Parallel Computation of Event-Based Visual Features Using Relational Graphs

Abstract

A graphical framework to approximate visual features from spike-based sensor data has been demonstrated. An event-based camera or dynamic vision sensor (DVS) provides the sensory input into the network which computes the intrascene optical flow, spatial gradient, and absolute intensity. The network uses the sparse, event-based input along with fundamental relations in parallel to converge upon quantities via incremental optimization. An event-based algorithm to compute optical flow was used to provide another stream of input into the network to aid convergence. A full network has been deployed in Python and parallelized to demonstrate its potential deployment on specialized hardware.