Spatial classification of range doppler maps generated from vehicles, humans and drones using 8.5 GHz FMCW radar

In this thesis, object classification techniques for human, drone and vehicle based on range doppler maps are investigated. The machine learning and deep learning model architectures used were determined by considering the spatial and temporal characteristics of consecutive range doppler images. The publicly available RDRD dataset is used to train and evaluate the proposed model architecture. The dataset contains range doppler images of different types of targets, such as humans, vehicles and drones, collected using an 8.5 GHz frequency modulated continuous wave radar. These images, which are obtained by performing certain signal processing steps for vehicles, drones and humans traveling along a given path, are stacked together consecutively for each class to form an image stack. There are several image classification techniques for processing these images, including machine learning and deep learning methods. In this context, machine learning based Random Forest and XGBoost algorithms as well as deep learning based 2D Convolutional Neural Network and Visual Transformer architectures are comparatively evaluated. Specific to these architectures, Random Forest architecture provided 91,45%, XGBoost obtained 92,54%, custom developed 2D Convolutional Neural Network resulted 94,34% and Visual Transformer architecture achieved 93,19% classification accuracy on the test dataset.