MAMOC –
Machine Learning Application for Motion Capture
Speaker: Michael Spijkerman
Your partner for software
development and consulting
In our development center and at
your site
• Software conception and development - tailored
to your needs
• Provision of individual specialists and entire
teams for your projects
• Experienced certified project management
• Your extended workbench - taking over task
packages
Tailored to your needs
• Quality, Security, Proof-of-Concepts,
Full agile customer proximity
The challenge – what happens on manual
production stations?
Input Output
Within a process step several actions are performed in a
meaningful order
A machine:
- Highly Integrated to digital production processes
- By digital Status Feedback (started, stopped, error, …)
- Optimized for best throughput and quality
A human
- Does things a machine cannot do
Input Output
We just need a smart supervision sensor
- provides digital process relevant Data
- Provide data for optimization projects (i.e. industrial engineering)
Challenge accepted!
Object
Detection
+ +
Hand-Pose
Estimation
Action
Recognition
Record
Depth-
Video
X
Z
Y
X
Z
Y
Our Approach:
- Not only RGB Video Material but also:
- Identified Objects
- Identified Handposes
- Depth Image
- Location data of the Hands and Objects
The experiment
Setup performed
Converted loupe lamp with Intel Realsense D435i
The experiment
Example chosen
Assembling a speaker kit
The experiment
Trained Object Detection
YOLOv4 (Cross Stage Partial Network, CSPNet) executed on darknet
The experiment
Enabled Hand Pose Estimation
MediaPipe Hands
The experiment
Implemented Action Detection Solution
Adjusted 2 Stream Advanced Graph Convolutional Network (2s-AGCN)
The experiment
Implement the MAMOC Framework
MIDIH Stack
Using MediaPipe Framework with some additional calculators
The experiment
Trained the Action Detection
VGG Annotator
The experiment
The feasibility of our approach could be proven!
Test and Validation
Conclusion
simple to mount
easy integration through midih compatibility
MIDIH Stack
cheap embedded Hardware
Input Output
A human:
- can be highly Integrated to
digital production processes
- can be optimized for best
throughput and quality
What we reached
- planned accuracy of object detection and hand
pose Estimation
- MIDIH Integration
- Self assessment: TRL 3-4
next Steps
- work on action detection accuracy
- (Applied) funding at federal ministry of
education and research “KI4KMU” program to
reach TRL 7
Presentations
- digital Hannover fair 2020
- 18th (Digital) Trade Congress: "Efficient and
innovative mechanical engineering
- Regional workgroup KI in medium-sized
businesses
- (applied) Embedded Software Engineering
Congress digital
THANK
YOU!
dmc-smartsystems GmbH Developement Center
Paderborn
Valentin-Linhof-Str. 8
81829 München
Tel: +49 89 42774-0
Fax: +49 89 42774-199
Ahornallee 9
33106 Paderborn
Tel: +49 5251 6824-441
info@dmc-smartsystems.dmc-group.com
Michael.spijkerman@dmc-group.com
www.dmc-smartsystems.com
MAMOC presentation:
https://youtu.be/Lzp5al4b3IQ
MIDIH marketing:
https://youtu.be/5GAsUu-uNhI

Mamoc dmc smart systems-midih_oc2_demoday

  • 1.
    MAMOC – Machine LearningApplication for Motion Capture Speaker: Michael Spijkerman
  • 2.
    Your partner forsoftware development and consulting In our development center and at your site • Software conception and development - tailored to your needs • Provision of individual specialists and entire teams for your projects • Experienced certified project management • Your extended workbench - taking over task packages Tailored to your needs • Quality, Security, Proof-of-Concepts, Full agile customer proximity
  • 3.
    The challenge –what happens on manual production stations? Input Output Within a process step several actions are performed in a meaningful order A machine: - Highly Integrated to digital production processes - By digital Status Feedback (started, stopped, error, …) - Optimized for best throughput and quality A human - Does things a machine cannot do
  • 4.
    Input Output We justneed a smart supervision sensor - provides digital process relevant Data - Provide data for optimization projects (i.e. industrial engineering) Challenge accepted! Object Detection + + Hand-Pose Estimation Action Recognition Record Depth- Video X Z Y X Z Y Our Approach: - Not only RGB Video Material but also: - Identified Objects - Identified Handposes - Depth Image - Location data of the Hands and Objects
  • 5.
    The experiment Setup performed Convertedloupe lamp with Intel Realsense D435i
  • 6.
  • 7.
    The experiment Trained ObjectDetection YOLOv4 (Cross Stage Partial Network, CSPNet) executed on darknet
  • 8.
    The experiment Enabled HandPose Estimation MediaPipe Hands
  • 9.
    The experiment Implemented ActionDetection Solution Adjusted 2 Stream Advanced Graph Convolutional Network (2s-AGCN)
  • 10.
    The experiment Implement theMAMOC Framework MIDIH Stack Using MediaPipe Framework with some additional calculators
  • 11.
    The experiment Trained theAction Detection VGG Annotator
  • 12.
    The experiment The feasibilityof our approach could be proven! Test and Validation
  • 13.
    Conclusion simple to mount easyintegration through midih compatibility MIDIH Stack cheap embedded Hardware Input Output A human: - can be highly Integrated to digital production processes - can be optimized for best throughput and quality What we reached - planned accuracy of object detection and hand pose Estimation - MIDIH Integration - Self assessment: TRL 3-4 next Steps - work on action detection accuracy - (Applied) funding at federal ministry of education and research “KI4KMU” program to reach TRL 7 Presentations - digital Hannover fair 2020 - 18th (Digital) Trade Congress: "Efficient and innovative mechanical engineering - Regional workgroup KI in medium-sized businesses - (applied) Embedded Software Engineering Congress digital
  • 14.
    THANK YOU! dmc-smartsystems GmbH DevelopementCenter Paderborn Valentin-Linhof-Str. 8 81829 München Tel: +49 89 42774-0 Fax: +49 89 42774-199 Ahornallee 9 33106 Paderborn Tel: +49 5251 6824-441 info@dmc-smartsystems.dmc-group.com Michael.spijkerman@dmc-group.com www.dmc-smartsystems.com MAMOC presentation: https://youtu.be/Lzp5al4b3IQ MIDIH marketing: https://youtu.be/5GAsUu-uNhI