optical flow and stereo information in a combined approach, yielding more precise and robust results than a separate analysis of the two information sources.The application of the 6D-Vision principle to real-world data is illustrated throughout the thesis.As practical applications usually require an object understanding rather than a 3D motion field, a simple yet efficient algorithm to detect and track moving objects is presented.This algorithm was successfully implemented in a demonstrator vehicle that performs an autonomous braking respifnextchar.. steering manoeuvre to avoid collisions with moving pedestrians.