Tim Sherry

Blind Source Separation of Speech

I developed a Real Time Blind Source Separation System for GN Resound. This system would separate up to 10 Voices speaking over each other within a reverberant environment (4mx4m room). It used frequency domain Independent Component Analysis (ICA), along with a backprojection algorithm to obtain an average +17dB improvement in SNR when compared to the inital recordings. Practially tested (the audio clips are from a real world test), with an end-end delay of ~100ms. developed in Matlab, ported to C

Close Window

Photohigher

FlyingFox cable cam from Photo Higher on Vimeo.

I developed the electronics and control system for the FlyingFox cable cam. The motor control, positional sensing, control input, digital LCD display and User interface were all included in the functional prototype I delivered at the end of my time at Photohigher.

Halo 2000 on Test from Photo Higher on Vimeo.

I contributed to the Halo series of high end gimbals. I developed the motor control signalling interface, which ensured the control messages got to the intended maxxon motor driver. I worked on the qarternion update to the gimbal stabilisation system alongside a senior engineer.

Close Window

EEG Source Localisation

adaptiveSurface from Tim Sherry on Vimeo.

EEG Source localisation is a powerful tool for diagnosis of a variety of neurological conditions. However, to do it right, an anatomically accurate conduction model of the subject is key, and the gold standard is based off MRI imagery. However, obtaining an accurate model for an infant is challenging, firstly because critical anatomical features, for example the skull are often not resolved in the image due to how thin they are. Secondly infants are prone to wriggling about, introducing movement artifacts and noise.
My work takes contemporary MRI segmentation algorithms and adds anatomical constraints around feature thickness to ensure models maintain anatomical accuracy even under adverse imaging conditions.

Close Window

Automated Seizure Detection Via Deep Learning

Seizure burden is a strong marker for adverse neurological outcomes. It is a powerful tool in quantifying injury during acute neurological trauma. However, accurately classifying seizure burden is a very time intesive task. An EEG expert needs to view the entirety of the EEG record to identify all seziures that are present, which takes roughly 1/3 of the original recording time . Recordings in clinical trails can exceed 72hrs, a huge burden on the scientists doing the markup. I have developed an SVM classification system which obtains 98.6% accuracy. This is in line with modern alogrithms. I have also developed a CNN classification system with similar performance. I am investigating the potential for LSTM's to further improve on the performance.

Close Window

RC Aircraft

I have built a number of fixed wing RC aircraft and multirotor drones.

Close Window

Composites work/Sailing

I have spent a significant amount of time working with composite materials. Most recently I have repaired a carbon fibre mountain bike after a major crash. This required significant work on the down tube and seat stay to get back to working condition.
I was head of maintennance at UQSail, which involved keeping the 9 Sailing boats and 1 Powerboat operational. During this time there were about 10 sigficant holes in hulls repaired, repairs to alloy masts, and a rebuild of the mast base area after rot had destroyed it. I directed a team of volunteers in this work, which kept the boats servicable and saved the club around $10,000, allowing the club to operate through financial difficulties.
I have also worked on my own sailing skiffs, as well as sailing them regularly.

Close Window

Ship Motion System

This project involved building a Data Acquisiton System using the NI Compact RIO platform, building strain + accelerometer sensor nodes and installing the whole system on a bulk tanker. In addition I have developed a bending moment estimation algorithm, which takes the strain gauge sensor data, fits that data to the FEM analysis from the ship design team, and estimates the first order bending moment experienced by the hull. This provides real world fatigue data for the owners asset management team.

Close Window

Classifynder automated miroscope system software update

The Classifyder automates the classification of objects of various forms on a microscope slide by taking high resolution 3D images of the objects and passing them through a proprietary classification algorithm.
I was tasked with reworking the software to make it more robust and ready it for coming hardware updates, and delivering improvements to usability for new commercial customers.

Close Window

Eight360 : Nova VR Motion Platform

Eight360 Nova (v2.5) VR Motion Simulator from Eight360 on Vimeo.

I worked with Eight360 to upgrade the Internal MEMS IMU for the NOVA VR motion simulation platform

Close Window