From Interdrone 2015

People

  • Diane

Areas for Further Research

Crowd Analytics through Facial Recognition

Eyeris

Software package called ImoVu - Deep Learning architecture for image processes that imitates human recognition

  • Vision for Embedded Systems
  • 97% accurate
  • 140 FPS on i7 processor
  • Ambient Intelligence (AML) - Electronic environments that are sensitive and responsive to the presence of people.
  • Minimal required resolution for a face: 50x50px, Ideal 70x70px
  • Does not use FACS
  • Uses OpenCV open source software to identify faces and for some devices, sometimes falls back to the devices facial recognition
  • This software is not useful for things or places (only people)
  • Allows for
    • Access to Hard to reach places
    • Large coverage area
    • Eye Level
    • Improved angles


  • Trained For
    • 5 Major ethnic groups
    • 5 age groups
    • 2 genders
    • 10 lightning conditions
    • 13 head poses
    • supervised and unsupervised
    • different camera structures
    • different resolution
    • different face attributes
      • Joy
      • Surprise
      • Sadness
      • Disgust
      • Anger
      • Fear
      • Neutral

3D Mapping

  • mark@visuallaw.com
  • sketchfab - 3d modeling software
  • pointcloud
  • DroneDeploy - android app for creating 3d mesh and orthomap ($300 a month)
  • Pix4D - the primary software for creating 3 modeling of terrain
  • Adjisoft Photoscan (?) - Competitor to pix4d
  • Autodesk is working on comparable software

DroneKit

Flying in Wind: State Estimation for UAS

  • Douglas Weibel, PHD
  • Wind - a vector velocity field with spatial and temporal variation
  • Wind has structured (predictive) and random properties (turbulence) that are difficult to model
  • State
    • Attitude (Roll, Pitch, Yaw, Roll Rate, Pitch Rate, Yaw Rate)
    • Attitude Rates
    • Position
    • Velocity
  • Other possible states
    • Wind Paramters
    • Height over ground
    • Battery energy
    • gyro bias
    • accel bias
  • Attitude Estimation
    • Stabilized Sub Platform
    • Whabba's problem
    • Integration
    • Combined
  • Contempory Sensors
    • Gyros
    • Accelerometers
    • Magnetomers
    • GPS Baseline
    • Thermopile
    • Elecstratic potential
    • Opical sensors
  • Federated Filter
  • Attitude Filter ( Cascaded Filter )

Applications of Rotation Matrices and Matrix-Vector Algebra in Autopilots

Bill Premerlani

  • A rotation matrix is a linear homegeneous transformation
  • Tensor

FAA Regulations: Past, Present, and Future

  • Jim Williams
  • Drones are named from the first unmanned aircraft in Britain during WW1
  • [1]
  • FAA cant regulate anything congress doesn't give them authority over
  • Required for an aircraft to be regulated
    • Approved Pilot
    • Aircraft Certificate
  • Code of Federal Regulations (91.113) - you must be able to see and avoid other aircraft (thus-- you need a 333)
  • COWA - Certificate of Waiver or Authorization
  • AC - 91.57A new AC that regulates RC hobby
  • AC's are not regulatory
  • CFR 14 91.13 Careless and Reckless Operation
    • Aircraft operations for the purpose of air navigation. No person may operate an aircraft in a careless or reckless manner so as to endanger the life or property of another.
    • Don't do stupid things that put people in the air at risk.
  • 14 CFR Part 107 creates new rules for small unmanned aircraft systems
  • Examples
    • CNN is looking at tethered aircraft (tether provides power) - Power Electronics 01
    • Railroad wants UAV to monitor tracks in front of trains looking for issues with tracks
    • Farmers are looking for ways to safely use UAV without LOS by monitoring airspace for incoming commercial aircraft
  • FAA - Department of Transportation interactions take a lot of time
  • Most insurance policies will not cover accidents caused by commercial use of drones
  • Some companies with 333 exemptions are suing companies operating without the proper exemptions

Embedded Computer Vision for Safer, Faster Drones

  • Goksel Dedeoglu, Founder PercepTonic, LLC
  • Computer vision algorithms analyze images to extract information about the world (allowing machines that perceive)
    • Size and shape
    • 3d position
    • Identity
    • Expression
    • Location
    • Motion
    • Illumniation, weather, etc.
  • Two things relevant for drone operation:
    • Motion
    • Depth
  • Inertial sensors
    • Identifies key points
  • GerbilBall is a safety sphere enabled by real-time depth sensing
    • Stereo-based
    • Has limitations on how close the two objects can be to each other
    • Can have as many cameras as needed to create a realtime look in all directions (expensive)
    • For objects that don't move often, may be cheaper to build an internal model of what has been seen
  • Best guess for power efficient vision computation is heterogeneous and multi-core architechtures
  • Examples
    • Fire Phone makes good use of computer vision with multiple cameras
    • Dyson 360 Roomba
  • http://www.cvisioncentral.com/evw2015
  • http://www.embedded-vision.com

DroneKit

  • A single API that runs onboard drones
    • Cloud REST API
    • Android SDK ("3DR Services")
    • Python SDK
  • Dronekit exists as a compatibility layer on top of the autopilot
  • Onboard and offboard MAVLink
  • Droneshare.com
    • REST API (speaks JSON)
    • Used for storing fetching displaying logs and other data
  • Full autopilot control from mobile
  • Built on Android AIDL infrastructure
  • Examples
    • Tower, DroidPlanner, Tower Pebble, Tower Wear, Solo
  • DK Python
  • "Solo is the best tool for aerial video".
    • Runs on PixHawk 2 (unreleased)
    • SDK for Solo dev.3dr.com
      • Prototype in Python without low level system knowledge


  • Pixhawk
  • tim@3dr.com

Structural Inspections

  • 6 weeks to 190 days to get 333
  • Several of the participants have run into issues with RF interference. One person had a fly away in downtown LA. It was identified as interference caused by a nearby consumer WIFI device. One person recommended an RF analysis as part of the pre-flight.
  • Military-grade components have much better shielding and seem to handle interference better.

Human Factors

  • In order to keep from overwhelming the operator
    • Multimodal - Use things like sound, or LEDS, to indicate status
  • Multiple HSC Tasks = divided attention problem - most people don't multitask well
    • Encourage task switching efficiency
    • Minimize the switching costs
    • Ensure that the operator is focused on the appropriate tasks
      • Procedures
      • Structure based abstractions - colors, arrows, etc.
      • Automation Cueing
      • Cues/alerts should be appropriate and not overwhelming. There is a cognitive cost to processing cues. Don't make the problem worse.
      • Consider Interruption Lag cueing-- an early warning about something that will eventually become important (the upcoming block in tetris) - studies are not conclusive on effectivity
      • Increase role of automation