Bernd Korn

A task driven 3D object recognition system using Bayesian networks

In this paper we propose a general framework to build a task oriented 3D object recognition system for CAD based vision (CBV). Features from 3D space curves representing the objects rims provide sufficient information to allow identification and pose estimation of industrial CAD models. However, features relying on differential surface properties tend to be very vulnerable with respect to noise. To model the statistical behavior of the data we introduce Bayesian nets which model the relationship between objects and observable features. Furthermore, task oriented selection of the optimal action to reduce the uncertainty of recognition results is incorporated into the Bayesian nets. This enables the integration of intelligent recognition strategies depending on the already acquired evidence into a robust, and efficient, 3D CAD based recognition system.

A fuzzy ICP algorithm for 3D free-form object recognition

We propose a new object matching algorithm which can separate overlapping objects and which is robust against erroneous data. The algorithm is based on the well-known iterative closest point (ICP) algorithm. However all published contributions to the ICP algorithm can not provide a proper segmentation of the input data. A fuzzy ICP algorithm can handle these problems by a fuzzy membership valuation at each iteration level. Furthermore, we introduce an evidence accumulation algorithm which allows a determination of the best match. In combination with search routines for the most common CAD models we provide powerful and efficient tools for CAD based object recognition systems.

Navigation integrity monitoring and obstacle detection for enhanced-vision systems

Typically, Enhanced Vision (EV) systems consist of two main parts, sensor vision and synthetic vision. Synthetic vision usually generates a virtual out-the-window view using databases and accurate navigation data, e. g. provided by differential GPS (DGPS). The reliability of the synthetic vision highly depends on both, the accuracy of the used database and the integrity of the navigation data. But especially in GPS based systems, the integrity of the navigation cant be guaranteed. Furthermore, only objects that are stored in the database can be displayed to the pilot. Consequently, unexpected obstacles are invisible and this might cause severe problems. Therefore, additional information has to be extracted from sensor data to overcome these problems. In particular, the sensor data analysis has to identify obstacles and has to monitor the integrity of databases and navigation. Furthermore, if a lack of integrity arises, navigation data, e.g. the relative position of runway and aircraft, has to be extracted directly from the sensor data. The main contribution of this paper is about the realization of these three sensor data analysis tasks within our EV system, which uses the HiVision 35 GHz MMW radar of EADS, Ulm as the primary EV sensor. For the integrity monitoring, objects extracted from radar images are registered with both database objects and objects (e. g. other aircrafts) transmitted via data link. This results in a classification into known and unknown radar image objects and consequently, in a validation of the integrity of database and navigation. Furthermore, special runway structures are searched for in the radar image where they should appear. The outcome of this runway check contributes to the integrity analysis, too. Concurrent to this investigation a radar image based navigation is performed without using neither precision navigation nor detailed database information to determine the aircrafts position relative to the runway. The performance of our approach is demonstrated with real data acquired during extensive flight tests to several airports in Northern Germany.

3D B-spline curve matching for model based object recognition

Introducing general CAD descriptions in object recognition systems has become a major field of research called CAD based vision (CBV). However, the major problem using free-form object descriptions is how to define recognizable features which can be extracted from sensor data. We propose new methods for the extraction of 3D space carves from CAD models and from range data. Object identification is performed by correlating feature vectors from significant subcurves.

Weather independent flight guidance: analysis of MMW radar images for approach and landing

We present a system which provides such navigation information based on the analysis of millimeter wave (MMW) radar data. The advantage of MMW radar sensors is that the data acquisition is independent from the actual weather and daylight situation. The core part of the presented system is a fuzzy rule based inference machine which controls the data analysis based on the uncertainty in the actual knowledge in combination with a-priori knowledge. Compared with standard TV or IR images the quality of MMW images is rather poor and the data are highly corrupted with noise and clutter. Therefore, one main task of the inference machine is to handle uncertainties as well as ambiguities and inconsistencies to draw the right conclusions. The performance of our approach is demonstrated with real data acquired during extensive flight tests to several airports in Northern Germany.

Handling Uncertainty in 3D Object Recognition Using Bayesian Networks

In this paper we show how the uncertainty within a 3d recognition process can be modeled using Bayesian nets. Reliable object features in terms of object rims are introduced to allow a robust recognition of industrial free-form objects. Dependencies between observed features and the objects are modeled within the Bayesian net. An algorithm to build the Bayesian net from a set of CAD models is introduced. In the recognition, entering evidence into the Bayesian net reduces the set of possible object hypotheses. Furthermore, the expected change of the joint probability distribution allows an integration of decision reasoning in the Bayesian propagation. The selection of the optimal, next action is incorporated into the Bayesian nets to reduce the uncertainty.

Extending enhanced-vision capabilities by integration of advanced surface movement guidance and control systems (A-SMGCS)

DLR has set up a number of projects to increase flight safety and economics of aviation. Within these activities one field of interest is the development and validation of systems for pilot assistance in order to increase the situation awareness of the aircrew. All flight phases (gate-to-gate) are taken into account, but as far as approaches, landing and taxiing are the most critical tasks in the field of civil aviation, special emphasis is given to these operations. As presented in previous contributions within SPIEs Enhanced and Synthetic Vision Conferences, DLRs Institute of Flight Guidance has developed an Enhanced Vision System (EVS) as a tool assisting especially approach and landing by improving the aircrews situational awareness. The combination of forward looking imaging sensors (such as EADSs HiVision millimeter wave radar), terrain data stored in on-board databases plus information transmitted from ground or other aircraft via data link is used to help pilots handling these phases of flight especially under adverse weather conditions. A second pilot assistance module being developed at DLR is the Taxi And Ramp Management And Control - Airborne System (TARMAC-AS), which is part of an Advanced Surface Management Guidance and Control System (ASMGCS). By means of on-board terrain data bases and navigation data a map display is generated, which helps the pilot performing taxi operations. In addition to the pure map function taxi instructions and other traffic can be displayed as the aircraft is connected to TARMAC-planning and TARMAC-communication, navigation and surveillance modules on ground via data-link. Recent experiments with airline pilots have shown, that the capabilities of taxi assistance can be extended significantly by integrating EVS- and TARMAC-AS-functionalities. Especially an extended obstacle detection and warning coming from the Enhanced Vision System increases the safety of ground operations. The presented paper gives an overview regarding those two assistance systems and discusses possible concepts and the potential of an integrated system with respect to taxi guidance operations.

Correct 3D Matching via a Fuzzy ICP Algorithm for Arbitrary Shaped Objects

We propose a new object matching algorithm which can separate overlapping objects and which is robust against erroneous data. The algorithm is based on the well-known ICP (Iterative Closest Point) algorithm. However, all published contributions to the ICP algorithm can’t provide a proper segmentation of the input data. A Fuzzy ICP algorithm can handle these problems by a fuzzy membership valuation at each iteration level. Furthermore, we introduce an evidence accumulation algorithm which allows a determination of the best match.

Erkennung und Bestimmung der aktuellen Konstellation von Objekten mit Scharniergelenken

Es wird ein Konzept zur modellbasierten Erkennung und Konstellationsbestimmung von Scharniersystemen vorgestellt. Scharniersysteme bestehen aus mehreren starren Einzelkomponenten, die durch Gelenke mit einem rotatorischen Freiheitsgrad verbunden sind. Zuerst wird bezuglich des festen Referenzmodells die Bewegungsfreiheit eines jeden Scharniers eines Scharniersystems erlernt. Da das Lernen wie auch das Erkennen der Scharnierarten aufgrund desselben Sensors erfolgt, erhalt man wahrend der Lernphase eine vollstandige Beschreibimg des Scharniersystems, die auch implizite Informationen uber Fertigungs- und Sensorungenauigkeiten enthalt. In einer anschliesenden Erkennungsphase konnen die erlernten Scharniersysteme in behebigen Szenen detektiert werden und ihre aktuellen Konstellationen bestimmt werden.

A 3D Object Recognition System with Decision Reasoning under Uncertainty

In this paper we propose a general framework to build a task oriented 3d object recognition system. To cope with noisy data under changing viewing conditions a 3d object recognition system has to acquire sensor data incrementally (active sensing) and has to choose appropriate actions to reduce the uncertainty in the recognition results (task driven recognition). To model the statistical behavior of the data we introduce Bayesian nets which model the relationship between obxad jects and observable features. Furthermore, task oriented selection of the optimal action to reduce the uncertainty of recognition results is incorxad porated in the Bayesian net. This enables the integration of intelligent recognition strategies depending on the already acquired evidence into a robust and efficient 3d model based recognition system.