UnivIS
Informationssystem der Friedrich-Alexander-Universität Erlangen-Nürnberg © Config eG 
FAU Logo
  Sammlung/Stundenplan    Modulbelegung Home  |  Rechtliches  |  Kontakt  |  Hilfe    
Suche:       
 Lehr-
veranstaltungen
   Personen/
Einrichtungen
   Räume   Forschungs-
bericht
   Publi-
kationen
   Internat.
Kontakte
   Examens-
arbeiten
   Telefon &
E-Mail
 
 
 Darstellung
 
Druckansicht

 
 
 Außerdem im UnivIS
 
Vorlesungs- und Modulverzeichnis nach Studiengängen

Vorlesungsverzeichnis

 
 
Veranstaltungskalender

Stellenangebote

Möbel-/Rechnerbörse

 
 
Einrichtungen >> Technische Fakultät (TF) >> Department Informatik (INF) >> Lehrstuhl für Informatik 2 (Programmiersysteme) >>

Deep Kalman Filtering to Reduce Error Variance in V2X Pose Forecasting

Art der Arbeit:
Master Thesis
Betreuer:
Feigl, Tobias
Lehrstuhl für Informatik 2 (Programmiersysteme)
E-Mail: tobias.feigl@fau.de

Philippsen, Michael
Lehrstuhl für Informatik 2 (Programmiersysteme)
Telefon +49-9131-85-27625, Fax +49-9131-85-28809, E-Mail: michael.philippsen@fau.de

Beschreibung der Arbeit:
Motivation. Vehicle to Everything (V2X) communication and 5G connectivity in vehicles enable advanced driver assistance systems (ADAS), collaborative perception, and vehicle platooning. V2X efficiently distributes object data, intentions, and route predictions and recognizes vehicles in situations without direct visual contact, so that future manual or autonomous vehicles can cope with traffic and roads more efficiently and safely. The basis of V2X, the 5G network, enables real-time applications that require high speeds, high availability, low latency, cloud computing, and artificial intelligence. In the cloud, enormous amounts of data are processed by many different sensors of vehicles, infrastructure, and pedestrians, which are shared by the participants, especially the vehicles. Computationally intensive learning algorithms can use this data and pass aggregated results on to the participants in almost real time.

Related work. The trajectory estimation for tasks such as object and vehicle location and tracking is generally noisy, and temporal filters have been used extensively for regularization. One of the most widely used methods is the Kalman filter (KF), which is both extremely simple and general. However, KFs require a priori the specification of a motion and measurement model, which burdens the modeler and at the same time requires the use of explicit models that are often only rough approximations of reality. E.g., in the position estimation task, it is common to use motion models that assume constant acceleration or velocity, and these simplified representations are known to be highly limiting. And similar to prominent models like ARIMA, KF only relies on the last past sample to predict the next future pose. However, the future movement is not only influenced by the last movement step, but long-term dependencies often form the shape of a trajectory and so affect the future pose.

Overall goal. Thus, in this thesis, the student will investigate a hybrid method that combines KF with neural networks that learn rich, dynamic representations of the motion and noise models instead. In particular, the student will learn these models from data using, e.g., long short-term memory, which enables representations that are dependent on all previous observations and all previous states. The student will evaluate the method using the latest trajectory prediction methods. The student will further examine how the number of historical samples, their intra-delay, changes and gaps in the sample rate and input stream affect the robustness and accuracy of the predictors. In a final step, the student will implement the model in a real demonstrator that uses, e.g., Time-of-Arrival (ToA) values or positions as input to the model. The student will show that the model learns easily from noisy labels and can be adapted to new environments or different vehicles or objects with very few samples.

Timetable (6 months, in person weeks [PW]):

  • [4PW] Literature and patent research; Familiarization with relevant work on the subject areas.

  • [4PW] Adaptation of the individual components.

  • [6PW] Methodological work: Data-driven fusion, learning process for complementing the absolute and relative components with RNN and KF.

  • [4PW] Evaluation and real-world demonstration.

  • [6PW] Transcript.

Expected results and scientific contributions:

  • Modular processing pipeline that consists of, e.g., an Extended KF plus deep (recurrent) neural networks that replace the measurement, process, and transition matrices. Based on the work of Becker et al. [1-5]

  • A localization pipeline that is real-world applicable and replaces state-of-the-art KF [2]. The real-world applicability is given by a live demonstrator.

  • The sensor information streams should consists of data such as ToA and position.

  • ToA or position values will be collected for different motion dynamics and different propagation scenarios.

  • The pipeline will be implemented in python with support from pytorch scipy, and scikit.learn.

  • KF + LSTM | HCNN (motion model), i.e., deep bayes, reduce error variance, share some light on uncertainty, delay and gap filling.

  • Uncertainty will be estimated implicitly by the KF and explicitly by the data-driven approach with Monte Carlo dropout, ensembles, Bayesian inference, SWAG or Laplace [1,4].

  • Information fusion architecture: AI-based multihypothese filter, i.e., when to fuse what and how, and with which architecture.

  • Fusion evaluation: classic (KF as is) vs. mixed (KF+AI) vs. "end-to-end" AI. [1]

Literature:
[1] Philipp Becker, Harit Pandya, Gregor Gebhardt, Cheng Zhao, James Taylor, Gerhard Neumann. (2019). Recurrent Kalman Networks: Factorized Inference in High-Dimensional Deep Feature Spaces. arXiv:1905.07357 [cs.LG]
[2] Zhuangwei Shi. (2021). Incorporating Transformer and LSTM to Kalman Filter with EM algorithm for state estimation. arXiv:2105.00250 [cs.LG]
[3] Scott Sun, Dennis Melamed, Kris Kitani. (2021). IDOL: Inertial Deep Orientation-Estimation and Localization. arXiv:2102.04024 [cs.RO]
[4] Shirli Di-Castro Shashua, Shie Mannor. (2021). Kalman meets Bellman: Improving Policy Evaluation through Value Tracking. arXiv:2002.07171 [cs.LG]
[5] Changhao Chen, Chris Xiaoxuan Lu, Bing Wang, Niki Trigoni, Andrew Markham. (2021). DynaNet: Neural Kalman Dynamical Model for Motion Estimation and Prediction. arXiv:1908.03918 [cs.LG]

Organizational matters:

  • This work is being carried out in cooperation with the Fraunhofer Institute for Integrated Circuits (Nuremberg, Department of Hybrid Positioning and Information Fusion) and Elektrobit Automotive GmbH (Department of Digital Sensor Fusion and Artificial Intelligence)

  • You may also have the opportunity to develop your previous knowledge as part of a HiWi job in preparation for this thesis.

  • The work also includes a lecture in the chair's colloquium at the end of the processing time.

  • If you are interested, I am available for a (personal) conversation to define the exact tasks together.

Schlagwörter:
Kalman Filter, Recurrent Neural Network, Long-Short-Term Memory, Historical Consistent Neural Networks, Vehicle2X, Machine Learning, Deep Learning, Classification, Künstliche Intelligenz, Artificial Intelligence, Maschinelles Lernen, Tiefes Lernen, ML
Bearbeitungszustand:
Die Arbeit ist noch offen.

UnivIS ist ein Produkt der Config eG, Buckenhof