Sensor Fusion Lab

Teams building drone detection and perimeter surveillance hear the same promise everywhere: add a second sensor, fuse the data, and tracking gets better. Whether it actually does depends on the sensors' noise, update rates, latency, missed detections and false alarms — and on how the tracks are associated and fused.

Testing this in the field is slow and expensive, and field data has no exact answer key: nobody knows precisely where every object really was. A single test run proves little, because one lucky or unlucky realisation of noise can flip the conclusion.

Sensor Fusion Lab walks through the whole chain in a browser: build a scenario with moving objects, attach simulated sensors with their own noise, rates, latency and false alarms, run a Kalman tracker, compare data-association strategies, fuse several sensors into one picture, and finally score everything against the known truth. Every stage shows its parameters, plots and an explanation of how to read them.

The key engineering idea is a strict wall between truth and algorithms: the simulator knows where every object is, the tracking and fusion code only ever sees sensor observations, and automated tests enforce that boundary. Comparisons are kept fair — the same data for every variant, the same rules for every configuration, scoring on a common time grid — and every experiment is stored with its configuration hash so it can be re-run and verified byte for byte.

The evaluation stage repeats a comparison over many random seeds and parameter sweeps and scores it with established multi-target metrics (GOSPA, CLEAR MOT, IDF1). A claim such as "fusion is better" appears only when the 95 % confidence interval of the paired difference supports it.