Adapting Object Detection Models for Multi-Target Detection Utilizing Radars

Abstract

This paper investigates the use of Deep Learning (DL) in multiple input multiple output (MIMO) radar target detection, focusing on azimuth and elevation estimation. Traditional methods face challenges like interference and reflections, especially in multi-target scenarios. Feature extraction conventionally relies on range correlation, Doppler filtering, and angle beamforming, followed by detection after constant false alarm rate (CFAR) processing. However, early data sparsification by bin selection often leads to information loss, particularly with large data cubes required for practical implementation. Deep Learning techniques offer an alternative, specifically in azimuth and elevation detection at earlier stages of radar data processing. We developed a convolutional neural network (CNN) model that achieved Mean Square Errors (MSE) of 0.149 for azimuth and 0.168 for elevation on single-target data from 5,000 samples. The model's performance in dual-target scenarios showed MSEs ranging from 0.838 to 1.845, tested on 8,000 samples from a dataset of 72,000. This paper details the model development process, its impact on radar target detection, and potential future research directions involving the substitution of multi-bin Deep Learning blocks with traditional methods.