TinyRadarNN Links Two Neural Nets on a RISC-V CPU for Low-Power RADAR-Based Gesture Recognition
GAP8 RISC-V processor powers a RADAR-based gesture recognition system built in just 92kB of memory and using 21mW of power.
A team of computing researchers from ETH Zurich and the University of Bologna have released a paper on creating an ultra-low-power gesture recognition system based on short-range RADAR sensors and neural networks running on the free and open source Parallel Ultra-Low Power (PULP) Platform processor design.
"This work proposes a low-power high-accuracy embedded hand-gesture recognition algorithm targeting battery-operated wearable devices using low power short-range RADAR sensors," the team explains in the paper's abstract, describing the TinyRadarNN system it developed. "A 2D Convolutional Neural Network (CNN) using range frequency Doppler features is combined with a Temporal Convolutional Neural Network (TCN) for time sequence prediction. The final algorithm has a model size of only 46 thousand parameters, yielding a memory footprint of only 92kB."
The extremely low memory footprint isn't the only notable thing about the team's approach: Implemented on the RISC-V-based GAP8 processor, an implementation of the free and open source Parallel Ultra-Low Power (PULP) Platform many-core processor, the system drew just 21mW of power in live prediction mode β meaning it's extremely usable in battery-powered devices.
"Two datasets containing 11 challenging hand gestures performed by 26 different people have been recorded containing a total of 20,210 gesture instances. On the 11 hand gesture dataset, accuracies of 86.6% (26 users) and 92.4% (single user) have been achieved, which are comparable to the state-of-the-art, which achieves 87% (10 users) and 94% (single user), while using a TCN-based network that is 7500x smaller than the state-of-the-art," the team notes, highlighting a dramatic improvement in network size for their work.
"Furthermore, the gesture recognition classifier has been implemented on Parallel Ultra-Low Power Processor, demonstrating that real-time prediction is feasible with only 21 mW of power consumption for the full TCN sequence prediction network."
The work is available on arXiv.org now under open access terms.