3. 3
Industrial IoT Automotive AI
Motor Condition
Classification
Challenges
Battery Powered,
Power Consumption
Achievements
Captured data sets, first
DL model
Arc Detection in DC
Power Distributions
Challenges
Accuracy,
Response Time
WiP
Creating good
learning data
Automatic Pedestrian
Emergency Breaking
Challenges
Response Time,
Robustness,
Distribution of work
Achievements
Captured data sets on airfield,
first DL model
Industrial IoT
4. 4
A “typical” smart home
4
Smart Kitchen
• Gesture, speech recognition
Intelligent Door
• Face, obstacle recognition
Dialog assistance
• Face, speech recognition
Smart Mirror
• Gesture, face, object
and speech recognition
Intelligent Chair and Fitness coach
• Gesture, face and speech recognition
5. 5
Display coated with a semi-transparent foil
Own mirror image + digital overlay
Intuitive control via gesture and voice
Show personalized information
Gaelic football scores, weather, news, …
Data Privacy as highest priority
Simultaneously computation of many
neuronal networks in the edge
Smart Mirror as Central Demonstrator
6. 6
Optimization of Neural Networks
Use EmbeDL toolchain
Distribute on embedded accelerators
Generate Virtual Mirror Image
3D representation
Frontal view without “real” mirror
Develop Secured Local Voice Assistant
Speech to Text, Text to Speech, NLP
Use VEDLIoT project hardware
t.RECS and u.RECS
Increase energy efficiency
Development & Optimisation Goals
Display Messages from/
to other devices
Face Recognition
Object Detection
YOLO
Gesture Detection
Smart Home
DNN
Efficientnet
Trained with WiderFace
Peripheral Input
Combinatorial Logic
Decision Maker
RGB-D Camera
ROS2 Bridge
Pointcloud
Registration/
Alignment
Person
Recognition
MagicMirror²
ROS2
/
t.RECS
Speaker
ROS2
/
u.RECS
Siamese Network
for Identification
Emotion
Recognition
Mask Detection
Tracker
(Sort)
RGB-D Camera
Intent Extraction
Skill System
Text to Speech
Tracker
(Sort)
Pointcloud Fusion for
Virtual Mirror image
Webstream
Microphone Input
Speech to Text
7. 7
Increase of energy efficiency
First Prototype
• Intel i7-7700K
• 2x GTX 1080 TI
• 16 FPS / 600 Watt
-> 37.5 Watt / Frame
Second Prototype
• Intel i9-9900k
• 2x RTX 2070
• 27 FPS / 450 Watt
Third Prototype
• 2x Nvidia Xavier
• 16 FPS / 55 Watt
Actual Prototype
• 2x Nvidia Xavier
in christmann t.RECS
• 28 FPS / 55 Watt
-> 1.9 Watt / Frame
19x
increase
8. 8
Thank you for your
attention
Micha vor dem Berge
christmann IT, Germany
micha.vordemberge@christmann.info