Dimitrios Chatziparaschis

I'm a graduate student in the ECE Department of the University of California, Riverside, working towards my PhD in EE under Konstantinos Karydis. Prior to that, I completed my MSc in Computer Science and Engineering and my Diploma in Electrical and Computer Engineering at the Technical University of Crete in Chania, Greece. In both my MSc and Diploma programs, I was honored to work with Prof. Michail G. Lagoudakis.

Email  |  Google Scholar  |  Github  |  LinkedIn

Research

My work lies at the intersection of computer vision, machine learning, and robotics. Main topics of my research include 3D perception, multi-modal sensing, landmark detection, and localization in outdoor and dynamic settings. In the past I focused on fully autonomous robotic applications for Search-and-Rescue, applying aerial and ground robot relative localization, vision-based object of interest detection, and decision making.

Demonstration of environment-aware robotic manipulation in dynamic target-reaching scenarios using a controller based on Dynamic Neural Fields (DNFs), Stochastic Optimal Control (SOC) theory, and monocular visual input.

Development of an algorithmic framework to perform real-time tree landmark detection and global association based on an underlying Kalman Filter for tree characteristics estimation and employed criteria for candidate matching based on association uncertainty.

Demonstration of a fully autonomous robotic platform for Stem Water Potential (SWP) measurements in avocado fields, based on Gaussian Processes (GPs) for modeling sampling area uncertainty and onboard leaf detection to perform 6D pose estimation and leaf cutting.

A lightweight keypoint detector based on a centroid point distance criterion, coupled with a multi-modal Non-Maximum Suppression (NMS) algorithm, to obtain salient and repeatable points in both colored and uncolored 3D point clouds.

Design and development of a robust semi-autonomous mobile robot solution to conduct soil apparent conductivity measurements in large fields, achieving high linearity in Pearson Correlation Coefficient (r) tests compared with groundtruth.

Development of a novel surveying approach using a custom-equipped UAV model, incorporating a gimbal with a monocular camera and a laser rangefinder. Ground target localisation based on multilateration, enhanced by a developed spatial filtering module to improve least-squares convergence speed and positioning accuracy. Real-time detection and ranging capabilities were demonstrated and evaluated in simulations, highlighting high applicability and localization accuracy across different environments and flight scenarios.

Development of a fully autonomous UAV behavior in Search-and-Rescue, utilizing machine learning approaches in both perception and control. The UAV's control and decision-making system was based on Deep Reinforcement Learning (DRL) and Deep Deterministic Policy Gradient (DDPG) models and was developed in a custom OpenAI Gym environment. For perception, object-of-interest detection was based on a proposed, designed, and trained feed-forward encoder-decoder network to perform pixel-wise classification and semantic segmentation.

Proposal of UAV system to assist surveys in GNSS-denied areas utilizing its clear-sky visibility of GNSS satellites. This study demonstrates the feasibility of this application through the proposal of system architecture and a theoretical evaluation analysis of positioning error.

Demonstration of an aerial-ground robot collaboration in Search-and-Rescue scenarios. By considering that the humanoid was “blind” in terms of sensing for localization, we introduced a UAV-assisted relative localization solution employing onboard multi-camera depth estimation with stereo correspondence and disparity filtering, 3D occupancy grid mapping, and Adaptive Monte Carlo Localization. The humanoid robot was equipped with a custom-trained YOLO detector for human detection and approached targets through a FootStep planner with R* search algorithm, guided by the UAV.