Event- Based Target Detection and Tracking for Remote-Sensing Applications

Neuromorphic event-based vision sensors have emerged as a promising technology for space-based target detection and tracking due to characteristics like high dynamic range and high temporal resolution. Most existing research focuses on the tracking of relatively large objects in terrestrial environments or the task of space situational awareness (SSA), where targets are small but background clutter is limited. This research explores the application of various detection and tracking techniques to remote-sensing event-stream data. Remote-sensing applications typically require tracking small objects embedded in the clutter of the Earth background. Adding to the challenge is the limited availability of neuromorphic remote-sensing datasets with reliable truth information. To overcome this challenge, we introduce an event-stream simulation pipeline leveraging existing simulators. The simulation pipeline is used to generate datasets for sensors in a geostationary orbit (GEO) and a low-Earth orbit (LEO). Existing methods designed for SSA are shown to generate high numbers of false alarms caused by structured clutter from ground-based features. Therefore, we propose a simple machine-learning (ML) detector that operates on events as they are produced. It is observed that this method can more effectively discriminate between events generated by targets moving in the scene and events generated by ground features or sensor noise. Aggregated frame-based approaches also show promising performance for applications where the full temporal resolution of the sensors does not need to be leveraged. Tracker performance for each method is evaluated with standard track metrics including precision, recall, multiple object tracking accuracy (MOTA), and target localization error. Performance is also characterized relative to target intensity and velocity. The event-by-event and frame-based ML methods demonstrate more generalizable performance in the presence of background clutter. ML methods achieve MOTA scores of up to 0.86 on simulated data generated compared to a maximum of 0.68 for non-ML methods.