PROCESS
The process was to first prototype a CNN running in real-time performing image recognition on a desktop class device. From there optimize the architecture, primarily the models so that they could run on an embedded device.
RESULT
Cross platform library that runs in real-time at greater than 25fps that performs image recognition on known datasets at >90% accuracy.
TECHNICAL DETAIL
TensorRT, C++
Initially running on ARM Cortex A-series.
Currently working on M-series.
