Embedded Systems (1000) and Roboics & …...Embedded Systems (1000) and Roboics & Intelligent Machines (2000) Project No. Student(s) Project Title 1001X12 Fadil Amiruddin Creaion of

Embedded Systems (1000) and Robotics & Intelligent Machines (2000)Project No. Student(s) Project Title1001X12 Fadil Amiruddin Creation of smart glasses that help the hearing impaired live more conveniently

everyday 1002X12 Matthew Newcomer Communicating Prior Knowledge of On-Road Incidents via Microprocessor

Packages for the Reduction of Reaction Times1003X11 Tarun Ravikumar Harnessing Mechanical Energy From Footsteps

1004X12 Jason Romps Propagation of WiFi through Different Materials

2001X12 Colin Fox Using Unmanned Aerial Vehicles and Mask RCNN to Geolocate Trash and Clean Beaches

2002X12 Siddharth Ganesh HelpAR

2003X11 Vinay Jagan Classifying Heart Arrhythmias Real-Time using Machine Learning

2004T12 Sophia Cruz, Cora McQuaid Shoo-B: The Engineering of a Robotic Scarecrow to Deter Roosting Vultures

2005T12 Alfred Premkumar, Kaitlyn Rahn DeepAV: Forecasting Pedestrian and Vehicle Path in Autonomous Vehicle LIDAR Scenes

2006X12 Logan Speckhard Designing and prototyping a 3D printed non invasive robotic prosthesis with voice controlled mimicking functions

2007X11 Rishi Vanga Identifying patterns in text in order to determine if articles were written by the same author

2008X12 Andrew Zhang Generating 3D Objects Using Images with Generative Adversarial Networks

Category Student Count: 14

1001X12Embedded Systems

1000

LCPS RSEF OFFICIAL ABSTRACT - 2020Creation of smart glasses that help the hearing impaired live more conveniently everyday Fadil AmiruddinOne issue deaf people face is the ability to not have normal face to face conversation. Google's solution to this issue is to write real time subtitles on a phone screen so deaf people can read the subtitles. Though this is a pretty effective solution, this product takes that solution one step forward by putting the subtitles on the glasses the user is wearing. The deaf can now have normal face to face conversations just like everyone else without looking at their phone. This product had 3 major parts which combined to form one product. The first part is the android application itself, in order to get audio data from its surroundings whenever sound is heard and translate that sound data to text. This text data is sent to the Arduino with a bluetooth connection. The arduino then sends this text data to the transparent OLED screen. The transparent display now shows text picked up from the phone. Next is displaying the image to the user's eye. This is done by simply putting the OLED screen below the user's eye. The reflection of the screen will then hit the reflective part of the sunglass lens and the reflection of the OLED screen will be shown to the user's eye. Doing this allows the user's eye to focus on the image projected by the OLED screen because reflection allows the screen to appear further than it actually is thus allowing the user to actually read the text.

Capps, D., Santos, S., Middelbeek, H., Santos, R., Fezari, Dale, … Silva, M. (2019, May 10). ESP32 Bluetooth Classic with Arduino IDE - Getting Started. Retrieved from https://randomnerdtutorials.com/esp32-bluetooth-classic-arduino-ide/

Duelbonna, Y. (2019). Bluetooth Communication using MIT App Inventor. Retrieved from https://roboindia.com/tutorials/bluetooth-communication-mit-app-inventor/

Mayon, I. (2017). Sparkfun Transparent OLED. Retrieved October 3, 2019, from https://learn.sparkfun.com/tutorials/transparent-graphical-oled-breakout-hookup-guide.

I/We hereby certify that the above statements are correct and the information provided in the Abstract is the result of one year's research. I/We also attest that the above properly reflects my/our own work (digitally signed).


1000

LCPS RSEF OFFICIAL ABSTRACT - 2020Communicating Prior Knowledge of On-Road Incidents via Microprocessor Packages for the Reduction of Reaction Times Matthew NewcomerThe goal of this project is to mitigate the loss of life due to car accidents by providing drivers advanced warning of the crash. Roughly 50 percent of crashes annually have driver inattention and/or skill mismatch listed as causes of the accident, both of which can be minimized with forewarning. Small ‘nodes’, or single board computers like the Raspberry Pi, possess enough processing power to proactively alert other nodes within the network when a hazard is detected. The system is able to recognize the crash, bundle external data like location, speed, and heading, and transmit it to other nodes. When this data is received, the node uses this information to determine whether to alert the driver or to bypass the alert. If deemed necessary, the receiving node alerts the driver through an auditory alert which contains instructions about the incident ahead and the suggested course of action. Preliminary testing has shown that a pair of nodes have transmitted data approximately 10 meters through open air with a 99% transcription accuracy. This number would likely fall if implemented in an actual vehicle due to the glass and metal surrounding the node, but a higher power antenna and external mounting would lead to real-world applicability. This system, if widely implemented, (e.g. in every vehicle) could transfer crash data seamlessly and work in remote environments. Next steps include implementing a text-to-speech engine for translating data into useful audio and expanding the size of the network by manufacturing more nodes.

Bella, F., & Silvestri, M. (2017). Effects of directional auditory and visual warnings at intersections on reaction times and speed reduction times. Transportation Research Part F: Traffic Psychology and Behaviour, 51, 88–102. doi: https://doi.org/10.1016/j.trf.2017.09.006

J. Jagannath, S. Furman and A. Jagannath et al. (2018). HELPER: Heterogeneous Efficient Low Power Radio for enabling ad hoc emergency public safety networks. Ad Hoc Networks, 89, 218–235. doi: https://doi.org/10.1016/j.adhoc.2019.03.010

National Highway Traffic Safety Administration. (2008). National Highway Traffic Safety Administration. National Highway Traffic Safety Administration. Retrieved from https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/811059



1000

LCPS RSEF OFFICIAL ABSTRACT - 2020Harnessing Mechanical Energy From Footsteps Tarun RavikumarThe average American walks approximately 6,000 steps per day; each step creates mechanical energy, energy which ends up being wasted and dispersed into the environment. Tapping into this wasted energy opens a door for opportunities to supplement the user’s actions. Varying amounts of piezoelectric sensors were used to generate this energy which gets stored in a LiPo battery through the aid of the BQ25570 chip. Other projects involving piezoelectric sensors are only capable of storing energy into a capacitor, while this experiment stores it in a full-fledged battery. For the final design, 28 piezoelectric sensors were used, which generated, approximately 0.001V or 0.025% of the total LiPo battery’s capacity, after just 25 steps. These results can be generalized to all humans of similar weight and footstep acceleration of the experimenter. The primary avenue of improvement to this experiment is to add more piezoelectric sensors; however, 28 sensors were used in this experiment, which is the most ever used compared to similar experiments, and the circuit design is currently the most efficient way to generate electricity that gets stored into a battery using piezoelectric sensors. An add-on to this experiment is being developed, where an accelerometer and gyroscope is placed into the shoe. The data from these sensors will be run through a neural network and will predict the action being done on the shoe, for example, if the user is jumping, the shoe will detect that behavior.

Caliò, R., Rongala, U. B., Camboni, D., Milazzo, M., Stefanini, C., de Petris, G., & Oddo, C. M. (2014, March 10). Piezoelectric energy harvesting solutions. Retrieved December 15, 2019, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4003967/

Hadzialic, R. (2019, July 29). Energy harvesting with the TI BQ25570 - part 1. Retrieved January 15, 2020, from https://www.lab4iot.com/2019/07/29/energy-harvesting-tutorial-with-the-ti-bq25570-part-1/

Zelisko, M., & Anderson, K. (2019, June 20). Piezoelectric generator - activity. Retrieved December 20, 2019, from https://www.teachengineering.org/activities/view/uoh_piezo_lesson01_activity1



1000

LCPS RSEF OFFICIAL ABSTRACT - 2020Propagation of WiFi through Different Materials Jason RompsSociety is becoming ever more dependent on being connected to the Internet. In the common household, users connect using the IEEE 802.11 protocol, better known as WiFi. When people think of a stronger WiFi signal, they think of the newest router with the newest technological capabilities, however, the strongest factor in determining WiFi signal strength is the position of the router. Finding the optimal router location in a house will benefit the user by giving them overall faster speeds throughout the entire home. While it is known that barriers correlate with a weaker signal of WiFi, there are no exact data points that compare WiFi signal strength to waves propagating through different materials. These variables include the absorption, reflection, and distortion of signals on walls of varying thicknesses and materials. It is known that waves experience path loss as they propagate through free space, but there is little information on how different materials react to WiFi waves at different distances. Certain materials exhibit much more resistance to wave propagation than others, while some materials rarely affect wave propagation. The experiment is currently being conducted and data collection has just begun. Results will be compiled to generate a mathematical formula that compares all materials’ inhibiting properties under a new constant known as the “WiFi Inhibitor Standard”. These results are paramount for an engineer who is given the materials and thicknesses of walls in a building and must find the optimal location for a router.

Wi-Fi. (2006, February). Database and Network Journal, 36(1), 16+. Retrieved from https://link.galegroup.com/apps/doc/A163332784/GPS? u=va_s_053_0900&sid=GPS&xid=cbe5e474

Wi-Fi Router Location: 802.11 Signal Coverage. (n.d.). Retrieved from https://www.electronics-notes.com/articles/connectivity/wifi-ieee-802-11/wi-fi-location- coverage.php

Wilson, R. (2002). Propagation Losses Through Common Building Materials. Reflection and Transmission Losses Through Common Building Materials.


2001X12Robotics & Intelligent

Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020Using Unmanned Aerial Vehicles and Mask RCNN to Geolocate Trash and Clean Beaches Colin FoxPlastics in the environment have led to the decline of marine life populations for decades. Plastic is an issue due to a combination of its low weight, durability, and resistance to water and microorganisms. Plastic has physical and chemical effects on the marine environment. Some examples of the physical effects are entanglement by fishing nets, blockage of intestinal tract, and bubbles in plastic prevent wildlife from diving for food. The chemical effect on the environment is due to the presence of polychlorinated biphenyls in plastics. Mask R-CNN was implemented on a designed and built drone to clean beaches of the plastics that plague them. The accuracy of Mask R-CNN at detecting trash vs non-trash accuracy, detecting the correct type of trash, and the Trash Retrieval Device’s (TRP) successful trash retrieval instances were all recorded. The implementation of the Mask R-CNN on the drone is currently being conducted and results are pending. In initial testing, it was found that the theoretical trash yield identified by Mask R-CNN was 70% of present trash, 87.5% of the objects were trash and the TRP’s successful trash retrieval percentage at 92%. The next steps of the project include using swarms of drones in testing and implementing longer lasting Lithium-Ion batteries that would replace the traditional Lithium-Polymer batteries. This solution provides valuable information about trash found on beaches and can mitigate the amount of trash in oceans. This also alleviates error factors regarding gathering information due to the consistency of a common trash surveyor.

Darrell, T., Donahue, J., Girshick, R., & Malik, J. (2014). Rich feature hierarchies for accurate object detection and semantic segmentation. Computer Vision Foundation.

Derraik, J. G. (2002). The pollution of the marine environment by plastic debris: a review. Marine Pollution Bulletin, 44(9), 842–852. doi: 10.1016/s0025-326x(02)00220-5

Ryan, P. G., Moore, C. J., Franeker, J. A. V., & Moloney, C. L. (2009). Monitoring the abundance of plastic debris in the marine environment. Philosophical Transactions of the Royal Society B: Biological Sciences, 364(1526), 1999–2012. doi: 10.1098/rstb.2008.0207



Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020HelpAR Siddharth Ganesh

Alzheimer's, otherwise known as Senile Dementia is the progressive deterioration of one's memories and mental functions. The people who are primarily affected by Alzheimer's are typically people between the ages 40+.

In the early stages of Alzheimer’s memories start to be forgotten, even going as far as the patient not even remember his name. Those who are close to a loved one with Alzheimer’s have experienced first hand how heartbroken and helpless they feel to help when their loved one's memories are slowly forgotten. The victim will also typically forget dress themselves, feed themselves and use the bathroom. Alzheimer’s has had a terrible effect on the victim’s families and the victims themselves. The survival rate for people with Alzheimer’s that are 70+ is about 49%.

There are two types of Alzheimer's Late-Onset Alzheimer’s and Early-Onset Alzheimer’s, in Late-Onset Alzheimer is the signs typically first appear when the person is in their mid 60’s, it is also known to be the most common type of Alzheimer's and is due to a gene called APOE e4. In Early-Onset Alzheimer’s signs typically first appear between a person’s 30 ’s and 60’s, Early-Onset Alzheimer’s is very rare and is usually caused by gene changes passed down from generation to generation. Some other symptoms of Alzheimer’s are having trouble focusing, having a hard time doing ordinary activities, being very confused or frustrated for no reason whatsoever, having very dramatic mood swings, feeling very disoriented at times and even having trouble communicating. Higher risk of Alzheimer’s is linked to a protein called ApoE, which the body uses in order to move cholesterol in the blood. People with high blood pressures and high cholesterols have a higher chance of getting Alzheimer's.

The use of Augmented Reality is by no means a revolutionary idea. The application of Augmented Reality has happened for decades, but what I am trying to do is a unique application that not only virtually interacts with real objects but also identifies them. The usability of this app extends more than just for Alzheimer’s patients, everyday people could use it. For example, if you misplace your keys then you can pull out your phone and use the camera of the phone to scan the room and then it will locate the keys. The research about Alzheimer’s will certainly help with the app development and this research project primarily focuses on the early stages of Alzheimer's and how to improve the patients standard of living by, creating an app that can both define the items in a given environment, and identify objects in the environment. It will help both the caregivers and the patients because it will help them identify item and objects that they may have forgotten, for example was if you scan a lamp and you forgot what it was and how to use it, it has a label and instructions of how to turn it on.

Understanding Alzheimer's Disease: The Basics. (n.d.). Retrieved September 23, 2018, from https://www.webmd.com/alzheimers/guide/understanding-alzheimers-disease-basics

What Happens to the Brain in Alzheimer's Disease? (n.d.). Retrieved September 23, 2018, from https://www.nia.nih.gov/health/what-happens-brain-alzheimers-disease

What Causes Alzheimer's Disease? (n.d.). Retrieved September 23, 2018, from https://www.nia.nih.gov/health/what-causes-alzheimers-disease

Brain Tour. (n.d.). Retrieved September 23, 2018, from https://www.alz.org/alzheimers-dementia/what-is-alzheimers/brain_tour

IOS - Augmented Reality. (n.d.). Retrieved September 23, 2018, from https://www.apple.com/ios/augmented-reality/

Kernisan, L. (2018, July 05). 8 Alzheimer's Behaviors to Track. Retrieved September 23, 2018, from https://www.aplaceformom.com/blog/12-17-15-alzheimers-behaviors-to-track/

Early-Onset Autosomal Dominant Alzheimer Disease: Prevalence, Genetic Heterogeneity, and Mutation Spectrum. (2008, January 09). Retrieved September 23, 2018, from https://www.sciencedirect.com/science/article/pii/S0002929707623179



Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020Classifying Heart Arrhythmias Real-Time using Machine Learning Vinay JaganHeart arrhythmias occur when there are irregular heart beats that are symptomatic of conditions such as high blood pressure and heart attacks. Currently, there are two types of models that exist for arrhythmia prediction: classification models and real-time prediction models. My project is to combine these two models in the form of a real-time classifier. A group of researchers from the University of California built a Convolutional Neural Network model in order to classify ECG signals retrieved from the MIT-BIH Arrhythmia Dataset. The ECG signals were split into individual heartbeats for input through a series of preprocessing steps. The group uploaded their data and model onto a Kaggle kernel which allowed for a quick replication of the model. During the replication of the model, an augmentation step was omitted in order to shorten inference time for predictions. The resulting model provided a 78% to 100% accuracy among the 5 classification targets. The model was then inserted into an application script in order to predict real-time. The mean inference time was 3.73 ms with a standard error of 0.006 ms.

Kachuee, M., Fazeli, S., & Sarrafzadeh, M. (2018). ECG Heartbeat Classification: A Deep Transferable Representation. 2018 IEEE International Conference on Healthcare Informatics (ICHI). doi: 10.1109/ichi.2018.00092

Rajpurkar, P., Hannun, A. Y., Haghpanahi, M., Bourn, C., & Ng, A. Y. (2017). Cardiologist-level arrhythmia detection with convolutional neural networks. ArXiv:1707.01836 [Cs]. http://arxiv.org/abs/1707.01836

Zhou, X., Ding, H., Wu, W., & Zhang, Y. (2015). A Real-Time Atrial Fibrillation Detection Algorithm Based on the Instantaneous State of Heart Rate. Plos One, 10(9). doi: 10.1371/journal.pone.0136544


2004T12Robotics & Intelligent

Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020Shoo-B: The Engineering of a Robotic Scarecrow to Deter Roosting Vultures Sophia Cruz, Cora McQuaid

In Loudoun County, a big problem that people encounter is vultures. Not only are they an unwanted eyesore, but they also cause property damage.

According to The National Public Radio, “Every few years, usually in the winter months, residents of the town of Leesburg, Va., come home from work to find their backyards overrun with turkey vultures. Not just a few birds, but hundreds of them” (Baker, 2013). Groups of vultures gather on roofs and there are only a few methods to disperse them. Hanging fake dead birds from roofs is one of the most common approaches but it is not visually appealing. Along with this, the Migratory Bird Act prohibits individuals from harming vultures.

Vultures are intelligent birds; when they realize that an obstacle is not harmful, they will return to their routines. A vulture exposed to an environmental change that is merely an inconvenience will realize that they’re not in danger and will return to their roosting habits.

Our solution to this problem was to create a robotic prototype that is programmed to spontaneously change variables such as its speed, flashing lights, randomized movement, etc. It is designed to scare off vultures in a non-harmful manner and prevent them from roosting on the peaks of roofs to limit property damage. We decided that it was vital to incorporate random functions so that the birds would not be able to decipher any pattern in the robot’s movements. The robot moves along a belt through the incorporation of a pulley system, an Arduino Uno board, and robotic coding. For demonstration purposes, we decided to include 5 buttons which will execute the 5 randomized functions that we have coded for. When the software registers an input (pressing the button), it will produce an output (the randomized functions). With each button, there is a different routine that the robot will carry out. The final iteration of the robot when mounted on a roof, would be edited to have a motion sensor as the input, and the robot would run through the various routines when the sensor is triggered. In addition, the final product would be weatherproof and built with lighter weight materials.

U. S. D. of A. (2019). Managing Vulture Damage. Retrieved from https://www.aphis.usda.gov/publications/wildlife_damage/content/printable_version/fs_vulture_damage_man.pdf

Layden, L. (2016, August 11). Black Vultures Are Protected By Treaty, But Eating the Profits of Oklahoma Ranchers | StateImpact Oklahoma. Retrieved from https://stateimpact.npr.org/oklahoma/2016/08/11/black-vultures-are-protected-by-treaty-but-eating-the-profits-of-oklahoma-ranchers/


2005T12Robotics & Intelligent

Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020DeepAV: Forecasting Pedestrian and Vehicle Path in Autonomous Vehicle LIDAR Scenes Alfred Premkumar, Kaitlyn Rahn

Current autonomous vehicle (AV) systems excel in object identification and localization. These systems use this to base autonomous path-finding on maintaining a set distance from surrounding vehicles and pedestrians. This current methodology is effective, but inherently inefficient compared to that of human autonomy. As market AV systems begin to introduce hardware better suited for neural networks, companies are racing to create systems that better imitate human autonomy, which requires accurately forecasting the motion of vehicles and pedestrians around a self-driving vehicle to better refine autonomous vehicle path-finding. This paper presents a novel approach to autonomous vehicle path prediction using a Recurrent Neural Network (RNN) accounting for situational context. The RNN is retrained from TrafficPredict, a recent RNN traffic forecasting algorithm. The inputs into the new RNN consist of two dimensional coordinates and a categorical representation of the location (e.g. Boston, Massachusetts; Holland Village, Singapore). The RNN is trained on two datasets preprocessed from the NuScenes AV dataset, consisting of vehicle motion data and pedestrian motion data, which produces two models. The output of these models, given past motion and location, are the coordinates of a vehicle or pedestrian one second in the future. Current training results show a -4.10 training loss and and a -5.22 validation loss in pedestrian motion forecasting, and a -4.11 training loss and and a -5.67 validation loss in vehicle motion forecasting. The test average and final displacement loss data are pending.

Ma, Y., Zhu, X., Zhang, S., Yang, R., Wang, W., & Manocha, D. (2019). TrafficPredict: Trajectory Prediction for Heterogeneous Traffic-Agents. Proceedings of the AAAI Conference on Artificial Intelligence, 33, 6120–6127. doi: 10.1609/aaai.v33i01.33016120

Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., & Savarese, S. (2016). Social LSTM: Human Trajectory Prediction in Crowded Spaces. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). doi: 10.1109/cvpr.2016.110

Caesar, Holger, et al. “NuScenes: A Multimodal Dataset for Autonomous Driving.” ArXiv:1903.11027 [Cs, Stat], Jan. 2020. arXiv.org, http://arxiv.org/abs/1903.11027.



Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020Designing and prototyping a 3D printed non invasive robotic prosthesis with voice controlled mimicking functions Logan Speckhard

According to the national archives, in 2015 it was estimated that 41,000 people in the United States experienced upper limb amputations below the elbow. At the same time, upper limb amputations accounted for the majority of trauma related amputations in the armed services (approximately 69 percent). The United States government archives states that the functions of upper extremities are more difficult to replicate than that of lower extremities due to the use of fine motor functions.

Of the prosthetics available, this test was meant to combine the best aspects of all of them. The low end price of a passive prosthetic, with the non invasive control in a body powered prosthesis, and the fine motor function of an electrically powered prosthesis. Along with this, adding the ability to control many individual joints off of one sensor. If successful, this robotic prosthesis will cost a total of four hundred dollars, and have individual control of all finger joints. The prosthesis will use a glove fitted with sensors on the opposite hand, making it non invasive. Using the Arduino Uno platform, it will complete the circuit when triggered with the word ON, and break the circuit when triggered with the word OFF. The implementation of this prosthesis would be huge not only in the civilian use, but in the use with veterans, and active duty servicemen injured in combat.



Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020Identifying patterns in text in order to determine if articles were written by the same author Rishi VangaOftentimes the author of old writings, such as Shakespeare’s plays, is debated or uncertain and historians spend countless hours tediously deciding whether the authorship is legitimate. This program could complete that process almost instantaneously and could be just as accurate. The independent variable was the part of the program that identified the author’s characteristics. The dependent variable was the part of the program that determines whether the writings were by the same person or not. The two main steps to this project were to first write a program that could identify certain variables that were manually decided on. The second step was to run a logistic regression program on these values and have the computer generate weights for them, with which it can then say whether other articles were written by the same author or not. One hundred data points were plugged into this program and it was able to achieve an accuracy of 91% across these hundred. The alternative hypothesis was that if a program is written to group articles by their author, it will be able to achieve an accuracy above 80%. This was supported because the accuracy was 91%. The independent variable influenced the dependent variable. If further research is to be done, more variables to identify in the writing could be found and more data points could be used.

Hsinchun, C., Nunamakera, J., & David, Z. (2014). Analyzing firm-specific social media and market: a stakeholder-based event analysis framework. Retrieved from https://www.sciencedirect.com/science/article/pii/S016792361400205X.

Kim, H., Chung, M., & Lee, W. (2008). Literary style classification with deep linguistic analysis features [PDF file]. Retrieved October 3, 2019 from https://pdfs.semanticscholar.org/19f2/63c908b6b087c9ae30241de7a3209fabe649.pdf

Suh, J. H. (2016). Comparing writing style feature-based classification methods for estimating user reputations in social media. Retrieved October 3, 2019, from https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4775724



Machines2000

LCPS RSEF OFFICIAL ABSTRACT - 2020Generating 3D Objects Using Images with Generative Adversarial Networks Andrew ZhangThis project aims to create a neural network model to create 3D models from images. Previous approaches used a type of neural network called generative adversarial networks. The discriminator in the previous tried to classify silhouettes of the object and the generator created voxel representations of objects and created images using a differentiable renderer. However the silhouettes don’t give enough information. My approach uses depth information to create point cloud representations. This is done by having the generator create known viewpoints and using those to solve for the point cloud. Then those are used to create depth images that are fed into the discriminator using a differentiable renderer. The result of this approach is a more accurate representation of an object. For example, it’s more able to represent smooth curves unlike the previous voxel representations without using a large amount of data.

Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Courville, A., … Ozair, S. (2014, June 10). Generative Adversarial Networks. Retrieved from https://arxiv.org/abs/1406.2661

Lin, C.-H., Kong, C., & Lucey, S. (2017, June 21). Learning Efficient Point Cloud Generation for Dense 3D Object Reconstruction. Retrieved from https://arxiv.org/abs/1706.07036

Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Radford, A., & Chen, X. (2016, June 10). Improved Techniques for Training GANs. Retrieved from https://arxiv.org/abs/1606.03498


Embedded Systems (1000) and Roboics & …...Embedded Systems (1000) and Roboics & Intelligent Machines (2000) Project No. Student(s) Project Title 1001X12 Fadil Amiruddin Creaion of

Documents