Research
AUTONOMOUS Driving
The NVIDA demo course at CES 2017
The NVIDIA PilotNet group has been working on the PilotNet lane-keeping effort for over the past five years. The PilotNet core is a multi-layer neural network that generates lane boundaries and the desired trajectory for a self-driving vehicle. PilotNet works downstream of systems that gather and preprocess live video of the road. A separate control system, when fed trajectories from PilotNet, can steer a vehicle. All systems run on the NVIDIA DRIVE™ AGX platform.
Internship Model Demo.MP4A Computer Vision Based Autonomous Driving Training Scheme through Selected Enriched Training Dataset
During my internship with Autonomous Driving team at NVIDIA in 2017, I trained a deep convolutional neural network (CNN) model to achieve car lane keep, lane merge and lane split using end-to-end learning approach. The model can also adapt reasonably well on a local road. Further, I developed a mechanism to improve the network performance by diversifying the training dataset and iteratively training the network with new but tough dataset.
An Vital Sign Monitoring System Through Physical Vibrations
HB-Phone: a bed-mounted geophone-based heartbeat monitoring system [PDF]
HB-Phone is an accurate, low-cost and easy-to-use bed mounted heartbeat monitoring system centered around a single commercial off-the-shelf analog geophone. The system senses the weak physical vibrations caused by heartbeats together with varies noises from the ambient environment. Then, it extracts heartbeat pulses from the noisy environment using an autocorrelation-based algorithm, by taking advantages that heartbeats are quasi-periodic but noises are not.
VitalMon: Monitoring a Person's Heart Rate and Respiratory Rate During Sleep on a Shared Bed Using Geophones [PDF]
VitalMon is a non-invasive heart-rate and respiratory-rate monitoring system which simultaneously tracks multiple people's heartbeats and respiration using two analog geophones mounted on a single bed. The challenges here are signals from multiple sources (people) are mixed together and geophones are insensitive to respiration vibrations. To enable vital signs tracking, the system first separates vibration signals from different sources based on the unique spatial signatures of the sources. Later, it segregates heartbeat signal and respiration signal using an amplitude demodulation algorithm, due to our observation that heartbeat and respiration vibrations are naturally modulated in human body.
BabyMon: Evaluating a Baby's Sleep Quality via Crib Ambient Vibrations Using Geophones
BabyMon aims to evaluate a baby's sleep quality via the vibration approach. The system perceives high fidelity vibration data and evaluates a baby's sleep quality from multiple physiology signs (heartbeats, respiration, body motions, sleep talking, etc.).
A Context Aware Monitoring System for sitting behavior study
A Capacitive Sensor Based Smart Chair for Understanding User’s Sitting Behavior
Improper sitting and prolonged sitting have grown to be a larger concern in the medical field. Our Touch-chair aims to learn a user's sitting state and context as much as possible. Especially, it focus on the person's identity, sitting postures, respiratory, and skin moisture level. The system utilizes extremely sensitive capacitive sensors to measure the capacitance changes due to micro variation caused by a user's posture, respiration, vapor pressure from the skin. By adapting a series of signal processing and machine learning techniques, the system can accurately classify the user's identy and sitting posture, as well as monitoring the user's respiratory rate and skin moisture level.
A Low Power Air quality monitoring System
Continuous Low-Power Ammonia Monitoring Using Long Short-Term Memory Neural Networks [PDF]
Environmental factors, especially ammonia concentration in a cage, greatly impact the reproducibility of rodent experiments. In this project, we propose an ammonia sensing system which is automated, ultra low-power and highly reproducible at the same time. The system is centered around an metal-oxide sensor which has reversible chemical reactions with ammonia in the air. We discover that the transient signature of the ammonia measurement may strongly correlated with the final concentration of the ammonia gas, and design a long short-term memory (LSTM) neural networks model to predict the final concentration based on the very-limited transient measurement. Our approach is accurate (within 0.12% error) and save more than 99.6% of energy.
Unobtrusive Motion Monitoring for Sleep quality/disorder Diagnosis
MotionTree, an in-bed Body Motion Monitoring system for Elderly Care Research [PDF]
Monitoring body motions during sleep can serve as a proxy to evaluate the sleep quality and to diagnose certain types of disease (periodic limb movement disorder, restless leg syndrome, etc.) We proposed to a load-cell based sensor systems to monitor a person's gross body motion on a bed. The system extract spatial features of gross motions and classifies them into 9 classes. We can estimate how those motion types impact a person's sleep and whether there is any sleep disorder that we should be aware of.
A Side-channel anomaly detection scheme for PLCs
NIPAD: a non-invasive power-based anomaly detection scheme for programmable logic controllers [PDF]
Programmable logic controllers (PLC) are widely used in industrial process control and are one of the most vulnerable devices in the pipeline. Traditional security methods can't easily apply to PLC due to further security concerns from manufactures while exposing the infrastructure. In this study, we propose to monitor the energy consumption of different command sets to detect any anomaly events inside of the PLC.