1 of 27

ADAPT: Drone Payload for Data Collection and Real-Time AI Processing in the Field

Ocean Sciences Meeting 2022

Website: https://kitware.github.io/adapt/

PI: Matt Brown, PhD, Kitware (matt.brown@kitware)

Co-PI: Peter Webley, PhD, U Alaska-Fairbanks

1

Supported by Federal funds under award NA21OAR0210112 from the National Oceanic and Atmospheric Administration, U.S. Department of Commerce

2 of 27

ADAPT Multi-Mission Payload

2

2

Image from https://www.weather.gov/otx/Our_Office

Oil Spill

An open source platform for deploying state of the art deep-neural-network computer vision in real time on small unmanned aircraft systems (sUAS).

ADAPT Multi-Mission Payload

3 of 27

The Problem: 2013 Galena Alaska flooding

3

  • $6 million in damage
  • 169 houses destroyed
  • Damage to public infrastructure
  • Flooding on the scale of hours
  • Residents had little to no warning

River 2km away

Ice blocks and debris the size of minivans, flowing down mainstreet

4 of 27

Develop UAV Payload to Optimize Data Collection and Exploitation

  • Automate the monotony of long-range monitoring
  • Real-time information
    • Take the intelligence to the field, rather than waiting to bring all the data back home
  • High-accuracy georegistration of events and features of interest
  • Inexpensive data collection
  • Open source platform
    • Eliminate lockin risk, enable open science

4

Process to acquire products from 2017 sUAS flights

5 of 27

Optimized Data Collection and Exploitation Life Cycle

5

5

VIAME

Imaging

Sensors FOV

INS

Embedded

Computer

Data Storage

Post-Flight Data Exfiltration

Data-Collection Payload

Upload New Deep Network Models

Train Deep Network Models

Real-Time Detections

6 of 27

Real-Time, In-Flight Segmentation Model

  • BiseNetV2
    • Real-time segmentation
    • 5320 x 3032 @2 fps….
  • Bilateral Architecture
    • Semantic Branch for high level semantic info
    • Detail Branch to capture low level info

6

7 of 27

Compression of Data Product

  • BiseNetV2
  • Real-time segmentation
  • Vectorize segmentation
    • 1.9MB png -> 20kb png
    • 100x compression of information!
  • 5,320 x 3,032 @2 fps

7

NVIDIA

Xaiver

8 of 27

Open Source Software and Documentation

  • All consumer-off-the-shelf (COTS) parts that are readily available
    • NVIDIA Xavier computer
    • Inertial Navigation System (INS)
    • Camera and lens
    • WiFi components
  • Total components Cost: $6574 - $8060
  • Open source software

8

Available on our Websitehttps://kitware.github.io/adapt/parts/

We are looking to build an open community of users

9 of 27

Completed System Build

9

Camera �adjustable from nadir to forward-looking

Inertial Navigation System (INS)

NVIDIA Xavier Computer

WiFi Downlink

GPS Antenna

Inertial Navigation System (INS)

10 of 27

System in the Field

10

Sept. 8th 2021, Test Flights Fairbanks, AK

at the Tanana and Chena Rivers

11 of 27

The System in Action

  • Live feedback on a WiFi-connected ground-station laptop
  • Quality control of recorded data via specialized low-bandwidth remote views
  • Real-time segmentation results (using an earlier prototype ice-segmentation model out of season)
  • Segmentation color codes:

11

System about 2,500 ft away along Tanana River

Ground

Sky

Trees

to test live segmentation chain

12 of 27

The System in Action

  • Live feedback on a WiFi-connected ground-station laptop
  • Quality control of recorded data via specialized low-bandwidth remote views
  • Real-time segmentation results (using an earlier prototype ice-segmentation model out of season)
  • Segmentation color codes:

12

Frozen Water

Background

13 of 27

The System in Action

  • Live feedback on a WiFi-connected ground-station laptop
  • Quality control of recorded data via specialized low-bandwidth remote views
  • Real-time segmentation results (using an earlier prototype ice-segmentation model out of season)
  • Segmentation color codes:

13

System about 2,500 ft away along Tanana River

Ground

Sky

Trees

to test live segmentation chain

14 of 27

The System in Action

System compatible with beyond line of site operation

14

Nearly Beyond Line of Sight

15 of 27

The System in Action

Real-time, remote image inspection for quality control

15

System about 2,700 ft away

16 of 27

Highly Synchronized Imagery and Navigation Data

  • PTP time synchronization between INS, Xavier, and camera
    • Sub-millisecond synchronization
  • Allows for powerful exploitation of visio-inertial correspondence for accelerated structure from motion and SLAM
    • Facilitate geo-registration and geometric assessment of environment

16

Comparison between camera orientation change (10 Hz) as measured from imagery via structure-from-motion (red) and INS (blue)

17 of 27

Analysis Over Time with Pixel-Level Registration

17

2017-05-05

2017-05-08

Significant ice breakup between observations

18 of 27

ADAPT Multi-Mission Payload

18

18

Image from https://www.weather.gov/otx/Our_Office

Oil Spill

An open source platform for deploying state of the art deep-neural-network computer vision in real time on small unmanned aircraft systems (sUAS).

ADAPT Multi-Mission Payload

19 of 27

Differentiating ADAPT: Annotation

  • Goal to optimize for annotation time
  • Experimented with CVAT, GrabCut, and GIMP
    • Polygon annotation - Intelligent Scissors and CVAT AI-enhanced tools
    • Dense annotation - semi-automatic (GIMP)
  • Identified Frozen Water vs Non-Frozen Water is most important boundary
    • Very complex and jagged boundaries
    • Variable texture although predominantly a very white color.
  • Developed a semi-supervised training and annotation workflow to annotate this specific boundary

19

Complex Topology

20 of 27

Novel Annotation Workflow

Sparse Human Annotations

New Training Image

Model-based annotation interpolation

Human Annotator�Corrections

20

21 of 27

Novel Annotation Workflow

Extant segmentation model bootstraps annotations requiring only minor corrections to rapidly achieve high-quality ground truth

21

22 of 27

Annotation Workflow

  • Goal to generate as many ground truth ready images
  • Reduce annotation time from 8.5 min/image to 1.5 min/image

22

Raw Image Data

GT-Generation Model Training

Ground Truth Segmentation Annotations

Sparse Annotation

Correction

Sparse Annotation

Model Output Human Adjudication

GT-Generation Model Deployment

23 of 27

Thank You!

23

Please visit our website for more informationhttps://kitware.github.io/adapt/

This presentation was prepared by Kitware Inc. using Federal funds under award NA20OAR0210083 from the National Oceanic and Atmospheric Administration, U.S. Department of Commerce. The statements, findings, conclusions, and recommendations are those of the author(s) and do not necessarily reflect the views of the National Oceanic and Atmospheric Administration or the U.S. Department of Commerce.

24 of 27

Ice Segmentation Results

24

WATER

ICE

SNOW

GROUND

Legend

Potential high-risk ice breakaway!

Results using BiSeNetV2 running real-time on a Xavier AGX

25 of 27

Retrain Model with Corrected GT

Original Hard Negative vs Retrained

25

26 of 27

AI-Assisted Annotation Workflow

Human inspection of model output on unseen examples allows

26

27 of 27

ADAPT Multi-Mission Payload: Oil Spills

27

27

Existing Approaches

ADAPT Approach

Pipeline Inspection / Spill Detection

Active Spill Tracking

Miles-Long Undersea Pipeline

Pilot

Spotter

Remotely Piloted UAV

Relay

Raw Video

Limited Range

Inefficient Voice Comms to HQ

Oil Spill

Manually �Surveyed Coordinates

Reduced Costs, Fewer Personnel, More-Timely Knowledge

Autonomous Operation

Low-Bandwidth

Long-Range

Visible– Thermal Imagery

Rapid Survey of Georegistered Boundary of Spill Sent Directly to HQ

ADAPT

Payload

Thickness Est.

Mapping

Classification

Detection

Inefficient and

Costly

Low-Bandwidth

Long-Range

Georegistered ALERTS

Inefficient and

Costly

Manned Boat

HQ

Inefficient Voice Comms to HQ

ADAPT

Payload

Thickness Est.

Mapping

Classification

Detection

Autonomous Operation

Active Spill

Potential Spill

Miles-Long Undersea Pipeline

HQ

ADAPT payload for optimized oil spill response operations�• Working with Alaska Clean Seas to define improved spill remediation approaches