Multi-Spectral Method for Hand Segmentation in Machine Learning

Back to all technologies
Download as PDF
2020-RAMA-68873
Researchers at Purdue University have developed a new method for hand segmentation in augmented reality and virtual reality to address the challenge in robot-human interaction where machines do not often understand the way human hands grip infinitely complex objects. Current technologies do not adapt to every possible degree of freedom of hand movement, nor recognize or classify grasping and using of objects with accuracy. The Purdue University approach overcomes these limitations therefore allowing the camera to distinguish various characteristics of human hands. Researchers were able to obtain seventy percent accuracy in comparison between manual hand motion and computer vision in a test involving individuals clasping thirty-six objects and using thirty different tools, where over four hundred thousand frames were collected. This resulted in a four percent improvement in Intersection over Union (IoU) with thirty-percent fewer parameters compared to similar machine learning architectures.

Advantages:
-Rapid human robotic communication
-Automated manufacturing
-Improved performance
-Fewer parameters than other architectures

Potential Applications:
-Machine Learning
-Pedestrian Detection
-Biological Image Processing
-Autonomous Vehicles
Dec 2, 2020
Utility-Gov. Funding
United States
(None)
(None)

Dec 2, 2019
Provisional-Patent
United States
(None)
(None)
Purdue Office of Technology Commercialization
The Convergence Center
101 Foundry Drive, Suite 2500
West Lafayette, IN 47906

Phone: (765) 588-3475
Fax: (765) 463-3486
Email: otcip@prf.org