Quadrocular, Neuromorphic Stereo Triangulation and Asynchronous Data Fusion for 3D Object Tracking

Abstract

Neuromorphic vision sensors (NVS) incorporate pixel-level circuitry that is only sensitive to local gradients in intensity. As a result, NVS' reduce background clutter and increase signal-noise ratio for long-standoff scenarios while providing the high temporal resolution needed for tracking of high-speed objects. This paper presents a hardware apparatus and novel algorithm to explore NVS passive ranging capability. A quadrocular, neuromorphic stereo system was constructed, calibrated, and used to capture temporal contrast signatures from high-velocity targets. Resulting data, computed calibration parameters, and a spatiotemporal correspondence metric were subsequently used to compute a pair-wise triangulation of object tip location. An asynchronous, adaptive Kalman Filter was then designed and deployed to fuse 3D track information from each camera pair and produce a high accuracy estimate of tip location and velocity. Preliminary results show this approach reduces track error over other standard methods by 30-50% while yield target tracks greater than 1MHz equivalent sampling rate under short standoffs.