Event-Based Noise Filtration with Point-of-Interest Detection and Tracking for Space Situational Awareness

Abstract

This thesis explores an asynchronous noise-suppression technique to be used in conjunction with asynchronous, Gaussian-blob tracking on dynamic vision sensor (DVS) data. This type of sensor is a member of a relatively new class of neuromorphic sensing devices that emulate the change-based detection properties of the human eye. By leveraging a biologically inspired mode of operation, these sensors can achieve significantly higher sampling rates as compared to conventional cameras, while also eliminating redundant data generated by static backgrounds. The resulting high dynamic range and fast acquisition time of DVS recordings enables the imaging of high-velocity targets despite ordinarily problematic lighting conditions. The technique presented here relies on treating each pixel of the sensor as a spiking cell keeping track of its own activity over time, which in turn can be filtered out of the resulting sensor event stream by user-configurable threshold values that form a temporal bandpass filter. In addition, asynchronous blob-tracking is supplemented with double-exponential smoothing prediction and Bezier curve-fitting in order to smooth tracker movement and interpolate target trajectory respectively. This overall scheme is intended to achieve asynchronous point-source tracking using a DVS for space-based applications, particularly in tracking distant, dim satellites. In the space environment, radiation effects are expected to introduce transient, and possibly persistent, noise into the asynchronous event-stream of the DVS. Given the large distances between objects in space, targets of interest may be no larger than a single pixel and can therefore appear similar to v such noise-induced events. In this thesis, the asynchronous approach is experimentally compared to a more traditional approach applied to reconstructed frame data for both performance and accuracy metrics. The results of this research show that the asynchronous approach can produce comparable or even better tracking accuracy, while also drastically reducing the execution time of the process by seven times on average.