An Event-Based Vision Sensor Simulation Framework for Space Domain Awareness Applications

摘要

Event-based vision sensors (EVS) provide a unique opportunity for Space Domain Awareness (SDA) applications. Inspired by the human eye, these sensors operate on the principle of change detection and each pixel functions independently and asynchronously from the other pixels. Pixels trigger events when a change in light intensity occurs. The sensor records events as a sparse time series output with microsecond level of precision. The space sensing community is interested in this technology due to wide dynamic range achieved by the operating in the log scale, the minimal data produced for a relatively static scene where changes in intensity are infrequent, and the temporal precision that opens opportunities to capture information on fast moving space objects where traditional frame imaging is not an option. As a relatively inexpensive sensor with technical capabilities well suited for tracking, they could augment existing ground-based systems or be a primary sensor on-board a spacecraft. EVS are uniquely suited for space-based operations. With low size, weight, power, and data requirements, they easily fit into tight engineering budgets for space systems. The data may even be well suited for on-board processing due to its sparsity. Despite all these advantages, the sensors are not ready to implement into SDA operations. Creating algorithms to handle the time series data and optimizing the sensor for low-light imaging are areas of active research to improve the utility of these sensors as tools for SDA. To support these efforts, I develop physics-based end-to-end model for event-based sensing of resident space objects (RSOs). This model adapts previous synthetic event generation methods to operate with photon flux input and precise measures of current. By implementing a model of pixel readout with microsecond-level precision and developing methods to model noise based on dark and induced current levels, I improve the accuracy of the frequency and polarity of the events produced. This accuracy is necessary to extrapolate sensor performance from the model and to generate synthetic events to feed algorithmic development. I also contribute to event-based algorithms through development of an online non-frame based tracking algorithm. In order to validate the sensor model and train the classifying portions of my tracking algorithms, I develop batch-based clustering methods that leverage the temporal dimension which improves the labeling of events between star and noise by 31.8%. Through exploration of different grouping and classifying methods for the tracking algorithm, I attain a maximum of 94.5% group agreement with the batch clustered data and a 97.6% true positive rate and 99.9% true negative rate when classifying satellites on a validation data set. Star classification performance is slightly lower at a 96.7% true positive rate and 96.5% true negative rate. The tracking algorithm's success on this one set of data suggests promising performance from these sensors in future SDA applications.