Current Research
My research focuses these days on self-supervised and multi-modal machine learning techniques for vision+radar perception, including representation learning.
I continue to occasionally contribute to wireless problems, particularly those intersecting with machine learning.
I have led a programme (over a period of 4 years) that envisioned and delivered a self-supervised learning algorithm for automotive radar perception.
Our algorithm uses contrastive learning and domain-specific radar augmentations to provide substantial improvements over supervised learning: 70% in label efficiency and 5.8% in mAP.
Performance characterised in the real-world under nontrivial MIMO radar artefacts.
|
|
Bootstrapping Autonomous Driving Radars with Self-Supervised Learning
Yiduo Hao, Sohrab Madani, Junfeng Guan, Mohammed Alloulah, Saurabh Gupta, Haitham Hassanieh
IEEE/CVF CVPR, 2024  
arXiv
Bounding box estimation from radar-only for autonomous perception by distillation from vision using contrastive learning and domain-specific radar augmentations.
|
|
Look, Radiate, and Learn: Self-Supervised Localisation via Radio-Visual Correspondence
Mohammed Alloulah, Maximilian Arnold
IEEE/CVF CVPR, 2023  
Project page
/
arXiv
/
Slides
Automatic localisation of objects in the environment by tapping into the cross-modal information captured simultaneously by vision and radar.
For our synthetic dataset MaxRay, contact my ex-employer Nokia Bell Labs.
|
|
Self-Supervised Radio-Visual Representation Learning for 6G Sensing
Mohammed Alloulah, Akash Deep Singh, Maximilian Arnold
IEEE ICC, 2022  
arXiv
Building a radar classification model by distillation from vision using contrastive learning.
|
|
Future Millimeter-Wave Indoor Systems: A Blueprint for Joint Communication and Sensing
Mohammed Alloulah, Howard Huang
IEEE Computer, 2019  
PDF
Position paper on why and how should we care to build a unified communication-perception network.
|
|
Benchmarking Learnt Radio Localisation under Distribution Shift
Maximilian Arnold, Mohammed Alloulah
arXiv, 2022  
arXiv
Benchmarking 8 flavours of real-world learnt RF localisation across three performance axes: (i) learning, (ii) proneness to distribution shift, and (iii) localisation, as well as against classic maximum likelihood estimation.
|
|
IMULet: A Cloudlet for Inertial Tracking
Mohammed Alloulah, Lauri Tuominen
ACM HotMobile, 2021  
PDF
/
Slides
Practical inertial navigation system using deep learning and cloud-device partitioning.
|
|
Deep inertial navigation using continuous domain adaptation and optimal transport
Mohammed Alloulah, Maximilian Arnold, Anton Isopoussu
arXiv, 2021  
arXiv
In-situ adaptation of deep inertial navigation through spatial diversity, optimal transport, and domain adaptation.
|
|
On goodness of WiFi based monitoring of sleep vital signs in the wild
Kamran Ali, Mohammed Alloulah, Fahim Kawsar, Alex X Liu
IEEE TMC, 21  
DOI
Measuring breathing and body motion markers wirelessly towards sleep monitoring.
|
|
ePerceptive: energy reactive embedded intelligence for batteryless sensors
Alessandro Montanari, Manuja Sharma, Dainius Jenkus, Mohammed Alloulah, Lorena Qendro, Fahim Kawsar
ACM SenSys, 2020  
DOI
Best-effort inferencing using a variable instantaneous energy budget by means of knobs at featurisation-level as well as at neural network-level.
|
|
KinPhy: a kinetic in-band channel for millimetre-wave networks
Mohammed Alloulah, Zoran Radivojevic, René Mayrhofer, Howard Huang
ACM SenSys, 2019  
PDF
/
Slides
How would future networks in mm-wave bands exchange information using vibrational (i.e. kinetic) modulation in order to support security and spontaneous interaction applications.
|
|
Degradable inference for energy autonomous vision applications
Alessandro Montanari, Mohammed Alloulah, Fahim Kawsar
ACM UbiComp/ISWC, 2019  
DOI
Degradable vision inference by means of knobs at featurisation- and neural network-level.
|
|
On Indoor Human Sensing Using Commodity Radar
Mohammed Alloulah, Anton Isopoussu, Fahim Kawsar
ACM UbiComp/ISWC, 2018  
DOI
Thoughts and guidelines on how to build sensing models for commodity radar HW indoors.
|
|
A Cross-Layer Coding for Scalable ECG streaming
Mohammed Alloulah, Mark Dawkins, Alison Burdett
EAI MobiCASE, 2016  
PDF
/
Slides
Joint compression and wireless coding scheme for robust and ultra-low power medical physiological sensing.
|
|
An efficient CDMA core for indoor acoustic position sensing
Mohammed Alloulah, Mike Hazas
IEEE IPIN, 2010  
DOI
Indoor acoustic localisation chip. Part of two further technical reports here and here.
|
|
Ultra-Low Power WiFi Modulator
Traditional direct conversion RF transmitters are not suitable for IoT devices with stringent power budget. A polar transmitter in wideband WiFi channels is challenging. We made it work with couple key innovations. Achieved 3x power reduction.
Silicon + several patents.
|
|
OFDM PHY Chip
Led work on multiple architecture-aware PHY implementations, part of a terrestrial DVB-T2 chipset. Keywords included inter-symbol interference (ISI) and inter-carrier interference (ICI).
Silicon in the field + several patents.
|
|