Ultra-low latency, low-power radar-based multi-object scene identification using spiking neural networks

The use of mm-wave radar for gesture, gait or pose recognition is a challenging and active field of research. However, future algorithms need to process multiple objects in a scene instead of identifying a single object, which requires segmentation and identification. Moreover, some application, such as drone or robotic autonomy, requires ultra-low latency, low-power implementations. The Ph.D. aims to investigate novel approaches, exploring faster, lower computation complexity and efficient radar signal encoding challenging the traditional radar signal processing pipeline. Moreover, we will use spiking neural networks (SNNs) to replicate closely how the brain process information. SNNs can be implemented on chips requiring ultra-low power and intrinsically processing temporal information, which is ideal for processing radar data.

In summary, the ideal candidate will focus on the following research questions:

• Explore novel neural encodings at the radar chirp level to reduce the traditional radar processing pipeline and extract ideal features for the SNNs
• Investigate SNNs to process radar information, enabling scene recognition and possibly understanding. For that, we will require multi-object identification and recognition, implying the segmentation of objects from the radar maps.
• Depending on the running projects, we will devise a demonstrator to showcase the different components of this research encompassing in an end-to-end solution.