Joyce Mo, University of California, Berkeley and Princeton Satellite Systems; Michael Paluszek, Princeton Satellite Systems; Angela Kwon, Princeton University
Keywords: Autonomous spacecraft, machine learning, target tracking, object identification, multiple hypothesis testing, autonomy, statistics
Abstract:
Currently, target tracking systems face significant inefficiencies when assigning measurements to multiple objects observed by multiple sensors, particularly in dynamic environments where the number of objects is unknown and constantly changing. In such systems, tracks are typically formed through two primary mechanisms: either by extending an existing track with a new measurement or by creating a new track from a single measurement without reference to prior data. These tracks are represented as sequences of measurements collected over multiple scans, with each track maintaining state and covariance estimates through the application of a Kalman filter. However, as the complexity of tracking scenarios increases—such as with multiple spacecraft operating in coordination—the need for more robust data fusion and association methods becomes increasingly apparent.
The goal of this study is to investigate improvements in multi-sensor data fusion for target tracking by applying advanced probabilistic techniques. Specifically, this paper explores the use of multiple hypothesis testing (MHT) as a solution for managing the uncertainty inherent in associating measurements with object tracks. MHT is a statistical framework that evaluates multiple competing hypotheses about which measurements correspond to which objects, maintaining a set of plausible track configurations as new data arrives. A valid hypothesis consists of a set of compatible tracks, where no two tracks describe the same object or share the same measurement, ensuring consistency across the track-tree structure. Each track is uniquely identified by an object ID, allowing for clear differentiation and hypothesis management throughout the tracking process.
To demonstrate the feasibility of this approach, this study will present simulations involving multiple space platforms, each equipped with infrared and visible spectrum cameras. These spacecraft will operate with six degrees of freedom, allowing for realistic modeling of aerial dynamics and sensor observations. In addition to the MHT framework, a convolutional neural network (CNN) will be trained to classify and detect target objects from the multi-sensor imagery, contributing semantic information that can further support the data association process within MHT.
Ultimately, this study aims to contribute to the broader field of space situational awareness and autonomous surveillance by demonstrating a method that combines traditional statistical techniques, such as multiple hypothesis testing, with modern machine learning approaches, such as convolutional neural networks and image classifiers. By doing so, it seeks to address ongoing challenges in multi-target tracking, particularly in environments with overlapping sensor coverage and ambiguous measurements. This work lays the foundation for continued work in scalable, real-time tracking solutions that leverage both probabilistic reasoning and learned perception models.
Date of Conference: September 16-19, 2025
Track: Astrodynamics