Sherif Abdulatif

Signal Enhancement via Generative Models 

My research is focused on using generative models in signal enhancement. More specifically generative adversarial networks (GANs) operating on signals in different domains either raw time-domain signals or in time-frequency (TF) representation or combination of both with a main interest in:

• Speech enhancement using GANs: in a realistic scenario, an audio signal is a mixture between the desired signal and an unwanted interference or background noise. Audio enhancement can be analyzed as a source separation problem between the desired audio signal and the unwanted intrusive effect. For instance, speech denoising (eliminating background noise from the desired speech) is a vital building block for automatic speech recognition (ASR) systems, hearing aids and teleconferencing systems. Some common background noise types (cafe or food court) can be very similar to the desired speech. In these cases, estimating the desired speech from the corrupted signal is challenging and sometimes impossible in low SNR situations. 
  My research is focused on introducing new generative models for denoising of speech signals by operating on raw time domain or TF representations of noisy speech inputs.
• Radar signal enhancement and super-resolution: radar sensors are widely used nowadays in the tracking and identification of different targets. This is due to its superior penetration capabilities and robustness against different lighting and weather conditions in comparison to vision-based approaches such as cameras and LIDAR. Additionally, radar is nowadays the dominant sensor used for autonomous driving as it combines high resolution in range, velocity and depth perception. This motivated the use of radar in vital tasks such as on-road pedestrians and cyclists recognition. However, most of the research assumes perfect measurement conditions where an ideal obstacle-free environment is preserved. Such requirements are not always applicable in many realistic use cases. Moreover, radar suffers from stringent constraints on its design parametrization leading to multiple trade-offs in radar resolution. 
  My research is focused on using GANs for denoising and super-resolution of radar TF representations.

Radar Micro-Doppler for Gait Recognition, Safety and Surveillance

This research is a cooperation between ISS and Fraunhofer IPA. Micro-Doppler (µ-D) spectrum introduces an additional capability in radar systems by allowing the analysis of the individual micro motions of moving targets. For instance, when studying the µ-D signature of a non-rigid body such as human motion, some frequencies reflect the translational velocity of the human and other frequencies reflect the swinging of different body limbs. 
My research is focused on using different learning-based techniques operating on the human µ-D signature for:

• Gait analysis and limbs decomposition
• Identifying and tracking humans from different moving targets
• Identification and recognition of different human subjects based on their respective motion signature