The overall objective is to introduce a generic Bayesian framework for robust sequential target discrimination from noisy time-series data. Drone recognition from radar data will be a key application area, for instance for unmanned air-traffic management. This research will leverage recent advances in sequential learning from time-series data (eg with kernel-based methods) and can involve incorporating any known priors on the features evolution over-time or the present noise or clutter using suitable (possibly hierarchical) stochastic state-space models. The key challenge is mitigating fluctuating and unreliable target (eg drone) classification results from supervised or unsupervised learning methods since: a) classifier uses features (eg kinematics or Doppler related) extracted from the sensor (eg radar) noisy time-series data, b) data and features quality significantly fluctuates over time (eg due to clutter, multipath, occlusions, etc) for a given target, c) some key salient features (eg related to micro-Doppler effects) are intermittently available.

The research will be conducted in close collaboration with Aveillant-Thales, who will offer guidance and potentially real radar data for various targets of interest to validate the developed recognition algorithms. This also includes the applicant spending at least 3 months at Aveillant's offices in Cambridge.

Applicants should have (by the start date) at least a good 2.1 degree in information engineering, mathematics or computer science. A background in statistical signal processing and/or machine learning is highly desirable, with experience in one or more of the following: tracking and SMC methods, supervised learning for classification and deep learning, anomaly detection and pattern-of-life inference. Applicants are expected to have a solid programming/modelling skills, for example in MATLAB/Python/C++.

Full details