20
Copyright © 2017 1 Using AI to Read Human Body Language In Real- Time from Standard Video

Using AI to Read Human Body Language In Real- Time from ...on-demand.gputechconf.com/gtcdc/2017/presentation/dc7174-paul... · Language In Real-Time from Standard Video. ... 80% of

Embed Size (px)

Citation preview

Copyright © 2017 1

Using AI to Read Human Body

Language In Real-Time from Standard

Video

Copyright © 2017 2

Current SolutionsKeyboardsVoiceSpecialty Sensors / Cameras

ShortcomingsExpensive to own / operateTime consuming to set up / useRequire specialty sensors / suits

How do we enable computers to communicate with us in a natural way

Friction: Humans are Physical / Computers are Digital

Copyright © 2017 3

Humans are Visual Creatures80% of language is non-verbal80% of that is body language

Dogs have co-evolved with humans to read human body language

Inspired by Nature

Security Work Companionship Entertainment Navigation

Copyright © 2017 4

Solution: Teach Computers to Read Human Body Language

How do we make the digital version of “Man’s Best Friend”?Solution:

Ordinary cameras + GPUs + deep learning Value Proposition

Natural InteractivityConsumer Hardware keeps getting cheaper and betterDeep Learning Brains keep getting smarter

Copyright © 2017 5

Why Now? Peace Dividends of the Game Console Wars

Kinect paved the wayGPUs enable deep learningGame Engines enable synthetic data and thousands of new interactive applications

Copyright © 2017 6

wrnch™: Human Pose Estimation Engine

User Application

wrnch™

In: Video

Out: Humans

Copyright © 2017 7

Deep LearningAccurateRobustKeeps getting smarter

Real-timeEnables interactivity

No Specialty HardwareConsumer grade cameras & GPUsTake as input: any video feed from anywhere

Wrnch™ Engine: Unique Features

Copyright © 2017 8

Live Demos

Copyright © 2017 9

Applications: AR / VR

Copyright © 2017 10

Applications: Robot Interaction

Copyright © 2017 11

In City for safetyIn Home for elder care, etc.In Store for retailIn Vehicle for awarenessIn Factories for accidents, etc.

Applications: Security / Human Monitoring

Copyright © 2017 12

Applications: Health & Wellness

Copyright © 2017 13

Applications: Eyes for Virtual Assistants

Copyright © 2017 14

Applications: Sport Analytics

Copyright © 2017 15

Fast allowing real-time interactivityAccurate tracking of 63 body parts per person including fingersRobust across large numbers of people in crowded conditions

BodySLAM™: Unique Features

Copyright © 2017 16

Wrnch Deep Learning Training Pipeline

Real Data Synthetic Data DL Models

wrBrain

Wrnch DL Training Framework

Copyright © 2017 17

Wrnch Deep Learning Inferencing Architecture

2D Pose

3D Pose

Tracker Activity Recognition

Gesture Recognition

Segmentation 2D Skeletons

3D Skeletons IDs Locations Activities Gestures

2D Video

2D Video

2D Video

3D Shape

3D Meshes

Identifier

TensorRT

Copyright © 2017 18

GPU CPU OS Total(mSec)

2D CNN 3D CNN Misc

Jetsen TX2 n/a Linux 297 200 84 13K80 Intel Xeon E5-

2686v4 2.3GHz Linux 131 90 36 51080 Intel i7-5930K

3.5GHz Linux 37 22 12 31080 Intel i7-5930K

3.5GHz Windows 41 24 13 4

Titan XP Intel i7-5930K 3.5GHz Linux 30 16 11 3

Titan XP Intel i7-5930K 3.5GHz Windows 34 18 12 4

Runtime Performance

Copyright © 2017 19

Gesture RecognitionPoint AtThumbs upFist

Activity RecognitionFall detectionPick up itemPut down item

Multiple Camera SupportTriangulationPersistent tracking3D

Roadmap

Copyright © 2017 20

Questions?