Master Thesis Artificial Intelligence
Master Projects
Content
-
Prove it.
Given a statement like "Wearing Burka's must be punished.", find in the UK Hansards (Parliamentary proceedings) all paragraphs in speeches in which the speaker speaks out in favor of or against the statement.
The Hansards are available at ILPS in a richly structured XML format for the period 1935 until 2011. Same for the Netherlands. Een voorbeeld is http://resolver.politicalmashup.nl/nl.proc.sgd.d.19770000073.2.2
Supervisor: Maarten Marx (ILPS). -
Sensor fusion on a mini Unmanned Aerial System
The Dutch Aerospace Laboratory (NLR) in Amsterdam has established an Unmanned Aerial System (UAS) lab; a creative environment with flexible workspaces for researchers and students with several mini-UAS and ground stations. At the lab students from different universities (Delft, Noordelijke Hogeschool Leeuwarden) are active, in close collaboration with end-users like the Dutch police (KLPD).
An mini-UAS is the combination of a remote controlled aerial vehicle, which can lift a certain payload (a sensor suite) and a ground station to control the vehicle and process the information from the sensor suite. Typically in the UAS lab the Pelican Quadrocopter is used, which can carry quite advanced sensors and has a quite extensive development base.
Part of the assignment will be to develop an algorithm to enhance the situation awareness of the operator by fusing the information of advanced sensors available for this platform; 3D camera’s (e.g the Point Grey Research Bumbelbee), time of flight cameras (Sick Ranger), surround vision camera’s (Point Grey Research Ladybug) and structured light 3D sensors (the Microsoft Kinect, the Mantis Vision F5).
Local contact is Gerald Poppinga, academic supervisor Arnoud Visser. -
Evaluation of Human Robot Interaction in Urban Search and Rescue
The European project NIFTi is focussed to improve robotic support in Urban Search and Rescue. Previous research has indicated that the situation assessment of the operator is really difficult and the stress-level very high. The NIFTi project tries to measure and improve the interaction between the operator and the robot in several ways.
It is the task of the student to prepare such Urban Search and Rescue scenario in simulation (based on the Unreal 3 Engine) and to set up an psychological experiment with several volunteers. The result should be that after the experiment it can be quantified how well the situation assessment of the volunteers was. The result of this experiments should be analysed and reported.References:Location:TNO SoesterbergContact: -
Bionics at the German Aerospace Center
At the German Aerospace Center several interesting robotic projects are available.A complete list is available at http://www.robotic.dlr.de/jobs/
.Our local contact is Dr. Patrick van der Smagt, Head of terrestrial robotics and bionics group.
If you are interested in doing research in Wessling, Germany, please contact Dr. Arnoud Visser.
-
Mobile Recommenders: Extending the PRIMER System
Attached Files: 
msc_projects_20102011.pdf (155.785 KB)
Abstract:
In this project, you will build on current research aimed at developing a mobile recommender system that can find, personalize, and present routes, places, and possibly itineraries to pedestrians exploring a city (like Amsterdam or Barcelona). The vision behind this is that people, when visiting a new city or exploring a familiar one, would sometimes like to get away from the tyranny of tour guides (e.g., Lonely Planet) and instead get a social and more local interpretation of the places and paths in the city.
Taking first steps towards realizing the vision behind the PRIMER system, a basic system that makes use of FlickR data (users, photos, tags, timestamps) has been developed that creates and presents popular routes on a given day using sequence alignment methods. Currently the system is being extended to provide personalized recommendations of places, using content-based collaboration approaches. There are a number of AI research areas that can significantly expand on this work:- Prediction & Forecast
- Social Network Analysis
- Group Modelling
Supervision : Frank Nack and Abdo El Ali
Contact information:
Abdo El Ali
e: a.elali@uva.nl
t: 020 525 8661
w: http://staff.science.uva.nl/~elali/ -
Masterclass MatchMakeRS
Attached Files: - Masterclass matchmakers.pdf (96.538 KB)
Matching people by Multi camera Remote Sensors
Summary
This is a MasterClass Research Project and will run from February until June 2011. The project will provide MasterClass Lectures, Workshops and company visits as well as a dedicated workspace where you will work together as a team.
The project offers room for 5 MSc Artificial Intelligence, Computer Science or Informatiekunde research students.
-
Real-time classification of rodent behavior
Attached Files: - Rodent Behavior Recognition.pdf (87.159 KB)
State-of-the-art in behavior recognition is the detection of behavior of humans, mostly based on spatial measurements like pose and speed. Unfortunately these techniques are not suitable for the detection of subtle rodent behaviors like grooming, sniffing and eating. A few systems have been described in literature that can recognize rodent behavior from a side view. In a recent study features are generated based on a computational model of motion processing in the human brain. Classification of these features and their temporal context is done using advanced event recognition techniques (HMMSVM).
This is an external project at the company Noldus Information Technology (www.noldus.com). Next to the support from the company (see attachment), also a UvA-supervisor is needed. Contact Arnoud Visser (www.science.uva.nl/~arnoud) for more information
-
3D rodent reconstruction
Attached Files: - 3D Reconstruction.pdf (108.109 KB)
The current set-up of monitoring rodents is done by a single top view infrared camera. To have a more detailed view of the rats and/or mice, a 3D reconstruction is desired which visualizes body parts such as ears, nose, paws and tail in great detail. The extension to multiple cameras offers the possibility to have these details because depth information can be added. State-of-the-art techniques in stereo algorithms use feature points to link pixels from one image to the corresponding pixels in the other image. The main challenge with rodents lies in the uniform appearance of theanimal itself and the grayscale image input.
This is an external project at the company Noldus Information Technology (www.noldus.com). Next to the support from the company (see attachment), also a UvA-supervisor is needed. Contact Arnoud Visser (www.science.uva.nl/~arnoud) for more information. -
Multiperson Tracking
Attached Files: - Multiperson tracking.pdf (120.573 KB)
The persons to be tracked are ordinary (consenting) people having their daily lunch at the Restaurant of the Future. They show a wide variety in clothing, hairstyle and/or skin color. The challenge lies in tracking and (re-)identifying persons by using multiple, uncalibrated cameras in a realtime setting. State-of-the-art methods combine person detectors and trackers in a probabilistic framework to estimate the position and identity of a certain person in the scene.
This is an external project at the company Noldus Information Technology (www.noldus.com). Next to the support from the company (see attachment), also a UvA-supervisor is needed. Contact Arnoud Visser (www.science.uva.nl/~arnoud) for more information.
-
Tracking in a Crowd
The project consists of the automatic tracking of specific persons in surveillance video of a crowd. The resolution in the video is assumed to be too low to perform face recognition and the density of the crowd is assumed to be too high to see the whole body.
Background
CCTV cameras are used for surveillance and forensics to improve public security. However, it is time consuming for a camera operator to track a person (e.g., a suspect). Much work has already been published about automatic people-tracking in simple ‘not-too-busy’ scenarios. However, only a few preliminary research papers have shown that tracking people in a crowd is becoming feasible (e.g., [Li e.a., Learning to associate..., CVPR, pp. 2953-2960, 2009]) and further research is needed before it can be applied. Automatic tracking of people in a crowd allows a more efficient tracking of people over a larger distance and in more realistic busy scenarios.
Activities
The tasks of the master student consists of (a.o.):
•&νβσπ;Literature survey which shows the current state-of-the-art in scientific publications.
•&νβσπ;Implementation of existing and new methods for tracking specific persons in a crowd.
•&νβσπ;Comparison of different methods by performing experiments.
•&νβσπ;Better performance than current state-of-the-art would be desirable.
•&νβσπ;Documentation in a M.Sc. Thesis, which contains a literature survey, description of implemented methods, experiments, results, discussion and conclusion.
Field
The project will be performed at TNO in The Hague for (at least) three days per week in conjunction with the UvA. TNO focuses on the research and development of new algorithms for image analysis, automatic scene interpretation and 3D reconstruction in the fields of defence, security, healthcare and mobility.
Profile
This is a project for a Master student in Artificial Intelligence with an interest in image analysis and pattern recognition. Experience with Matlab is expected.
Duration of the project
8 month full-time (40h/wk) or an equivalent.
Supervision
TNO: dr.ir. Henri Bouma, e-mail: henri.bouma@tno.nl
UvA: dr. Marcel Worring, e-mail: m.worring@uva.nl
-
Machine Learning for Information Extraction
Project with Philips Research
Philips Research is the source of many advanced developments in Healthcare, Lifestyle and Technology. Building on 90 years’ experience in industrial research and our world-leading patent position, we’re dedicated to meaningful innovations.
In the healthcare domain, we are enhancing imaging and monitoring systems, as well as exploring innovative personal healthcare. In lifestyle, we’re helping people see, hear, remember and share content, anywhere and anytime. Our vision focuses on simplicity, making technology an integral – but invisible – part of everyday life.
Recently, there is an increasing interest in making the contents of medical data machine understandable. It is envisioned that machines can play an important role in filtering, analyzing and data mining medical data, once it is made structured and understandable. This, in turn, will improve the quality of patient care, enhance the workflow of clinical professionals and produce data for evidence-based medicine
Computer-aided clinical decision support has received more and more attention in Philips, as it is understood that this type of technology contributes to its Sense and Simplicity corporate strategy. Clinical decision support is part of Philips’ Healthcare division. This internship will be carried out in Philips Research, which is one of world’s largest corporate research organizations. Philips Research conducts fundamental research, in collaboration with the product divisions, to leverage the competitive position of Philips as a whole.
The internship will be executed in the Health Information Management department of Philips Research and will be embedded in a Philips Research project on clinical decision support for radiology.
For more background information about Philips Research click here.
Background assignment:
This project we will focus on analyzing the contents of medical narrative reports. Reports are an important source of information in the clinical workflow. In most cases reports are the prime means of communication between clinical professionals. It is thus of eminent importance to structure their contents to enable automated extraction of relevant pieces of information and classification.
For this project we have defined a classification problem on the contents of medical narrative reports. This problem is state of the art and relates to providing easy access to the contents of reports in a Philips clinical workstation, currently developed in Philips Healthcare. Depending on the outcome of the internship, the results will be used to improve the workstation.
Assignment:
You will help to create a ground truth for the classification problem in conjunction with the Philips Research team and will apply several machine learning techniques to the classification problem. Furthermore, you are expected to write a report describing the classification problem, the results obtained and the lessons learned. Since the problem is new and clinically relevant, you will be encouraged to write up your results in a scientific publication.
Duration:
6 months. We prefer students able to start as soon as possible.
This assignment is suitable to do your thesis.
It is negotiable that the student work from home/university one or more days a week. However, due to its sensitive nature, the training data cannot leave the premises of Philips Research.
Note:
When you apply we expect to receive both Cover Letter (outlining your motivation and informing your availability) and Resume. Please also note that in order to be applicable for an Internship, it should be compulsory (outside EU/EER) by your education and you need to be registered as a student, formal documentation of which may be requested at any time.
Skills Required:
- You are currently studying towards a University Master with one or more qualifications: Computer Sciences (In particular Machine Learning, to conduct research in the area described above) and Medicine.
Note:
No medical knowledge is required for this assignment.
To be successful in this internship:
• You are strong in oral Dutch and/or English and written English
• You are accurate
• You are able to work independently
• You have a practical approach and are pragmatic
*
Philips: Merlijn Sevenster
+31 (0)40 27 49655UvA: Maarten van Someren (M.W.vanSomeren@uva.nl) and Sofia Katrenko (S.Katrenko@uva.nl)
- You are currently studying towards a University Master with one or more qualifications: Computer Sciences (In particular Machine Learning, to conduct research in the area described above) and Medicine.
-
3D model enhancement using GIS data
Building 3D models of urban areas is one of the hot topics right now at both the industrial and
the scientifi c level. The range of applications of these 3D models run from a simple Google-map
approach for visualization, to forensic investigation where 3D models are used for analysis of the crime scene. The approach to building such 3D models can be categorized in 3 areas: classical approach, high tech approach and scienti fic (or elegant) approach.
The classical approach covers techniques based on laser scanners, active stereo vision (where
markers are used or a pattern is projected into the scene). The high-tech approach is what
the industry is using. Companies such as Google or Cyclomedia spend thousands of dollars in
fancy equipment to obtain a set of geolocalized images. Later, a set of highly skilled modelers
manually build the 3D model. The elegant approach is however more scientifi c. The goal is to
use state of the art mathematical methods to build a 3D model from a simple set of images.
In this scenario no expensive equipment is required (dGPS, laser, scanners, IMU, etc). With
something as simple as a pocket camera a 3D model can be built. These models however lack
some information regarding the ground for instance and typically consist of planes or surfaces.
The goal of this MSc project is to integrate existing sparse 3D models obtained with state of
the art computer vision methods with additional information in order to complete the 3D models
and o er a more appealing and realistic look. In particular the student will have to integrate
GIS (Geographic Information System) data with 3D data.
Steps needed for this task are:- Develop the method and techniques to geolocalize the GIS data and the existing 3D model.
- Detect the ground level and roof level in the sparse 3D models.
- Use the integrated models to project aerial images into the ground.
- Use GIS data to match the sparse 3D models with buildings and fit more complex primitives (such as blocks) and project the texture.
more appealing and visually attractive 3D model. The complete process should be automatic
and computationally feasible.
If you are comfortable working with Matlab, have a strong interest in 3D modeling and want
to get your hands in one of the hot topics in computer vision, please contact us and we will
consider your application.
For more details: isaac.esteban@tno.nl Example of 3D models can be found here: - Develop the method and techniques to geolocalize the GIS data and the existing 3D model.
-
Pose estimation and tracking with range cameras
The use of cameras that give range information increases rapidly. One type of cameras is the time of flight camera (http://www.mesa-imaging.ch/), the other type is stereo cameras (http://www.ptgrey.com/products/stereo.asp).
The data generated by the camera is a set of 3D points that describe surfaces that are visible from the camera. These data are noisy or can be unreliable in other ways..
In the M.Sc project the student will develop methods to infer the pose of the human body on the basis of range data from a camera. In particular recognition of body parts (head, hands, feet) will be a central issue. If this is done with sufficient robustness, recognition can be extended to multiple persons and their body parts.
We have the availability of both a time of flight camera and a stereo camera, and thus can provide realistic data.
As a first step the data will have to be separated into background and foreground. The foreground data will then have to be clustered into different objects. As a third step is has to be determined whether it is a person and what the pose of the person is. This will be the largest challenge.
We will use probabilistic methods that are common in our group and use and extend methods that are used in regular or multi camera pose estimation. The application field is natural interaction and gaming.
Information:
Ben Kröse b.j.a.krose@uva.nl
-
Looking at humans: recognizing humans from shape
Background
When robots engage humans in natural circumstances, it is important that a robot recognizes the human in front of him. The recognition can be vision or voice based, but this assignment should concentrate on the shape of the human.
Assignment
>
>
>
Images courtesy of Aldebaran Robotics and Hokuyo Automatic Cooperation Ltd.In our laboratory is a small humanoid robot, Nao, equiped with a laserscanner on top of its head. By bending his body, the Nao can make a full 3D scan of the space in front of him. It is the task of the student to apply machine learning techniques on this point cloud and to detect if a human is front of the robot and classify this human in broad categories (adult/child, male/female).
References
- Aleksandr V. Segal, Dirk Haehnel, Sebastian Thrun, Generalized-ICP, Robotics: Science and Systems 2009 PDF.
- Rob Suikerbuik, Hans Tangelder, Hein Daanen, Aernout Oudenhuijzen, Automatic Feature Detection in 3D Human Body Scans, SAE Digital human modeling for design and engineering conference, June 15-17, 2004, Rochester. PDF
Contact:
-
Hot Topics in Video Search
For students interested in graduation projects on the frontier of video search, call me or drop me an e-mail so we can discuss the many possibilities.
Contact:
dr. Cees Snoek
Mobile: 06-24277812
Email: cgmsnoek@uva.nl
Web: http://staff.science.uva.nl/~cgmsnoek -
Probabilistic Tracking using Stereo Cameras
Short Description:In order to manage crowds, plan building infrastructure, evaluate the positioning of utilities in public spaces, etc, it is necessary to be able to track the motion of individual people over long distances and multiple cameras. Modern algorithms begin to solve this problem, but require a precise representation of the appearance of the individuals. Obtaining this is a challenging task, because it requires detecting when a person enters the field of view of a camera, identifying all pixels that correspond to that person over all frames where that person is visible, dealing with occlusions, and detecting when the person has left the field of view.
This project will use stereo cameras to capture videos of large crowds, and will use probabilistic models of (local) appearance and motion (such as linear dynamical systems, Kalman filters or particle filters) to solve the problem of association over multiple frames (which pixels belong to what person?) and of ingress and egress detection. The resulting tracks will then be used to compute global appearance features that can be used to solve the problem of association between different cameras.
Requirements:
A strong background in Machine Learning and good programming skills are essential for this project. Knowledge of C/C++ is a plus.
Contact information:
References:
-
ILPS assignments
-
Older IAS assignments
At the IAS-website, a list of existing assignments (Nov 2004 - Aug 2009) is still available. This list will become obsolete by the list on blackboard.
Contact: Arnoud Visser -
Machine Learning and Information Retrieval
Supervisor: Maarten van Someren in collaboration with Krysztian Balogh, Sophia Katrenko and Marc Bron.
Some information needs are of the form “give me all X that have property Y” or a relational counterpart “give me all X that have relation R with Y”. Often it is not hard to give a few results and the problem is to find more. For example, if we look for all “capital cities” or for all “proteins that influence the production of insulin” we can give a few examples that are well known but the problem is to find the rest. In a version of these settings we have information about the classes of objects involved. For example, we may have a list of cities, or of proteins and want to know which of these satisfy the query. In information retrieval this is called “entity search”
One approach to this problem is based on co-occurrence. We look for words or elements of our list that frequently co-occur with “capital” or “insulin”. If we search a large collection of documents then this is a simple and powerful way to answer such queries.
A more sophisticated way is to use the context of terms. We take the few examples that we have, find where they occur in documents and collect the contexts in which they occur, and perhaps a few more features (e.g. word morphology). From this we induce patterns (using Mchine Learning) and then we use these patterns to find new results. There are several possible schemes for this. This second approach is very expensive.
In a Bachelor project (see Overgoor, 2010) a first comparison was made between the two approaches that shows that they are indeed not perfectly correlated. Something can be gained by combining them.
The goal of this project is to unify these two approaches. If data are plenty then approach one is probably cheap and effective but for less frequent categories the second approach may be better.
This project builds on methods and software developed in previous projects by Krizstian Balogh, Sophia Katrenko, Viktor de Boer. References are below. Beside the datasets used in the papers below, there is a possibility of using TREC data
Viktor de Boer, Maarten van Someren, Bob J. Wielinga: A redundancy-based method for the extraction of relation instances from the Web. International Journal of Man-Machine Studies 65(9): 816-831 (2007)
Sophia Katrenko, Pieter W. Adriaans, Maarten van Someren: Using Local Alignments for Relation Recognition. J. Artif. Intell. Res. (JAIR) 38: 1-48 (2010)
Jan Overgoor (2010) The value of learning linguistic structures in relation extraction, Bachelor Thesis, UvA (http://staff.science.uva.nl/~bredeweg/pdf/BSc/20092010/Overgoor.pdf)
Krisztian Balog, Edgar Meij, Maarten de Rijke (2010) Entity Search: Building Bridges between Two Worlds, Semantic Search 2010.
-
External projects
There are possibilities for external (outside UvA) projects. Many people at UvA have contacts with companies and institutions where you may do a thesis project. Some such contacts of myself are for example:
Textkernel (www.textkernel.com)
Sentient group (www.smr.nl/)
Xerox (www.xrce.xerox.com/About-XRCE/Internships)
D-CIS (www.d-cis.nl/career/internship)
Lely (http://jobs.lely.com/en/students)
Computer Science Department, University of Rome "Sapienza", Italy
- maarten van someren -
Semi-Supervised Learning for recognizing bird behaviour
Supervisor: Maarten van Someren (in collaboration with Willem Bouten, IBED)
Semi-supervised learning means that a system learns a classifier from both labeled and unlabeled data. In one project we are developing new algorithms for semi-supervised learning. In another project we used data that are measured by sensors attached to birds to recognize what the bird is doing. It is hard to obtain sensor data that are labelled behaviour categories but there are many unlabeled data. This therefore looks like an excellent setting to test our semi-supervised methods in practice. Such experiments always give inspiration for improvements and this can make a good thesis.
