AI-Based Target Detection For Modern Radar Systems

Abstract

Radar Systems continuously emits electromagnetic pulses to illuminate object of interests in the environment and receive back-scattered signal. These back-scattered signals called echoes are collected by the Radar receiver and are integrated in a two-dimensional grid to formulate a raw 2D-map. This data is further processed using FFT based MTD techniques and coherent integration is performed. Afterwords, other signal processing and detection algorithms are applied to extract moving targets in the presence of clutter and interference. Convectional target detection generally used in radar include Constant False Alarm Radar (CFAR). CFAR detector is based on statistical models and may not provide optimal results in dynamic environment scenarios which includes strong clutter,jamming ,interference, weather & atmospheric affects. Furthermore, CFAR detector is sub-optimal for targets with rotors due to micro-doppler effects. AI based technique have potential to be used in Radar for target detection/ classification. Due to unavailability of actual radar data, simulated datasets are widely used in literature for the training and testing of AI based models which results in biased trained networks. In this thesis, research is done for usage of AI algorithms for Radar target detection with special focus on training datasets. For the generation of simulated dataset, a software simulator is developed to emulate radar transmitter, antenna, receiver, targets, and clutter scenario using MATLAB. Two deep learning based models namely Faster-RCNN and YOLO-v4 are trained and tested on the simulated radar dataset. Actual radar data files are utilized to produce a large dataset and both deep learning models are evaluated on it. Subsequently, training on actual radar data is performed and results are compared with conventional target detection technique as well as with the networks trained on simulated dataset.