Building a Machine Learning Tracker

How are trackers usually designed?

Classical tracker designs use Kalman Filters or similar approaches to combine noisy sensor measurements (target detections) with an in-built movement model based on the rules of physics to produce a smoothed, lag-free estimation of the current target position. These do a good job generally, but the limitations do become apparent in GMTI (Ground Moving Target Indication) tracking applications. Ground moving targets are often capable of quite sudden manoeuvres such as sharp turns which the internal models of the Kalman Filters do a poorer job of predicting. Compounding the problem, radar detections are often missing just at the point when a turn is executed due to the target slowing down (small, slow moving targets can more easily disappear into the background ground clutter, depending upon the radar in use and its operational mode).

Applying Machine Learning (ML) techniques

There is an opportunity to do a better job, by taking advantage of other data sources to enrich the model predictions. For example, modern mapping data can suggest which paths are more likely. A car approaching a T junction will normally slow down and turn left or right, rather than coasting straight ahead at constant speed into the field beyond the junction. There is also the possibility of improving the motion predictions of the model by learning from data rather than simply relying on pre-conceived motion models.

The difficulty of applying ML to tracking problems is that of securing sufficient real-world labelled data for supervised learning and validation. This can be mitigated by augmenting real world data with synthetic, simulated tracker input.

Simulator development and deep learning training

We began with a simple simulator that could generate simple target sequences and add noise and detection probability effects. It was initially based on reading in fixed road networks to use for target paths. This has now been extended to read mapping data from anywhere in the world, and to generate collections of realistic target movements for ML training and validation. Using this data, it has been possible to train a deep learning network to do a good job of predicting target positions from the recent movement history. These trained ML models will be integrated into the real-time tracker application.

Having built a simulator for ground targets, we now plan to extend it to provide airborne and satellite targets, as we build towards a general purpose, high capacity, “track everything” capability.

If you would like to find out more about either the simulator or tracker itself, please get in touch.