Summer 2015

Download the SUNFEST 2015 report.

Project Archive

Christopher Alicea-Nieves (Computer Engineering) - University of Puerto Rico

Caffe Framework on the Jetson TK1: Using Deep Learning for Real Time Object Detection

Advisor: Camillo J. Taylor

Deep Neural Networks are an approach of using state-of-the-art machine learning algorithms in order to improve the applications of the computer vision field. These neural networks are complex mathematical models which can be trained to identify certain objects within a frame, and some of them are able to identify also where on the image is the object that they identified. This technical report focuses on presenting the implementation of a deep neural network on a low-power embedded computer system, specifically the NVIDIA Jetson TK1, in order to run convolutions in real-time for object detection. This can be achieved implementing Fast R-CNN, a state of the art model for convolutions, in the machine learning framework called Caffe.

Project report | Presentation slides

Samantha Burns (Materials Science) - University of Pennsylvania

Preparation of a Mechanochromic System using Liquid Crystal Dispersed in Polymer

Advisor: Shu Yang

We have developed an inexpensive and sensitive mechanochromic system. Our system is prepared by a simple emulsion procedure of nematic liquid crystal in an elastomeric polymer. The resulting polymer dispersed liquid crystal (PDLC) exhibits a transmittance change in response to strain. We aim to characterize the mechanical, optical, and morphological behavior underlying this mechanochromic response. We propose a new model that illustrates how mechanical strain modifies the polymer-liquid crystal interface and alters the nematic director fluctuation within the liquid crystal droplets. We propose that both of these factors play key roles in changing the light transmittance of PDLC samples. Understanding the mechanism behind this mechanochromic behavior is critical for developing precise and reliable methods to tune a material’s optical properties. This capability is critical for commercial applications, such as smart windows and biomedical pressure sensors.

Project report | Presentation slides

Hannah Corcos (Mechanical Engineering) - University of Maryland

Exercising Caenorhabditis elegans

Advisor: Haim Bau

Muscles deteriorate with age, restricting mobility. It is thought that regular exercise can lessen the rate of muscle deterioration, but it is not known if it affects longevity. Using the nematode Caenorhabditis elegans, we wish to determine a relationship between exercise and the aging process. To do this, we need to find a way to induce the animals to exercise. A method of exercising C. elegans was identified by exposing them to a fluid flow. Without food, the C. elegans maintained a constant low activity level until the external flow reached a threshold velocity. Once this threshold velocity was exceeded, the C. elegans assumed a heightened activity level. The effect of food on the activity level was also examined. In the presence of food the animals’ activity level declined since they were not actively searching for sustenance. Lastly, using different strains with mutated sensory neurons, it was determined that an elevated activity level is caused by the animals’ cilia. This opens the way find how exercise affects mobility level as a function of age and longevity.

Project report

Santiago Gonzalez (Electrical Engineering) - Case Western Reserve University

Watts App: An Energy Analytics and Demand-Response Advisor Tool

Advisor: Rahul Mangharam

Real-time electricity pricing and demand response has become a clean, reliable and cost-effective way of reducing peak demand on the electricity grid. Annual revenues to end-users from demand response markets are more than $700 million, making demand response the largest virtual generator in use [1]. DR-Advisor, an open source software tool created at the University of Pennsylvania, acts as a recommender system for building’s facilities manager. Using historical data from a building, DR-Advisor uses data-driven models to suggest suitable control actions to meet the desired load curtailment during demand response events. Using data sets from several buildings on the University of Pennsylvania’s campus, we enhance the capability of DR-Advisor by adding plug-ins for data-preprocessing and energy analytics.

Project report | Presentation slides

Mario Gutierrez (Physics) - North Park University

Hearing Loss: Piezo-polymers Nano-fibers for Bio-medical Applications

Advisor: Jorge Santiago

Hearing loss happens when hair cells inside the cochlea are damaged and about one third of people in United States have some degree of hearing loss. This can possibly be solved by replacing those damaged hair cells with polyvinylidence fluoride (PVDF) nanofibers. PVDF is a highly non-reactive and pure thermoplastic fluoropolymer produce by the polymerization of vinylidene difluoride. PVDF have great piezoelectric properties and this means that it has the ability to generate an electric charge in response to an applied mechanical stress. In this research, I used a computer program to simulate a cylindrical cantilever, which resembles a hair cell in the cochlea, and analyze the response of the frequencies by changing parameter values (length and diameter). Here, I report that I was able to analytically verify the frequencies obtain from the computer program are similar with those obtain using an equation that I was able to derive. I also simulated two different conditions of the fiber to obtain the current that each fiber produce when stimulated in the audible frequency range.

Project report | Presentation slides

Alexander Hunt (Materials Science) - University of Pennsylvania

Graphene-Boron Nitride Heterostructures: A platform for the synthesis of the nanotransistor

Advisor: Charlie Johnson

Graphene-Boron Nitride (G-BN) heterostructures can lead to the realization of nanoscale electronics that will be smaller than the dimensional limit—14 nanometers—of silicon transistors and provide higher mobilities. However, the graphene-boron nitride heterostructure although self-insulating cannot function as a transistor alone due to not having a second conducting pathway. Thus, the utilization of the graphene-boron nitride heterostructure within a side gated FET would require two G-BN flakes to randomly nucleate together. In contrast, the growth of a graphene ribbon around the BN ribbon if fabricated would allow for a single flake to act as a side gated FET. In order to grow such a G-BN-G heterostructure, the hydrogen etching of BN at high processing temperature must be overcome which means the use of methane as a carbon precursor for graphene growth is no longer a viable option. As a result, benzoic acid will be studied in the following report as a potential carbon source to synthesize the outer graphene ribbon due to its low thermal breakdown temperature with present findings showing that hydrogen etching of BN has been reduced.

Project report | Presentation slides

Timothy Kindervatter (Electrical Engineering) - The College of New Jersey

Design of a Wireless Power Transfer System for Wireless Sensor Networks in Biomedical Applications

Advisor: Jan Van der Spiegel, Milin Zhang

Analysis and design of a wireless power transfer system for application in a brain implant is presented. The design consists of a monitoring system and three main stages: the power transfer stage, the power management stage, and the energy storage stage. RF energy harvesting and inductive charging are identified and tested as viable alternatives for wireless power transfer. Designs for a custom-designed power management PCB as well as a prototype monitoring system are presented.

Project report

Timothy Linscott (Electrical Engineering) - Seattle University

Cross-CUT Interference Present in Timing Extraction

Advisor: Andre DeHon

Built-In Self Tests such as those developed by Wong, Sedcole, et al. and Gojman for measuring the internal path delays of a reprogrammable chip require maximizing isolation to eliminate interference between their components. This work demonstrates that placing anything more than a single measurement circuit on a chip at a time can influence the results of a measurement. The load placed on the clock by the measurement circuits is explored as a possible cause along with the ways that the different clock quadrants can be exploited to reduce the clock’s influence on the measurements. This work also begins characterizing the noise introduced by running Timing Extraction measurements in parallel and demonstrates how this noise can be minimized.

Project report | Presentation slides

Erica Sadler (Bio-Engineering) - Cornell University

Photo-Actuation: Determining the Effect of Blue Light on the Actuation of Micro-Beads by Serratia

Advisor: Vijay Kumar

MicroBioRobots (MBRs) are robotic systems with both an animate and inanimate component. The animate, or biological, component of MBRs acts as an actuator, providing the source of power for the MBR’s movement. Since this biological component is, in fact, a living organism, it is generally very responsive to external stimuli. For this reason, microorganisms, such as bacteria, act as sensors in addition to actuators. In order to utilize the innate sensing of microorganisms effectively, their response to particular stimuli must be understood in a way that is both measurable and predictable. Light is a stimulus known to affect the motility of bacteria. I developed a 2D tracking algorithm in MATLAB in order to analyze the motion of MBRs made up of micro-beads actuated by Serratia marcescens. I report on the effect of blue light on this motion, finding that a high intensity of blue light leads to an increase in the tumbling of bacteria and an overall decrease in the movement of MBRs, sometimes working strongly against a current. Through my attempts to obtain accurate and useful data, I also examine the effectiveness of several techniques for creating functional MBRs and acquiring successful video data. This is a complex process of its own and does not yet have a generally agreed upon protocol. I take into account the pros and cons of several techniques, finding that the smaller bead size I observed (3μm) was easier for the bacteria to move and control than the 6.8μm beads. I also found that washing the micro-beads using sonication did appear to promote the natural attachment of the bacteria to the micro-beads. Due to the adhesion of both beads and bacteria to surrounding surfaces, it was best to use Percoll to give the MBRs buoyancy and then observe them on some plane other than the top or bottom to prevent any beads or bacteria that had become stuck in place from affecting the data. Drift is another common problem when observing objects of this size, and the best way I found to control it was through the use of a sealed chamber. Even with these improvements, future research is necessary to find the optimal way of collecting and analyzing data on MBRs. Future studies should also be conducted with MBRs using lower intensities of blue light to determine if different light intensities will affect the MBRs’ motion differently. It would also be useful to conduct further research on how bacteria move beads of varying sizes so that the use of bacteria as an actuation technique is not limited solely to beads with diameters of 3μm.

Project report | Presentation slides

Eleanor Tursman (Physics) - Grinnell College

Geometric Stereo Increases Accuracy of Depth Estimations for an Unmanned Air Vehicle with a Small Baseline Stereo System

Advisor: Camillo J. Taylor

A small stereo camera is a light and economical solution for obstacle detection and avoidance for unmanned air vehicles (UAVs). It is possible to create depth maps of a scene given a pair of stereo frames from such a camera. However, the smaller and lighter the stereo camera, the smaller its baseline, which in turn limits its ability to discriminate objects that are farther away. Maximizing the effective range of the stereo setup is essential for real-time applications, where quick decisions need to be made to avoid obstacles approaching the UAV. To overcome this difficulty, we use knowledge of the camera’s position over time to mimic large baseline disparity calculations, a technique we dub “geometric stereo.” This paper outlines a simulation which shows that this technique is able to obtain better results for depth estimation with smaller confidence intervals than those obtained by simply averaging discrete depth maps over time.

Project report

Jerrell Walker (Electrical Engineering) - University of Pennsylvania

Piezoelectric Polyvinylidene Fluoride Nanofibers

Advisor: Jorge Santiago

This work focuses on electrospinning nanofibers for use as an alternative solution to cochlear implants which address a loss of hearing. The focus of this work is to generate self-poled fibers. Self-poling requires a diameter of 100 nm. Polymer solutions of Polyvinylidene Fluoride (PVDF) and a light-emitting polymer, F8BT, are dissolved in combinations of dimethylformamide, and acetone. Conventional also termed far-field electrospinning is employed in order to generate the nanofibers. Average fiber diameters were larger than desired. Two notable solutions yielded average diameters of 180 nm to 200 nm. Characterization of diameters was pursued via high resolution scanning electron microscopy(SEM). Aluminum foil was used as both a collection substrate and a substrate for the SEM work. Further work will have to be done modifying electrospinning parameters to obtain the proper diameter for this application.

Project report | Presentation slides