sensing what matters

project page of the PhD research by Wilbert van Norden


 

about me


Wilbert van Norden (born in 1981 in the Netherlands) has been working at CAMS Force Vision in Den Helder since January 2006. Force Vision is a part of the Dutch Defense and Material Organisation. Force Vision designs, implements and maintains all operational software for the Royal Netherlands Navy.

At Force Vision, Wilbert works in the Planning and Decision support department where he focuses on Sensor Management. Based on his research Wilbert is persuing his PhD title in cooperation with Delft University of Technology, initially under supervision of prof.dr.drs. Rothkrantz and since 2007 under supervision of prof.dr. Catholijn Jonker. Since the work is based on a previous PhD study (also partly executed at CAMS Force Vision) by Lt.CDR Fok Bolderheij, he is also involved in this research.

Recently, Krispijn Scholte started his MSc project where he helps me to enable the real-time implementations of some of the concepts that resulted from the research.

Topics include:
  • Classification;
  • Information Fusion;
  • Sensor coverage estimation;
  • Sensor Management.


 contact information

Address               CAMS – Force Vision
t.a.v. Wilbert van Norden
MPC 10A
P.O. Box 10000
1780 CA  Den Helder
the Netherlands

phone                 : +31 223 653894
mail                     : w.l.van.norden[AT]forcevision.nl

 


the research

Execution of more complex missions in more complex environments with advanced systems gives rise to a high level of automation due to less readily available knowledge on board Royal Netherlands Navy (RNLN) ships. Growing complexity in sensor systems requires more knowledge to utilise those systems. This available knowledge on board ships of the RNLN however is decreasing due to a strive to reduce ship's complements and to reduce education and training time. This (growing) discrepancy between available and required knowledge requires research in command and control processes and sensor management in particular.

Complex missions call for a well considered deployment of sensors since not all objects in the environment can be observed at the same time and it is critical to mission success to identify threats as soon as possible. Based on timely observed threats the appropriate actions can be chosen to ensure mission success and safety of the ship. Identifying threats correctly means that the available information on objects is sufficient to make these distinctions. For mission success it is therefore essential to optimally deploy sensors to obtain the relevant information about objects.

In order to determine which objects pose the most risk against mission success the classification process is essential. Based on the classification solution objects can be ruled out as a threat or identified as one. The focus of this research therefore is to automate the classification process as well as using uncertainty in classification as an input for sensor management. Sensor management should try to reduce the uncertainty on the most dangerous objects as best as possible and should therefore start with reducing the uncertainty about the classification.

Reducing uncertainty to improve the classification process not only implies knowing what information to obtain but also how this information may be obtained. In order to reduce uncertainty as best as possible, the sensor settings need to be adapted to the situation. Knowledge on how the sensors operate in different environments is required to optimise performance to fulfil the information requirements. Existing tools for calculating sensor performance are needed within the sensor management system to determine the required settings for particular information needs.

Due to the large amount of object in today's complex missions, the information requirements are substantial which requires a prioritisation mechanism for sensor tasks. Determining which information is considered to be more important than other information is directly related to how much threat the object under consideration poses to the mission. For prioritisation purposes the worst possible case in assumed which gives the upper boundary of the risk estimation. This maximum possible risk an object poses is used as the priority for the information need that is determined for that object.

Describing the information requirements for classification requires a new classification model-based methodology. Traditional classifiers assume a fixed label-set whereas the complex missions require this set to be flexible. Classifiers should also be able to cope with uncertain input which is not the case in traditional systems. Since training sets are not widely available in the maritime military domain a system based on models of possible objects is a logical step. Model-based classifiers are build to deal with uncertain information and can deal with changing label sets as well as indicate which information is needed to improve the classification accuracy.

The model-based classification methodology leads to a number of classifiers that are combined in a single system. In this system the operator may influence the process by either changing the label set based on his knowledge of the situation or by providing an actual classification solution. The chosen combination rule is Dezert-Smarandache theory (DSmT) since this theory can handle highly conflicting sources that express belief on solutions sets with overlapping labels. The user preferences can be taken into account as well by some additional rules that have been developed. Using DSmT to combine the different classifiers leads to a single classification system.

For any classification systems it is important to verify performance based on suited evaluation criteria. These criteria are determined by the application domain in which the classification system needs to operate. In this case, the overall system needs to deal with a large and changing label-set that may be hierarchical, uncertain input, and it needs to admit to uncertainty and/or confusion between labels. The latter characteristic is referred to as soft classification for which evaluation criteria have been developed. These however are not suited for non-hierarchical label-sets. Criteria that are suited for these types of sets however cannot handle soft classifiers. New criteria have been developed to appropriately evaluate the performance of the classification system based on the characteristics of the problem domain.

Tests with the new classification system shows an improvement of classification results over traditional systems and that the new system is capable of describing the information requirements. In a simulated environment several classification systems have been tested and compared. This comparison is done on existing criteria as well as the newly developed ones. On all accounts, the new classification system outperforms existing methodologies.

In short, combining different model-based classifiers with DSmT leads to improved classification of objects as well as an improvement in sensor deployment and this system is therefore essential to successfully execute the missions the RNLN is faced with today and in the future.

 

 

my publications

2009

2008

2007

2006

2005