Fachbereich Informatik

64-476 Oberseminar Technische Aspekte Multimodaler Systeme

Raum F-334
Zeit Dienstag 16 - 18 Uhr (c.t.)
Veranstalter Jianwei Zhang


07.04. Vorbesprechung
Vortragender: Lasse Einig
Zusammenfassung: Master Thesis - Oral Defense

Hierarchical Plan Generation and Selection for Shortest Plans based on Experienced Execution Duration

This thesis presents the Plan Evaluator, a method to find the optimal plan, executed in parallel, from a set of sequential HTN generated plans for reaching a goal state from an initial state. Plan decisions made by HTN planners are either based on counting the amount of steps within a plan or by statically assigned weights to the plan steps. The Plan Evaluator uses dynamically generated task execution durations, based on experience from experiments or simulations, to calculate the shortest plan considering the total plan execution duration. In addition, the plan is arranged in a parallel executable order and the benefit from this parallel execution is incorporated to the decision made by the Plan Evaluator. The proposed method is evaluated with two scenarios and the results are verified with additional scenarios. The results show that for any scenario a configuration exists at which the Plan Evaluator will find a plan which is executed faster or as fast as the original decision made by the HTN planner.
Folien Slides 1.0 M pdf

17.04. Active Constraint for a Robotic Spine Surgery System
Vortragender: Haiyang Jin

21.04. Netzwerkanbindung von einem FPGA an einen PC
Vortragender: Alexander Krichewski

28.04. Visual-Audio Object Recognition using Hidden Markov Models
Vortragender: Weipeng He

Comparing to traditional visual object recognition, multimodal object recognition is advantageous in that different modalities provide complementary information. This work aims to implement a system for object recognition given videos of interactions with objects and investigate different modality fusion methods.

The bag-of-words model with SIFT descriptors and the MFCC are used as visual and audio features. The system classify objects by computing the probability with learned hidden Markov models. The system incorporates two different fusion methods: feature level fusion and decision level fusion. The former method learns a joint probability distributions with one HMM, while the latter method learns two separate probability distribution with two HMMs and combine them under the conditional independence assumption.

Experiments based on a dataset of 33 different household objects are carried out to evaluate the performance of these two fusion methods as well as single modality approaches. The result shows that both fusion methods improved the performance over single modality methods, while these two methods are mostly comparable.

05.05. Progress report: 3D-printed circuits
Vortragender: Florens Wasserfall
Folien Slides pdf

12.05. Previous Research Project Outline
Vortragender: Jinpeng Mi
Folien ....pdf 1.0 M pdf

19.05. Adaptive Gesture Recognition System Integrating Multiple Inputs
Raumänderung R133
Vortragender: Tobias Staron
Abstract: Master Thesis - Oral Defense
Although the ability to recognize gestures is an obvious feature for robots, there are not many gesture recognition systems that actually exploit the possibilities offered by robotics. This work makes use of the possibility that a robot can have more than one sensor and presents a system able to process multiple inputs, depth images and information about the users skeletons provided by the OpenNI tracker in this case, by applying a sensor fusion method. Furthermore, the possibility of interactions between a robot and its users might lead to feedback provided by them. This way the system can learn online and adapt its internal models, for example to a changed environment or new users. The tests will show that multiple inputs as well as adaptivity leads to performance improvements. The best gesture recognition results can be achieved by the primary contribution of this work, a system that combines multiple inputs one the one hand and adaptivity on the other hand.
Folien kolloquium_new.pdf 1.0 M pdf

02.06. canceled, moved to 16.06

09.06. tbd
Folien ....pdf 1.0 M pdf

16.06. Entwurf und Implementierung einer Multitasking-Umgebung für den Arduino Due
Vortragender: Enno Köster
Zusammenfassung: Seminarvortrag zur Bachelorarbeit

Folien ....pdf 1.0 M pdf

30.06. Accessibility on the Android Platform: Subtitles on Media Content
Vortragender: Gökhan Kalender
Zusammenfassung: Seminarvortrag zur Bachelorarbeit

Folien ....pdf 1.0 M pdf

07.07. Waveform-Viewer App for Android
Vortragender: Lars Lütcke
Zusammenfassung: Seminarvortrag zur Bachelorarbeit

Folien ....pdf 1.0 M pdf