PostDoc Internship - RAISE

Key Skills

Algorithms, Deep Learning

Job Description

Radars are a key modality in the race to realize a L4-L5 self-driving car. Traditionally, the low angular resolution has been a limiting factor for radars. Recent imaging radars are however pushing the boundaries of radar capabilities further than what was possible before. In this context, there is great interest in techniques that allow for a higher (super-resolution) angular resolution through advanced signal processing. The SOTA technology is sparse array processing, which uses various compressive sensing techniques and array design to achieve a high angular resolution under feasible data rates through sub-Nyquist sampling.

The performance of current compressed sensing methods is however strongly limited by the following.

1) The iterative optimization algorithms (such as sparsifying Bayesian techniques) that are used to estimate the sparse code are very time-consuming, leading to high latencies.

2) These methods make strong assumptions about the data distribution (priors). Performance is thus bounded by the accuracy of these heuristic priors and their parameters.

3) In practice, the actual underlying distributions cannot be described analytically.

4) These priors are not task-adaptive, i.e. they do not consider the downstream system task (e.g. radar object detection).

To address these challenges, we propose physics deep learning methods for fast signal recovery in radar compressed sensing. Our methods learn task-adaptive statistical priors directly from data, while also incorporating all available knowledge of signal structure and the physics of the measurement environment. This circumvents formalizing a highly-complex (and likely intractable) statistical model and the challenges in its estimation, by learning tractable models directly from data. We will use deep algorithm unfolding methods to exploit known signal structure or acquisition physics in network designs. Along this line, we have already shown that advanced signal processing algorithms for sparse recovery in clutter (robust PCA), and super-resolution (sparse coding) can be recast as trainable deep networks.