Skip to content

ESR2

Intelligent Ultra Low-Power Signal Processing for Automotive

There has been a long history of solving Sound Source Localization (SSL) and Sound Event Detection (SED) problems with both conventional signal processing methods, including the beamformer-based search, subspace methods, probabilistic generative mixture models, and independent component analysis. Recently, the advent of modern deep neural networks has shed light on building multiple-purpose model in this field. Since 2015, the number of DNN models for SSL is explosively increasing, covering all major types of network layer types, such as the Multi-Layer Perceptron (MLP), convolutional neural network (CNN), convolutional recurrent neural network (CRNN), encoder-decoder neural network, attention-based neural network, and etc.

While most research in this field focuses on indoor scenarios, the I-SPOT project aims to provide solutions for outdoor automotive. The ESR project targets at the development of low-power real-time processors with the capability to:

  • Execute the Sound Event Detection and Localization algorithms with low hardware overhead
  • Support multiple-mode algorithms with optimal hardware efficiency
  • Enable programmability for algorithm updates via agile hardware architecture

Research Team
Early Stage Researcher: Jun Yin
Host Institution: KU Leuven
Supervisors: Marian Verhelst, Andre Guntoro