Annotation tool for object detection and tracking applications

(Bachelorarbeit)

Many recent object detection systems rely on supervised learning algorithms. Those require manually annotated training data to adapt a general model for an object class.
Moreover, ground truth labels are not only valuable while training a detection system, but also to determine a system's performance at runtime.
For videos, single detections are often fused across time and with other sensor's output by means of tracking. For a proper evaluation of those algorithms, not only a label of the object's position in a single frame is required, but also the association of the same object in subsequent frames.

The project goal for this bachelor thesis is to extend an existing annotation tool with the capabilities to annotate video sequences for tracking applications and multiple object classes.

Environment

Published 25.05.2007, Christian Wojek

Long-Term, Fine-Grained Actigraphy

Actigraphy is a method of activity and sleep study, achieved by strapping a small watch-like unit on a person for an extended period of time. The unit continually records the movements it undergoes, and when the data is later read to a computer, it can be analysed for interesting patterns. This project aims at making a 'smarter' actigraphy unit that not just logs the sensor data, but does a bit of the recognition itself already...

In this thesis, the student is actively involved in a development and research cycle which aims at making a module that is as small and light as possible, and can be worn for long periods (days to weeks). This includes a large amount of embedded programming, learning about basic signal processing and classification algorithms, and a bit of elementary hardware design (including microcontroller interfaces, taking measurements and some surface-mount soldering).

The student will need to evaluate (and wear) his/her own self­built unit for extended periods, and, if the final  work is  qualitative  enough,  scientific  publication  of  the  student's  research will  be encouraged,  as well as practical usage of the actigraphy module by others.

Our main interest in this project is on the algorithms and research questions,  but the unit's purpose is to be applied to problems in circadian rhythm analysis, wake-sleep patterns monitoring, and psychiatric trials with bipolar patients.

• be motivated to learn, and have a practical mindset
• in possession of creative thinking skills, not afraid to touch hardware and electronics
  
• know, or have an interest, in embedded sensor signal analysis and machine learning
• be eager to become proficient in C and C++
  

Motion Features for Object and Pedestrian Detection in Image Sequences

(Diplomarbeit, Master Thesis, Studienarbeit)

Recognition of objects and people has become a popular research topic in recent years. In this context pedestrians pose a number of particular problems. Not only may pedestrians have a wide range of possible articulations, moreover different clothing changes their appearance considerably. Additional challenges result from occlusions and varying lighting conditions.

The goal of this master thesis is to extend the object features, used within two pedestrian detection systems for still images, with motion information. Next to the implementation of the features a thorough evaluation is to be performed to demonstrate and analyze the influence of the motion information.

Applicants should enjoy working on state-of-the-art computer vision problems and have to become acquainted to the C++ source code of the systems.

Environment

Published 08.01.2007, Edgar Seemann