ARI develops sensor-agnostic model to measure track uncertainty
The defense and intelligence communities have a lot of interest in accurately tracking vehicles and dismounts. While there are many tracking algorithms for a variety of sensors, there has been little research into measuring the inherent uncertainty of tracked trajectories and how to combine these trajectory uncertainties with other sensor information.
Connecting this information and geospatial information is generally done by a human analyst. The analyst does not have access to useable uncertainty data, so there cannot be total confidence in their conclusions.
Researchers at the Illinois Applied Research Institute (ARI) are working to change this by developing a sensor-agnostic model. This model will be able to measure track uncertainty as a whole.
Uncertainty in tracking is driven by two factors. One is detection level uncertainty, which is determined by sensor sensitivity and environmental phenomenology. The second is motion uncertainty, which comes from object trajectory.
Track-level uncertainty is a combination of the two factors.
ARI made the model by convolving the detection position-uncertainty measurement with a polar-coordinate trajectory propagation model.
The detection uncertainty model can be measured directly from the data or inferred from known parameters such as a camera point-spread function.
The trajectory propagation model assumes a constant velocity object trajectory with Gaussian uncertainty in speed and heading, however both the motion model and uncertainty distribution function could be modified for other uses. Much like a Kalman filter, the model consists of a measurement phase, prediction phase and an update phase.
The baseline motion model uses a constant velocity assumption, but it is able to adapt well to more complex trajectories as well as accurately reflect the increased uncertainty about object position under those circumstances.
ARI’s model allows analysts to make quantitative assessments about the likelihood that a tracked object was in a given location. They will also be able to answer questions about the probability of detections belonging to a given track or two tracks stemming from the same object.
These features enable persistent tracking over longer period of times which is essential for pattern of life and activity characterization.
Uncertainty information about tracked objects allows analysts to make more informed, quantitative decisions regarding geospatial information.
ARI is now interested in testing this model on a wider array of data, as well as pushing the limits of tracking in low SNR environments and cluttered tracking scenarios such as urban areas. As systems move toward true autonomy, incorporating accurate uncertainty information will aid algorithm robustness and improve performance.