Research

AccCall leverages the smartphone's built-in accelerometer to monitor and enhance real-time phone call quality. By detecting device orientation, user motion, and acoustic coupling conditions, the system dynamically adjusts audio processing parameters to improve speech clarity and reduce noise during calls without additional hardware.

MDLdroidLite proposes a release-and-inhibit control mechanism that dynamically allocates computational resources for deep neural networks on mobile devices. The system selectively activates and inhibits neural network layers based on real-time resource availability, achieving significant energy savings while maintaining inference accuracy.

MDLdroid presents ChainSGD-Reduce, a distributed deep learning framework that enables efficient on-device training across a chain of smartphones. The approach reduces communication overhead and memory requirements by propagating gradient updates sequentially through the device chain, enabling collaborative model training without a central server.

mmFER presents the first system to perform facial expression recognition using millimetre-wave radar, enabling privacy-preserving, contact-free emotion sensing. The system captures subtle facial muscle movements from reflected mmWave signals and classifies them into seven universal expression categories with high accuracy, even through occlusions.

mmBP introduces a non-contact blood pressure measurement system using commodity mmWave radar. By capturing pulse wave velocity and arterial stiffness indicators from radar reflections on the wrist area, mmBP estimates systolic and diastolic blood pressure without any physical contact, opening new possibilities for continuous health monitoring.

Conference version of MDLdroidLite, presenting the release-and-inhibit control mechanism for resource-efficient deep neural network execution on mobile devices.

Conference version of MDLdroid, introducing the ChainSGD-Reduce approach for distributed mobile deep learning across a chain of smartphones.