Jason K. Oberg
Ph.D. at University of California, San Diego
M.S. at University of California, San Diego
B.S. at University of California, Santa Barbara
Hardware Implementation of Microsoft Kinect's Body Part Recognition
The Microsoft Kinect has received little to no attention by the hardware community, specifically those interested in reconfigurable hardware via field-programmable gate-arrays (FGPAs). This is likely due to both the lack of interfaces and to potential hardware applications. This research project takes a step forward in both directions. It presents our findings with attempting to interface the Kinect device to a modern Xilinx MLxxx development board; a common platform for hardware development using FPGAs. In addition, and most importantly, we present a fully-functional hardware implementation of a body part recognition algorithm written entirely in a hardware descriptive language (HDL). The algorithm uses randomized decision trees for computing the probabilistic location of body parts on a human. This research presents a complete architecture and hardware simulation for the algorithm. It predicts significant speed-ups over even parallelized software versions of the algorithm and discusses future optimizations.A nice detailed talk I gave on the project at the end of my internship at Microsoft Research can be found here:
The Kinect's Body part Recognition Algorithm on an FPGA--MSR presentation
Relevant Publications
[FPL'12] Random Decision Tree Body Part Recognition Using FPGAsJason Oberg, Ken Eguro, Ray Bittner, and Alessandro Forin
The International conference on Field Programmable Logic and Applications (FPL 2012)
[US Patent] Decision Tree Computation In Hardware
Jason Oberg, Ken Eguro, Victor Tirva, Padma Parthasarathy, Susan Carrie, Alessandro Forin, and Jonathan Chow.
US Patent App. 13/344,473, 2012