tinyML Asia 2021 Video Poster: Efficient inference of low-resolution optic flow on low power...

tinyML Asia 2021 Efficient inference of low-resolution optic flow on low power neuromorphic hardware Felix BAUER, R&D Engineer, SynSense Motion can be inferred from visual scenes by determining the optic flow. For tasks like ego-motion regression the spatial resolution of the optic flow field can be very low. This work presents a Spiking Neural Network (SNN) that infers a low-resolution optic flow field from Dynamic Vision Sensor (DVS) input. Based on a non-leaky Integrate-and-Fire (IAF) neuron model, it can be deployed on neuromorphic hardware such as Speck. Combining an event-based sensor with a low power event-driven asynchronous processor, Speck is suitable for efficient use in autonomous agents, such as drones, where low energy consumption is crucial. The model is trained and tested on a standard PC with data from drone-mounted DVS sensors and then deployed on Speck. We demonstrate that the network is capable of performing the highly temporal task of low-dimensional optic flow inference, mak