mWidar (microwave detection and ranging) replaces expensive radar hardware with simple antennas and heavy signal processing. Instead of sweeping one direction at a time, it images an entire scene at once, at thousands of frames per second. That makes it attractive for counter-UAS work and, eventually, for tracking hypersonic vehicles.
That speed and low cost come with harder perception problems. Working with high-fidelity mWidar simulations, I build the full detector-tracker pipeline:
- Detection. Computer vision and deep learning methods that find targets in microwave images, even when targets are occluded or interfere with one another.
- Data association. Keeping identities consistent across frames when detections are missing, cluttered, or ambiguous.
- Estimation. Predicting target motion with Kalman and particle filters, and quantifying how confident the tracker should be.
The broader goal is to move sensing from hardware-intensive to software-intensive systems, so that capable tracking becomes cheap enough to deploy widely.