Dynamic Vision-Based Satellite Detection: A Time-Based Encoding Approach with Spiking Neural Networks

Nikolaus Salvatore, Justin Fletcher

Abstract: The detection of residence space objects (RSOs) is a challenging task due to the lack of distinguishing features and relative dimness of targets compared to the high brightness and high noise of data collection backgrounds. Dynamic vision sensors present a possible solution for ground-based optical detection of RSOs due to their high temporal precision and dynamic range. However, the visual data collected by these sensors is asynchronous, and in the context of high-contrast remote sensing applications, often contains extraneous noise events. We propose a temporally-weighted spike encoding for dynamic vision data which reduces the density of event-based data streams and enables the training of deeper spiking neural networks for classification and object detection applications. The encoding scheme and subsequent spiking neural network architectures are evaluated on available dynamic vision datasets, with competitive classification accuracies on N-MNIST (99.7%), DVS-CIFAR10 (74.0%), and N-Caltech101 (72.8%), as well as showing state-of-the-art object detection performance on event-based RSO data collections.

External IDs:dblp:conf/icvs/SalvatoreF23