We focus on nine societal areas.
TNO cooperates with companies, the public sector and other organisations, to apply our knowledge and expertise with and for others.
We develop knowledge not for its own sake, but for practical application.
TNO offers you the chance to do groundbreaking work and help customers and society with innovative, practical and smart solutions.
On our TNOTIME platform you can find blogs, interviews and background stories. Discover the field of TNO captured in photographs, animations, video and text. Share interesting articles and powerful infographics, or download whitepapers.
This invention is designed to evaluate the motion of multiple cameras with (at least partially) overlapping views of a 3D space. Many methods for estimating this type of motion exist, but they are inaccurate due to misperceptions of landmarks, as well as inefficient due to the high volume of processing power they require. The method succeeds empirically compared to a well-used existing approach known as RANSAC.
This invention adds extra parameters to existing techniques for estimating the motion of a camera array. These include the calculation of likely error in order to reduce the impact of outlier landmarks on the overall estimate. The method uses a loop of estimates, moving back and forth between the large dataset and the single motion estimate; an algorithm will check the motion estimate against the evaluations of landmarks. The method also outputs a final uncertainty value associated with its estimate.
One particular use for the invention is in estimating the motion of a moving vehicle using two visual inputs arrayed in stereo. Motion of the vehicle is determined by establishing corresponding landmarks between the two cameras. The invention can be hardwired into a camera system or hosted in a computer program, potentially stored on a CD or DVD, that can be installed to perform this type of estimate.
The method begins with simple left-to-right matching of pixels between the different cameras, followed by triangulation used to estimate position relative to a primary object. The method calls for a set of uncertainties associated with each point. These data allows the algorithm to weight its motion estimates on the basis of likely data quality. The utility of landmarks for the estimate will be weighted according to a system of minimal squared distances. The algorithm evaluates the Bhattacharyya distance between the probability distributions found by the initial evaluation.