This project aims at using environmental cameras to detect a user’s activities inside a gym, e.g. push-ups, squats, etc. It could also count and record different activities a user has performed, The goal is to analyze a user’s action and give user real-time feedback.
Similar to the “Camera-based recognition” project below, this project could use existing implementation on volumetric template matching. The students may also develop new algorithms (e.g. DNN). The real-time system can be built on top of Gabriel, but the client part has to be rewritten if we use non-Android based cameras.
This project provides naive assistance to visually impaired Google Glass users by providing spoken cues of objects they are looking at. The Google Glass device will stream video to a cloudlet, which will detect objects in the video frames and send object names back to the device. This project will use existing open source implementations of state-of-the-art computer vision models using deep neural networks such as Teradeep: https://github.com/teradeep/demo-apps and https://www.youtube.com/watch?v=_wXHR-lad-Q.
The streaming can optionally be done with the Gabriel framework:
Environments: Android and Linux
Skills Learned: Python, Caffe (C++/Python) or Torch (Lua), OpenCV
4. Interactive Google Street View hyper-lapse simulation
In addition to the latency benefit, a cloudlet is a perfect place for caching large amount of data with locality. In this work, the students will implement interactive hyper-lapse simulation using Google Street View. A reference prototype is implemented and open sourced at https://github.com/TeehanLax/Hyperlapse.js. The students will extend this work to enable an interactive hyper-lapse system. A user will choose the route interactively as the hyper-lapse video proceeds and the cloudlet will provide a cache for the Google Street View to speed up downloading of the image data (reference demo video: https://vimeo.com/63653873).
This project seeks to use camera based sensing of gestures in an interactive game setting. The idea is to allow the user to control the game using large hand and body motions, observed by a fixed camera (e.g. on a laptop). Example: Gestris (gesture controlled tetris) from a few years ago. This system used an unmodified tetris-style game, and a pair of cameras to allow two players at a time. The computer vision code used volumetric template matching – a fairly simple, but computationally expensive algorithm that is fairly easy to train. This will not run in interactive time on a laptop, and will need cloudlets. The goal of this project is to demonstrate some sort of gesture-controlled interactive game. Alternatively, the system can be designed to respond to other camera-sensed inputs, such as a laser pointer on real objects or projected screen.
This project will attempt to apply superresolution techniques to produce a “live” view that is higher resolution than the native camera on a mobile phone or Google glass device. The idea behind superresolution is that multiple images that are slightly offset can provide clues to the ground truth at a sub-pixel level. There are many algorithms to do this, including some that are supported in OpenCV. The goal of the project is to demonstrate some sort of application showing an enhanced live view of what the camera is capturing (e.g., a magnifier app).
Finding a missing child in the crowd is a crucial and non trivial task. This project aims to reduce the burden on desparate parents and alleviate their unpleasant experience. Parents will have to share a few recent photos of the missing child with a cloudlet server and ask for help from the crowd around the area where the child is believed to be missing in. The cloudlet will build a classifier from the child’s images and use it to match faces on videos and images coming from the crowd’s input. Once a match is detected the parents will be notified of where the child is last seen. The processing must be super quick as the child might move far from the current location until detection is confirmed. Alternative (extention): upon confirmed detection, the cloudlets might also provide a tracking functionality of the child.
Tasks: Build a face recognition model (from the child’s images) Stream video from the Google glass to the cloudlet server Nexus 6 sends scene images each 10 secs to the cloudlet server Find the child in the video frames or images coming from the Glass or the phone Send a notification (with the location) to the parent’s smartphone (another Nexus 6) Hint: Accuracy is important but not a major deal in the project – the focus is on how fast the process would be (with reasonable accuracy).
Used together, mobile/wearable sensor readings and task/time/motion scripts could become a powerful yet straightforward way to detect and log a person’s activities. As an early step, we want to develop an interactive system for associating those two types of data.
This project has four main pieces:
- Design and implement a scheme for a user to specify his/her activities in terms of a sequence of timed actions. It should allow for recursive decomposition and for re-use of “atomic” actions (e.g., sitting down, lifting an object, opening a door, walking between refrigerator and stove, etc.). This might be captured in or transferred to a calendar program for convenience.
- Design and implement a scheme for interrogating the user* via his/her smartwatch about the task that is about to start, currently being executed, or just completed. It could be an inquiry with a limited number of responses or a simple yes/no confirmation.
- Design and implement a scheme to associate smartwatch sensor data with the tasks as logged via EMA. Doing so establishes a labeled data set for a supervised machine learning algorithm (see below).
- Design and implement a machine learning approach to automatically identify some actions or whole tasks based on the smartwatch sensor data. As a bonus, upgrade the user interrogation scheme to do EMAs for confirming/refuting the identified actions/tasks.)
(Note: the general term for this form of user interaction is “ecological momentary assessment” or EMA. There is a substantial literature on the technique to which the course instructors have contributed. A good introductory slide set and paper are available here: https://www.dropbox.com/sh/76x5uaekcd5xo38/AADg2KOeVTmxePK5bOzAk6Nsa?dl=0)
Blind individuals often struggle to make geographic sense of unfamiliar locations. Tactile maps can provide information about an area, but are typically passive objects with no awareness of the user’s current location. Smartphones with location-based navigation apps using text-to-speech functionality can direct individuals from point A to point B, they provide little context or sense of place. This project will explore how tactile maps and location based apps can be combined to provide a better real-time navigation system.
Students will use existing tools (www.tactilemaps.net http://www.tactilemaps.net) to create tactile maps. They will explore strategies for embedding touch responsive regions in the maps for use with a companion smartphone application. They will explore how to leverage the smartphone’s location sensors, audio and haptic IO capabilities to provide relevant and useful feedback to the users.
Most of the work itself will be in designing an Android application, though some amount of 3D printing will be required.
Blood pressure (BP) is one of the key measurements used to monitor patient health. The operation of current blood pressure monitors are usually described in manual with small font, unclear images, and complex interactions. These manuals, often over 30 pages long for even basic BP monitors, also contain several pages of error codes in an attempt to inform the patient of operational mistakes.
Glassistant is a Google Glass based vision application to guide the use of a wrist worn BP monitor. The current prototype, based on cloudlet infrastructure, uses vision processing to determine if the BP monitor is correctly positioned (right or left wrist, bottom of wrist, buttons facing user, proper elevation with respect to the heart). Voice commands and visuals are used for each step in the process. Machine vision techniques have been used to understand room light level, background clutter, detecting device orientation, detecting the device screen, and using optical character recognition to extract the contents of the screen.
The project will focus on developing and implementing a protocol to instruct and track the use of the BP monitor, to recognize errors, and to guide corrective actions. Usability evaluation will consist of a pilot study.
Gait variability, defined as fluctuations in gait characteristics from one step to the next, has been estimated to affect 35-45% of community-dwelling ambulatory older adults. Greater gait variability is related to less confidence in walking and lower levels of daily physical activity, and it is an independent predictor of falls and future mobility disability.
While slow gait speed may be a global indicator of balance problems in walking, altered timing of steps (gait variability) and abnormal postures (biomechanical abnormalities) signal the onset of mobility disability. Sensors available for this project include a Kinect, a video camera, and an Inertial Measurement Unit (IMU – includes accelerometers and gyroscope).
Exercises are used to clinically assess gait. For example a typical exercise, the Sit to Stand, measures how long it takes a person to sit down then stand up a certain number of times. The Sit to Stand exercise can be measured with a Kinect while the Get Up and Go (a timed walk) exercise can be measured by the IMU. Using these typical gait exercises we will develop a gate analysis and abnormalities detection algorithm. The algorithm will be evaluated in a pilot study using mechanisms to vary gait such as adding a weight to one leg.
Using an Automatic External Defibrillator (AED) is a time-critical task. If done correctly, it can save many lives. However, there is currently no easy way to guide a novice user through the procedure without on-site support from trained personnel. But Wearable Cognitive Assistants can change this. With wearable devices like Google Glass, it is possible to continuously capture what the user is looking at. An assistance system for tasks like using an AED, can be built on top of this such that 1) The system tries to understand the user’s progress with computer vision techniques; 2) The cognitive assistant provides step-by- step guidance to the user; and 3) The system gives user feedback based on progress. The video captured by Glass will be streamed to a Cloudlet and processed in real-time.
This project requires some computer vision background (if you do not know anything, you can learn along the way). We have a Gabriel platform that takes care of communications between mobile clients and cloudlets. You’ll mainly be using python to program cognitive processing for the AED. Small customization for Android client may also be needed. It also involves Android programming (on Google Glass). The real-time video transmission part is already built and open-sourced.
Description same as 1a
Consumer acceptance of IoT deployment is shadowed by privacy concerns. Users lack control over raw data that is directly streamed from sensors to the cloud. Current cloud-based IoT architecture does not have a clean method to let users filter and retain their sensitive data. A vision of using cloudlets to preserve privacy has been proposed. Imagine the following scenario in the future. You’re in a room talking with friends. Microphones in your smartphones and security cameras are recording your conversation and sending them to the cloud for analysis. Although you rely on cloud audio analysis services to automatically take note and set up reminders from conversations, you want to exclude sensitive information, for example, salaries and social security numbers, from raw audio data that is uploaded. You leverage cloudlets as privacy mediators to preserve your privacy. Audio data is transmitted to a trusted cloudlet before going into the cloud. On the cloudlet, audio analysis is performed in real-time to filter out sensitive information. For example, the cloudlet changes the audio containing the email password you said during a conversation into beeping.
In this project, you’ll build such a privacy mediator for audio data using cloudlets. We already have a framework named Gabriel that takes care of data transmissions between mobile clients and cloudlets. You’ll focus on audio data analysis on cloudlets. There are a few widely-used open-source automatic speech recognition frameworks you can leverage (CMU Sphinx, Kaldi, etc).
One likely deployment model for Internet of Things is to have centralized hubs that can offer devices network connections, check for firmware or software updates, and monitor traffic for anomalous behaviors. This project seeks to develop new ways of adding new devices to this hub in a simple and secure manner, as well as offering new kinds of services, such as linking different devices together or doing simple kinds of end-user programming.
This project will use a wearable biomonitoring device that is adhesively mounted to the hand to estimate position and monitor the user’s heart rate and blood oxygen saturation. The device will be used for stroke rehabilitation to understand when the hand becomes impeded due to muscle stiffening. Specifically, the project will aim to answer the following questions: 1) What is the position of the hand when muscle stiffening occurs, 2) what motion of the hand initiated muscle stiffening, and 3) how was the muscle stiffening alleviated. In addition, the user’s heart rate and blood oxygen saturation will be monitored at the fingertip. For this project, data from the wearable device will be transmitted to a smart phone via Bluetooth link. The cloudlet infrastructure will be used to off-load data processing and storage. An interactive user interface will display relevant biomonitoring signals and suggests ways to prevent or alleviate muscle stiffening.
In the Internet of Things (IoT) applications are based on large amounts of sensor data and making that data usable to everyday persons is becoming a challenge as well. The project uses context-aware and Internet of Things computing technology to aid care coordinators in keeping their patients healthy, happy, independent and safe in their respective homes. Taking care of elderly persons is becoming a real challenge in many countries. The system will (1) allow to view and add their patients’ information, (2) provide some data analytics, (3) gather medical and social/emotional data to provide a holistic view of the health status of patients, and (4) provide alerts and notifications when a patient deviates from their baseline health. We combine mobile and stationary sensors, as well as EMA (Ecological Momentary Assessment) surveys for parameters that cannot be sensed (e.g., social activity, mood). We will aim to model people’s sleep patterns, physical activities, stress levels and social activities, showing end users details of their own behaviors and offering the community aggregated summaries. These technologies should enable self-monitoring and sharing of progress with healthcare providers.
The Coda Distributed File System (Coda) has multiple features that make it desirable in poorly connected environments. By its aggressive, persistent, whole-file caching strategy and ability to continue read/write operation even when the Coda servers are unreachable it makes it very well suited to provide file services for cloudlets that have possibly intermittend connectivity which are available close to a mobile end user. However with several modern workloads such as MapReduce or video stream analysis that work with large, append only type files, the whole file caching strategy makes it inefficient to work with large files because we possibly send a multi-GB file back to the server multiple times. Several things need to improve for Coda to tackle such large file more effectively. First of all, in the existing design Coda has no insight into individual read and write operations. By adding support for FUSE either directly or through a proxy process it would be possible to observe which parts of a file are updated or accessed. Secondary there has to be some sort of immutable file storage on the server that allows clients to fetch data from an older version of a file (to be able to maintain the existing open-close consistency model. This could be implemented as an S3(-compatible) storage pool, or possibly a more efficient delta packing format, similar to git packfiles.
From there on, there are many possible directions this work can take. On the write path it now becomes possible to track which parts of a file have actually been modified, and a binary delta can be generated and sent to the server instead of the whole file. On the read path it will be possible to fetch file contents on-demand as they are being accessed, with the caveat that we will probably lose performance and our ability to survive network failures because now only fragments of files may been cached. Implementing this may require some kernel level work, deep knowledge of C/C++ programming and a lot of careful thinking about consistency in a distributed system.
Independent Study: Virtual Window Shopping on Business Streets (Video)
Student: Student: Tan Li Mentor: Zhuo Chen
Independent Study: Edge Computing for Live Video Style Transfer (Video)
Student: Shilpa George Mentor: Satya
4. Just In Time Sign Translation (Video)
Students: Anagh Singh and Vignesh Shankar Mentor: Junjue Wang
Travelers in a foreign language environment often have troubles navigating through surroundings. Sometimes failure to navigate can result in severe consequences, for example, going through wrong declaration lanes at Customs in airports. More often, it causes inconvenience and frustration, for instance, missing a highway exit. This project aims to perform just-in-time sign translation in places including airports, highway, and hotels, to help travelers navigate in a foreign language environment.
Leveraging a head-mounted smart glasses (e.g. Google Glass), the system translates phrases on signs as a user looks at them and displays the translation on the screen. The function is similar to Google and Translate, but the project should reduce end-to-end latency significantly leveraging cloudlets, since a user is likely to be walking or even driving as she passes through a sign. To build such a system, OCR is needed to convert images of text into text. Machine translation is also required to convert the text into another language. Fortunately, there are well established open-source projects for both OCR (tesseract-ocr) and machine translation (OpenNMT). The source and target language can be fixed to make the machine translation easier. The focus of this project is on overall system design and implementation, not on improving OCR nor machine translation.
8a. Opportunistic Search of Edge Sourced Data (Video)
Students: Sandeep Agarwalla and Sindhu Simhadri Mentor: Jan Harkes
Cloudlets receive multimedia data feeds from IoT sensors, video cameras, drones and other data sources that they are associated with. Searching this data at Internet scale with relatively low result latency will become important in the future. The goal of this project is to create a search framework for edge-sourced data that is based on the publish-subscribe paradigm. Cloudlets can subscribe or poll for any appropriately scoped searches, run the search and send results back to the cloud.
8b. Improving Search Efficiency In Edge Computing (Video)
Students: Ziqiang Feng and Yanzhe Yang Mentor: Jan Harkes
Some edge computing nodes (e.g., on a small, autonomous drone) may be under-powered, and incapable of running expensive algorithms such as those based on convolutional neural networks. The cloud has the necessary resources, but shipping all the data from the edge to the cloud is wasteful of bandwidth. However, the edge node could run a cheap “early-discard” algorithm that has a higher false positive rate than the best algorithm. The goal of this project is to explore this tradeoff space and to demonstrate a search capability based on these principles
9. ActiveMaps (Video)
Students: Ticha Sethapakdi, Diana Zhang Mentor: Jason Hong
The goal of ActiveMap is to make it easy for people to quickly create and deploy many kinds of location-based services. As an analogy, just as a web browser can read and render web pages, we want the ActiveMap to be able to read and render Maps. Example Maps might include a map of a location (e.g. a building floorplan), or content about where static or dynamic objects are (e.g. nearby Starbucks and discounts, real-time location of a person or members of a family, or location of buses as they move about). Systems issues include specifying the Map functionality and syntax (e.g. defining floorplans, defining locations inside, update rates, diffs from last read version, text snippets to show), creating a GUI to display and interact with multiple Maps, and creating backend software for making these Maps easy to deploy on web servers
For individuals who undergo partial arm amputations, robotic myoelectric prosthetic devices can allow them to recover a great deal of their arm and hand functionality. A significant challenge in adapting to a prosthetic device is learning to use their brain to control the device effectively and safely.
In this project, we will use a Microsoft Kinect and a skin EMG reader to help provide feedback to users learning to use a prosthetic device. Participants in this project will develop machine learning tools to determine what feedback to provide to a user performing physical therapy exercises to help them learn to use their prosthetic device correctly. Example exercises are: lifting a light object, lifting a heavy object, passing an object from one hand to the other, and lifting a tray. Using the Unity game engine we have developed three 2-dimensional games that users can control using the EMG sleeve, as well as two virtual reality games.
Data was collected from 12 volunteers, play-testing these games for 10-20 minutes at a time. Additionally several subjects performed activities of daily living, such as passing an object from one hand to the other, while recording simultaneous Kinect and EMG data. For this dataset we collected 233 instances of an activity of daily living, specifically the act of lifting a light object from one hand to the other.
With these provided datasets, we would like to address the following machine learning classification tasks. The first is to identify the type of muscular activity a user is performing given 8 channels of EMG data (e.g. wrist extension or arm rotation). The second is to identify if a subject is correctly performing a physical task, such as transferring an object from one hand to the other, given EMG data coupled with Kinect 2 depth and RGB data.
Augmented reality / mixed reality is an emerging technology that may revolutionize mobile gaming. The idea is to mix elements of the real world and the user’s movements and actions in the real world, along with elements of a virtual world to produce an immersive gaming environment. For example a game may place a virtual monster or treasure chest at some real location, with which the user can interact. The virtual elements should be displayed and moved consistently with the real world surrounding as the user moves. To do this well will require a reasonably powerful device with cameras, sensors, and displays, along with cloudlets to do the heavy computational steps. Although a complete game is beyond the scope of a semester-long project, several projects can be defined to demonstrate various aspects of AR gaming. All of the projects will use an Android device as the front end (maybe with Cardboard, or VR headset adapter), and use a Linux-based cloudlet for computational offload. OpenGL will be used to display mixed reality scenes, and a combination of OpenCV and other visual libraries, along with custom code (C++ or Python) will be used on the cloudlet. In a mixed reality game, the user will interact with objects and environment with their hands. We need a way to detect movements of the arm and interpret actions. Potential demo: user aims and shoots at AR targets using virtual gun held in hand.
Objects help people remember. Many things are valuable because they carry part of the past. People recall emotions and experiences when they see these objects. It could be a tarnished birthday card which reminds one of an old friend. It could be a magnet buried inside of a dusty box that reminds one of a past trip. While people’s memory could fade, digital records do not. What if there is a wearable time machine that could help people relive their past experiences?
Leveraging head-mounted smart glasses (e.g. Google Glass and Microsoft HoloLens), this project aims to build an object detection-based system that helps people relive their past experiences. The application displays short video clips from the past to users through head-mounted smart glasses when users see special objects. To create such an experience, the application would record short video segments throughout a day using the smart glasses. Users or the application itself would mark some objects as “memory triggers”. The application builds object detectors for these objects and associates video segments with them. Then, when the application detects a memory trigger, it retrieves and displays relevant video segments that are associated with the object to augment memory recollection and help users relive their past.
This project provides you with an opportunity to work with smart glasses, cloudlets, and deep neural networks (DNNs). You will learn through practice how to design and build a real-time video streaming and analysis application using deep learning based object detection. Depending on your interests, features can be changed. Familiarity with DNNs is preferred but not required.
Recent developments in deep learning have made it possible to automatically alter and synthesize realistic images. One interesting application is Face Swap, which superimposes a person’s face, including facial movements, to another person in a natural-looking way. Existing open-source projects leverage autoencoders and generative adversarial networks to achieve such effects. However, they require significant computation power and the processing happens offline. The project aims to build a real-time deep face swap application on mobile devices by offloading computation to a cloudlet, a small data-center that is one wireless hop away from the mobile device. The application would stream the camera feed from the mobile device, perform face swap on the cloudlet using deep neural networks (DNNs), and transmit the altered video stream back to the mobile device for display.
This project provides you with an opportunity to work with cloudlets and DNNs. You will learn through practice how to design and build a real-time video streaming and deep learning-based analysis application. Depending on your interests, features can be changed. To make existing face swap DNNs run in real-time can be challenging. Strong system optimization skills and familiarity with DNNs are preferred.
A child has gone missing. An AMBER alert is issued. The authority launches a visual query including face recognition over a network of public surveillance cameras. The story so far is well understood.
Now, some good citizens offer to help by making their personal video feeds searchable – dash cameras, smart glasses, smartphones, recreational drones, home surveillance, etc. On one hand, these citizens may join the search at any point of time. On the other hand, they may leave the search at any point of time, when, for example, they enter private zones or run low on battery. So, these citizens should be able to come and go easily when the search is in progress.
In this project, you will develop a system that realizes the above vision. You are encouraged to base your project on OpenDiamond (https://github.com/cmusatyalab/opendiamond ) but it is not mandatory. OpenDiamond is a system for searching non-indexed data on an edge computing infrastructure. Although this project involves searching image or video data, you are not required to have prior computer vision knowledge. OpenDiamond comes with a number of visualfilters (e.g., RGB histogram, SIFT matching, DNN) that you can reuse off-the-shelf.
What you will learn:
- Formalize the design requirements of a system from a motivation
- Design an execution model that can especially facilitates the required agility
- Using VM technologies (e.g., Docker container) in a larger system
- Tradeoff between data shipping and command shipping
What you need to already know:
- Programming with Python, JSON, XML, etc.
- Basic programming with networked systems (e.g., TCP sockets, Flask)
- Concepts about remote procedure calls (RPC)
This is variant A of the project described above. This project focuses on de-coupling edge nodes’ participation time from a search’s life time. Specifically, different edge nodes may join or leave the search at arbitrary times. The front-end should display a stream of search results without any interruption whenever new nodes join the search or a node stops sharing data.
This is variant B of the above project. This project focuses on handling the heterogeneity of participating edge nodes. While some powerful edge nodes (e.g., dash camera with an on-vehicle cloudlet) may be able to run expensive search pipelines (e.g., including a DNN), other weaker edge nodes (e.g., drones) may not. The students need to develop a systematic and principled way such that all edge devices can contribute to the same search, but possibly undertaking different processing onboard.
1. Automotive Model Assembly Assistants with AR Feedback (Video)
Student: Chinedu Ojukwu Mentor: Junjue Wang
Manual assembly still plays a crucial role in automotive production despite increasing automation. With the human in the loop comes two challenges: 1. Novices need to be trained for sophisticated tasks and closely supervised 2. Even with experienced workers, human error is still a major source of lost productivity in manufacturing. Wearable cognitive assistants (WCAs) are a potential solution to these challenges. WCAs guide users step-by-step through a task, catch mistakes that a user makes, and tells the user how to fix them. Previous wearable cognitive assistants have been developed for several assembly tasks, including Disk Tray Assembly and Sandwich Models.
In this project, you will build upon existing works to create an Automotive Model Assembly Assistants with AR-style feedback. You will work on Tamiya car models, which are miniature versions of racing cars and its assembly process resembles car assembly in a factory. Previous work has created multiple object detectors to [identify user states] (https://github.com/montiblanc97/gabriel-car). This project will build upon the previous work to create assistants for the entire assembly procedure and provide AR-style feedback by overlaying virtual object onto the physical ones. You will explore and gain insights on how to push existing AR applications to incorporate semantic visual understanding and how to effectively communicate instructions using AR.
3: Speculative wizard for edge-based interactive image search (Video)
Student: Haithem Turki Mentor: Ziqiang Feng
Eureka is a system that enables a user to perform content-based interactive search of unindexed image data using filters. It trades off compute cycles for better use of a human user’s time. This project pushes that concept a bit further. When running a search, the “best” filters to use are unknown, so the user usually does trial-and-error. You are asked to add new features to Eureka that speculatively “try out” new filters and make suggestions. The basic idea is that, when the user labels the returned candidates, it creates a quasi test set, which can be used to guide the suggestion. With speculative execution of new filters using additional compute cycles, it has great potential to further improve a user’s productivity.
5. Cognitive Assistant for Training in Additive Manufacturing (Video)
Student: Tianyu Gu Mentor: Asim Smailagic and Dan Siewiorek
The project will develop a Cognitive Assistant that will guide a user with only a limited amount of previous training by a human operator through the steps necessary to safely and successfully operate metal additive manufacturing equipment. We will accomplish this using a laser power bed 3D printer set up. The Cognitive Assistant will be developed and tested. Because its architecture allows cognitive engines to use a variety of programming frameworks and operating systems, expansion to support operation of other 3D printers will be possible.
We have done some early work in this area, including virtual coaches and automated help-desk with remote expert paradigm that the project can benefit from and create new value. We have created and built over a dozen of different virtual coaches. Our Virtual Coaches represent a new generation of attentive personalized systems that can continuously monitor its client’s activities and surroundings, detect situations where intervention would be desirable, offer prompt assistance, and provide appropriate feedback and encouragement. Virtual Coaches are intended to augment, supplement and, in some circumstances, be a substitute for an expert by offering guidance and help.
Sometimes a worker requires assistance from experienced personnel. Apprenticeship programs let novices learn by observing and working with experienced workers. More recently, help desks have evolved to provide audio and visual access to experienced people for help with problem solving. The help desk can service many field workers simultaneously. Additive manufacturing machine operators perform a number of procedures that are repeated including setup of the equipment in preparation for a build, cleaning of the equipment after a build, filter changes, routine maintenance of the equipment and a number of other regular operations. To support the equipment operator, the Virtual Coach will have the tasks and documentation needed for these procedures such as text and schematic drawings. Because it is centrally maintained, even if this information changes daily or hourly, workers still get accurate information. Sometimes a worker requires assistance from experienced personnel. Apprenticeship programs let novices learn by observing and working with experienced workers. More recently, help desks have evolved to provide audio and visual access to experienced people for help with problem solving. The help desk can service many field workers simultaneously.
7: Multi-modal mixed-initiative AR (Poster, Video)
Student: Fanglin Chen Mentor: Junjue Wang
Guest Demo (Verizon): Mobile Edge Computing for Computer Vision (Video)
Guests: Verizon Researchers
Guest Demo (InterDigital); AdvantEDGE Platform (Video)
Guests: InterDigital Researchers
1. Low-Bandwidth Video Calling (Poster)
Student: Ananya Joshi Mentor: Roger Iyengar
The Covid-19 pandemic has forced many people to conduct meetings over video calls, using products like Zoom and FaceTime. This allows participants to see each others’ faces, unlike traditional phone calls. Displaying faces enables nonverbal communication in the form of facial expressions. However, transmitting video requires substantially more bandwidth than audio. In this project, you will leverage edge computing to build a communication application that shows facial expressions but uses much less bandwidth than video calling.
A client will transmit video to a cloudlet, which will extract information about the facial expression of the person in the video. The cloudlet will then send this facial expression information to the other users. The other user’s devices will synthesize an avatar depicting the original facial expression. Transmitting facial expression information requires less bandwidth than transmitting whole video frames. You will run experiments to quantify the true bandwidth savings. Information about people’s facial expressions will be extracted using the Azure Face Cognitive Service, which can be run on cloudlets using Docker.
Working with an expert to complete a task is much more pleasant than trying to follow written instructions by yourself. Products such as Ikea furniture and children’s toys require assembly by end users who have never done these tasks before. If someone makes a mistake when following the instructions, they might not realize it until several steps later. However, an expert can immediately tell somebody whether they have completed a step correctly or not.
In this project, you will create an automated system that acts as an expert for an assembly task. The user will wear an AR headset with a camera, and processing will be done using a cloudlet. Your system will leverage several research tools for wearable cognitive assistance applications and computer vision models. The following videos depict similar applications:
The goal of this project is to build an app and edge service to use your phone like a pair of binoculars. The idea is to take the live camera view and zoom in to show an enlarged view of the scene. At high zoom levels, the image will become blocky or blurry, and (when the phone is hand-held) very shaky. To address these issues, we will build an edge service that will take the video stream from the device, stabilize it, and perform super-resolution processing, allowing one to present an output image at a higher resolution than the original camera image.
Various techniques for super-resolution have been developed in the past 20 years. Early techniques used multiple, slightly-shifted images (like those from a shaky camera phone) to mathematically hypothesize the high-resolution source that would best explain the images seen. More recent techniques employ deep-learning models that have been trained to hallucinate plausible high resolution images from lower resolution ones, and can be used with just a single source image. Any of these techniques can be used, depending on student interest (classical image processing or Deep Learning) and availability of code. All of these are computationally intensive, and will need computational power of the edge to support the mobile device.
A final demo will show a live view on the phone, comparing a simple zoomed in view with the stabilized, super-resolution view. Depending on student interest, this could be a custom-built Android application, or it could be a modified version of the Gabriel client built in Satya’s research group. Or the whole project could be http-based, and the “application” can just be a browser-based interface.
Previous efforts in computer vision help to identify and segment common objects in a collection of images or videos. Example of such work are:
- For Video: SiamMask ( paper, demo video), and code) is used to detect and segment objects from videos in real time, from a single bounding box annotation and produce segmentation mask and bounding boxes as output.
- For Images: Deep object co-segmentation (code), segments common objects (ignoring background such as sky, etc) of the same class within a pair of images.
A user can upload such a collection of images to Delphi to learn a just in time object detector. Mask RCNN pretrained on COCO can be customized to have good performance using just 70 images (model AUC of ~0.6). Delphi notifies the user once the training of the model is completed (~15 min on a V100 GPU). The DNN can then be deployed for real time detection.
During an assembly task, a user can collect and upload videos of a part he wants to detect. The user then uploads the collected images or videos to Delphi and trains a new detector for the object of interest. With minimal labeling (one bounding box per video for SiamMask), the user can obtain ground truth annotations for the required object. She can start the training session in Delphi and once done, deploy the model for inferencing.
Requires knowledge of Android SDK development, minimal computer vision. Computer vision codes required are publicly available on github.
The goal of this project is to create an app that runs on a smartphone, leverages cloudlet resources over a wireless link, and walks users through a physical assembly task. The app will use computer vision to determine when steps of the task have been completed. The phone offloads images to a cloudlet for processing. You can see an example of such an app here. To keep the workload of the project manageable, your app will likely only be able to detect a modest number of steps. However, this project will give you hands-on experience in creating an edge-native application and a good sense for the challenges of getting computer vision models to work well in real systems.
Microsoft Seeing AI offers a feature that captures a picture with a phone camera and describes the scene in this photo to a blind person. You will implement this feature yourself using publicly available machine learning models. You can get creative about which models you chose to use, and how you interpret the model’s output. For example, you can use an object detector that will tell you labels and bounding box coordinates for objects in the image. But then how do you summarize the spatial relations of these objects as text? Are there models that can provide more general information (such as “this is a bedroom” rather than just “there is a bed, a dresser, and a lamp”)?
Recent advances in neural rendering, such as neural radiance fields (NeRFs), open a promising new avenue to model arbitrary objects and scenes in 3D from a set of calibrated images. NeRFs can faithfully render detailed scenes from any view while simultaneously offering a high degree of compression in terms of storage. However, up until recently, runtime performance was a critical limitation of NeRFs which can take tens of seconds to minutes to render a single image. A recent follow-up work proposed a framework that enables real-time rendering of NeRFs, which expands the frontier of NeRF applications into use cases such as VR/AR.
Your project will be to leverage this work to build an end-to-end pipeline that allows a user to capture a video of an object of interest, train a NeRF at the edge, and overlay the rendered 3D representation onto an AR headset. This project requires proficiency in both systems building and computer vision. Although you are not required to invent novel computer vision algorithms, you should be comfortable working with deep learning code and frameworks such as PyTorch.
This project will explore mobile augmented reality using edge computing. The project team will create an edge-native application for a mobile pedestrian heads-up-display (HUD = e.g., Hololens/Google Glass) user. The application will overlay historical images on a real-world outdoor scene during a walk across the CMU campus. For example, see these historic CMU images).
This project will explore the sources of latency in 4G and 5G mobile networks. The project team will design and implement a framework to measure where latency arises as user traffic passes from the user device to the edge, the cloud and back to the device. The main focus will be on segmenting and measuring the latency introduced by each of the wireless link, the radio access network and the wireless core network. The framework will be built as much as possible from off-the-shelf components (e.g., wireshark) and tested using a real application (e.g., OpenRTIST) in a real environment (The Living Edge Lab). The project will produce the framework, Living Edge Lab measurements using the framework and an analysis of the sources of latency in the Living Edge Lab. The project will require knowledge of mobile network architecture, networking, system measurement and benchmarking.
Detecting fake news is an important problem. Fundamentally, the detection part requires data that leads to a privacy concen. We want to detect fake news in a way that is privacy-promoting, meaning that the platform never sees user data. We use federated learning to achieve this. Each individual user trains his own model and then sends that model to the platform, which combines its inputs to create a global model. User data never leaves the user device. Can fake news detection be done in a low-bandwidth setting using federated learning? Answering this question is the goal of this project.
OpenScout is a pipeline for automated object detection/facial recognition. Android clients send image frames and GPS coordinates from the device to an GPU-enabled edge node where two cognitive engines perform object detection (using objects from the COCO dataset) and face recognition (using either OpenFace or Microsoft’s Face Cognitive Service). Results are pushed into an ELK (elasticsearch-logstash-kibana) stack for visualization and analysis. This project would consist of adding new cognitive engines to perform functions like OCR on any text in the image stream or perform pose estimation on the actors in the scene. Microsoft has an OCR cognitive service that could be leveraged, however another OCR framework could be proposed. Pose estimation could be done with OpenPose.
OpenRTiST is a real-time style transfer application that takes image frames from the device and returns images that are stylized to resemble famous works of art which are then displayed on the user’s screen. Generative adversarial networks have been used to perform unpaired image to image translation. This project would extend OpenRTiST to create a new cognitive engine using CycleGA or CUT to add new styles via unpaired image to image transfers such as winter to summer or day to night.
Google recently released AR-based walking directions. This uses training on Google’s streetview image dataset to get more accurate location information than what you would get from just the phone’s GPS. The goal of this project is to create a similar system based on Deep Neural Network object detectors and classic SIFT feature extractors, and to compare their effectiveness as a potential basis for an AR application for exploring indoor points of interest (POI). GPS is not available indoors.
The initial goal would be to capture video walking across campus, and collect seasonal image data of places like the fence and other recognizable landmarks and then labeling the various POI with CVAT, the computer vision annotation toolkit. The labeled datasets can then be used to train object recognizers with the OpenTPOD pipeline (tool for painless object detection) as well as the classic SIFT/SURF feature detectors. With this the relative effectiveness of both approaches can be evaluated. How robust are these methods when faces with seasonal or weather changes, or when students repaint the fence? A stretch goal would be to turn this into an Android application, by building on the Gabriel framework which captures image data from Android devices, offloads to the cloud, or a nearby cloudlet, for any heavy computation and returns annotated frames with markers for any POI that were found within the image. An example Gabriel application is OpenRTiST which performs real-time style transfer on the input images.
A virtual coach that monitors and provides feedback to rehabilitation exercises has the great potential to improve patient compliance with therapy. Development of these systems requires a setup of a motion capture sensor. The goal of this project is to use interaction techniques and edge computing to create a more capable virtual coach to monitor and provide intelligent feedback to patient exercise performance. The coach would use a web camera and/or IMU wearable sensor. The project would include two main parts: (1) Interaction Techniques: Explore more advanced AI techniques to provide transparent feedback to the user; Develop audio or gesture recognition models to control a system; (2) Edge Computing: Create a system to transmit a video/image of an exercise to a server and receive tracked body joints using the OpenPose library; evaluate the effectiveness of edge computing techniques in providing real-time feedback. For the development of this system, we have collected a dataset that includes video recordings of three upper-limb exercises from 15 post-stroke survivors. This project requires proficiency in implementing machine learning models and interactive systems (e.g. using audio or gesture modality).
Sinfonia is a system for mobile applications to discover and deploy compute intensive operations on nearby infrastructure. The focus of this project is to improve the decision making in Sinfonia about what infrastructure can be considered ‘near’ or ‘sufficient’ based on available metrics and bringing in additional metrics to allow us to make better decisions. Examples would be:
Currently we just report % used, explore if better decisions can be made with the USE method (Utilization, Saturation, Errors) for each resource.
Include additional resource metrics that are currently missing, f.i. disk I/O capacity and utilization.
Expose Docker cache state, f.i. by using a bloom filter populated with the hashes of cached images.
Include additional network related information to help estimate which nodes may require fewer network hops, f.i. direct Verizon mobile phones in specific ‘5G Edge regions’ to cloudlets in their respective AWS Wavelength zone. Or have latency metrics to significant nearby network peering points, mobile operator / ISP. specific egress points.
Granting edge connectivity to an autonomous drone allows us to get the capabilities of a traditional autonomous drone without carrying a heavy compute payload. Instead, the drone offloads computation on data from its sensors to a nearby cloudlet. For applications that do not have tight latency requirements (e.g., detecting objects along the drone’s flightpath and taking pictures at various GPS locations) our design performs comparably. For more dynamic tasks, high latency can make things difficult. In this project, you will develop an edge-based application that performs latency-resistant (up to 1s) tracking on drone video feeds. Your application should be able to account for transmission delays and sudden changes in the motion of the target.
Skills learned: drone simulation using Parrot Sphinx, DNNs, and tracking algorithms
Coding language: Python
With world health crises such as monkeypox and COVID-19, there have been closures of fitting rooms to minimize interactions. DressUp is an AR application that allows consumers to virtually “try on” clothing of their choice. Instead of simply superimposing the user’s face on a mannequin, the application uses state-of-the-art deep learning techniques to estimate the pose of the user and warps the clothing to generate a photo-realistic fitting result. The app leverages edge computing to obtain seamless interaction which is not attainable if run on the mobile-device or at the cloud.
When a mobile user moves location, they expect a seamless application experience. But, movement can also mean migration of application functionality and state to a different edge cloudlet. This project will explore the triggers, methods, and performance impacts of application migration between edge nodes. The team will create or modify a stateful edge-native application, develop a migration prototype, and test the function and performance of that application during mobile user movement in the real world Living Edge Lab network. This project can build off our existing work in application migration (e.g., nephele) and orchestration (Sinfonia).
Edge computing can supplement the mobile device computing power by offloading application functionality to a cloudlet. In this project, you will use this approach to create an edge-native media composition application that will automatically create short videos similar to IG Reels and Tik Tok. You will apply multiple machine learning models like body tracking, hand tracking and face tracking on the edge for real time video capture and compositing to produce a single video from a set of short clips. You may make use of the Living Edge Lab “OpenScout” platform for your edge-native application. You will test your application on the Living Edge Lab mobile network. Experience in Android application development and machine learning are important.
The possibility to evaluate and even train large deep networks on mobile devices significantly enhances the potential applications for mobile and pervasive computing, but standard deep networks are not designed for mobile computing. In this project with significant machine learning content, we will develop a saliency-based layer and channel pruning method to convert deep convolutional networks into more efficient mobile-ready models. We will prune, quantize and compress models for use on mobile devices, and then evaluate them in classification tasks with medical image data. We will provide the code and method for a deep network channel pruning and a saliency-based explanation of a deep network’s inefficiencies. You will be expected to build on this to develop an improved deep network channel and layer pruning method, and also to implement a quantization method to obtain a minimally compressed model. The model will be compared to existing literature.
Experience in linear algebra, deep learning, pruning, quantization and design/application of mobile deep networks is desirable
Keywords: Deep Network Compression, Pruning, Quantization, Medical Image Analysis
There are multiple devices that have to communicate data wirelessly and efficiently in IoT systems. Of the different kinds of wireless networks available, LPWANs (Low-Power Wide Area Networks) are particularly useful due to their ability to support numerous streams of small data packets coming from devices over a large area. However, increased capacity and overhead when upscaling LPWANs become a key issue. This project aims to build upon recent work on LPWANs at CMU (WiSE Lab). The work explored packet recovery using multiple receivers to correct corrupted data, as well as the use of directional antenna arrays in order to reduce interference among clients.
We will use the cloudlet to bring these sensing devices closer to their data-centers, while also ensuring that the data received is coherent and intelligible. This project will attempt to integrate the use of the cloudlet system with this LPWAN infrastructure, utilizing the techniques of packet recovery and interference isolation to maximize performance. Our goal is to see improvements in successful packet delivery and network capacity by the end of the semester. The final deliverable will include a demonstration of the system to illustrate these performance improvements compared to published results. Implementation tools include using the LoRaWAN protocol, deployed on the appropriate hardware for testing.