A Cognitive Radar for Classification of Resident Space Objects (RSO) operating on Polarimetric Retina Vision Sensors and Deep Learning

Abstract

A novel cognitive radar, operating on Polarimetric Dynamic Vision Sensor (pDVS) and deep learning principles, aimed at discriminating moving targets, based on their motion patterns, is presented. The system consists of an asynchronous event-based neuromorphic imaging sensor coupled with polarization filters which enable better discrimination; a spinning light modulating wheel, operating at varying angular frequency, is placed in front of a static object. A pipeline has been designed and implemented in order to train a neural network for motion pattern classification using event data. This pipeline first extracts features using a pre-trained convolutional neural network and then feeds these features into a single-layer long short-term memory recurrent neural network. The outcome of this study indicates that deep learning combined with pDVS principles is well suited to classify accurately motion pattern-based targets using limited set of data; thus opening the way to many innovative bioinspired-based vision applications where feature extraction is complex or precognitive vision-based applications for the detection of salient features. The proposed cognitive radar would be able to operate at high speeds and low bandwidth, while maintaining low storage capabilities, low power consumption, and high-processing speed.