Radar Data Compression through Model-based Deep Learning
Introduction
Automotive radar sensors are required to adhere to the functional and safety standards, meaning that these sensors are required to operate with high update rates (>25 Hz) and high resolution in range, Doppler, and angular domains. The significant dynamic range of radar signals promotes the use of high-performance analog-to-digital converters (ADCs). All these factors lead to an enormous amount of data collection for a single radar frame, i.e., generally tens of gigabits per second (Gb/s). This increases both the on-chip memory and data transfer expenses. The student will look into enhanced deep learning methods that can encode and decode radar data, having real-time and memory constraints.
Scope
Data compression is being applied in a broad range of applications. Each application may use different approaches to compress data by exploiting application-specific features in the data. In contrast, JPEG for RGB images use, for example, down-sampling, block splitting, and a discrete cosine transform to enable a lossy compression. Deep learning is commonly applied nowadays and has shown outstanding performance in some tasks; hence, also in the context of data compression in the following domains: communications [1], medical imaging [2], seismic sensing [3], and SAR imaging [4]. Likewise, in automotive radar sensing, similar and more sophisticated/pruned methods can be applied. The student is free to define the architecture of the encoding and decoding model and exploit radar-specific features to further reduce the data rate.
#LI-0d06