SHALINI GUPTA, PAVLO MOLCHANOV, KIHWAN KIM, KARI PULLI, JAN KAUTZ NVIDIA RESEARCH
GESTURE RECOGNITION: USING A MULTI SENSOR APPROACH SHALINI GUPTA, - - PowerPoint PPT Presentation
GESTURE RECOGNITION: USING A MULTI SENSOR APPROACH SHALINI GUPTA, - - PowerPoint PPT Presentation
GESTURE RECOGNITION: USING A MULTI SENSOR APPROACH SHALINI GUPTA, PAVLO MOLCHANOV, KIHWAN KIM, KARI PULLI, JAN KAUTZ NVIDIA RESEARCH DRIVER DISTRACTION GESTURE INTERFACE (http://www.softkinetic.com) DAY DAY NIGHT NO SUNLIGHT NO SUNLIGHT
DRIVER DISTRACTION
GESTURE INTERFACE
(http://www.softkinetic.com)
DAY
DAY NIGHT
NO SUNLIGHT
NO SUNLIGHT SUNLIGHT
sensors
COLOR + DEPTH RADAR
MULTI-SENSOR SOLUTION
gesture UI sensors
COLOR + DEPTH RADAR
MULTI-SENSOR SOLUTION
3D shape color velocity
3.2% INCREASED ACCURACY
+1.5m/s
- 1.5m/s
0 m/s
v t 0.15W velocity power
16X POWER EFFICIENCY
v t 0.15W vT gesture gesture velocity power 2.5 W
16X POWER EFFICIENCY
radar prototype
SHORT RANGE FMCW RADAR
x y z v radar prototype 4D vector
SHORT RANGE FMCW RADAR
+1.5m/s
- 1.5m/s
0 m/s
POSITION RESULTS
VELOCITY RESULTS
+1.5m/s
- 1.5m/s
0 m/s
GESTURE NETWORK
3D convolutional layer fully connected NN logistic regression subsampling layer 60 frames Trained on GPU
10 GESTURES
PALM SWIPE SHAKE CALL left right up down left right clockwise counter-clockwise ROTATION
INDOOR CAR SIMULATOR
OUTDOOR CAR
ERROR RATE
C D R D+C R+D R+C R+D+C 39.90% 9.10% 10.90%
D – depth C – color R - radar
ERROR RATE
C D R D+C R+D R+C R+D+C 39.90% 9.10% 10.90% 7.90% 8.30% 7.40%
D – depth C – color R - radar
ERROR RATE
C D R D+C R+D R+C R+D+C 39.90% 9.10% 10.90% 7.90% 8.30% 7.40% 5.90%
D – depth C – color R - radar
ERROR RATE
D – depth C – color R - radar
Night Evening Day (shadow) Day (sunlight)
6.70% 3.00% 9.70% 20.90%
D+R (CNN) D+R+C (CNN) D+C (HOG)
ERROR RATE
D – depth C – color R - radar
Night Evening Day (shadow) Day (sunlight)
6.70% 3.00% 9.70% 20.90% 6.70% 1.50% 8.30% 7.50%
D+R (CNN) D+R+C (CNN) D+C (HOG)
ERROR RATE
D – depth C – color R - radar
Night Evening Day (shadow) Day (sunlight)
6.70% 1.50% 8.30% 7.50% 22.20% 2.45% 13.00% 20.90%
D+R (CNN) D+R+C (CNN) D+C (HOG*)
*Ohn-Bar and Trivedi, IEEE Trans. on
Intelligent Transportation Systems, 2014.
52ms on Quadro 6000
DEMO
CONCLUSION
GESTURE UI2 COLOR + DEPTH RADAR1
1Multi-sensor System for Driver’s Hand-Gesture Recognition, IEEE Automatic Face and Gesture Recognition, May 2015. 2Short-Range FMCW Monopulse Radar for Hand-Gesture Sensing, IEEE International Radar Conference, May 2015.