Mahdi Zaman

I am an engineer and researcher, applying my interdisciplinary experience to medical imaging, computer vision, multi-agent intelligence, and communication. As a researcher, I am currently affiliated with HadleyLab, UCF College of Medicine and CAVREL-UCF, College of Engineering and Computer Science. I leverage deep learning architectures to develop models for strong visual recognition in medical applications (3D volumetric segmentation, precision surgery) and for collaborative perception in connected agents (3D object detection). I have also co-developed protocols for vehicular applications such as tele-operated driving, automated tolling, and dynamic control.

I am humbled to get a chance to collaborate with industry researchers from Microsoft, Ford, Honda and Hyundai. Currently, I am in the process of finishing up my PhD and on the lookout for interesting opportunities to join full-time . Please reach out for potential collaboration opportunities or to share ideas; my preferred mode of communication is Email, LinkedIn, X - in respective order.

Email  /  CV  /  Google Scholar  /  Github  /  LinkedIn  /  X

My core curiosity is to explore our abilities as a species, currently exploring vision. I enjoy thinking and working on algorithms and systems aimed at robot vision and learning; at both micro and macro scale. In my spare time, I enjoy learning diverse concepts (current points of interest: evolution, fitness, neuroscience, psychology, spirituality) and experiment on myself; chasing the core of epistemology where it all comes together. I often share my summarized thoughts and personal experiences through my blog.

Research & Projects

My research projects have been wonderful opportunities to learn about medical imaging, robot perception, remote driving (both automated and humandriven), scaling vehicular communication and writing product-oriented systems for specific applications. Currently I am working on the highlighted projects.

Health Science & Medical AI

I develop AI-driven models for surgical video analysis and medical image processing, focusing on action recognition, operation outcome prediction, and efficient deep learning architectures for 3D medical imaging.

Developed a novel transformer architecture utilizing frequency domain representation of high-dimensional features for multi-resolution feature extraction. Our model achieves state-of-the-art performance in 3D volumetric segmentation for organ and tumor detection from CT scans and MRI images.

Working on an action recognition model for robotic surgery videos to assess procedural steps and predict operation success/failure, leveraging VLM architectures and temporal analysis.

Perception & Control for Autonomous Agents

Working on scene understanding / representation learning, for vehicles / warehouse robots / agricultural tractors to make better decisions.

Enabling remote driving in controlled terrain for manufacturing/warehouse/parking applications with V2X connectivity. We developed novel medium access protocol for no-loss communication in indoor warehouse which proved to be reliable on synthetic scenarios. We're currently scaling for larger vehicle capacity with equivalend bandwidh-efficiency.

Developed channel-efficient feature generation, sharing and fusion to efficiently leverage visual cues from neighboring agents in a multi-agent setting. Our model is currently being tested on several collaborative perception datasets for proving its generality and scalability.

We present a novel neural network model to leverage local (from on-board sensors) and look-ahead (via V2X) perception for efficient steering maneuver. It empowers the host vehicle with enhanced safety and prior knowledge for navigating a wide range of road scenarios.

Infrastructure-assisted Tolling

An infrastructure-assisted transaction procedure is presented. Potential use-cases are: toll collection, road user charging, remote driving, automated valet parking etc. The application leverages high-speed communication via Cellular-V2X. Outcomes aim for development of SAE J3217 standard.

Scaling V2X

Vehicle-to-everything (V2X) communication enables vehicles to share their cognition and constitute a form of mass intelligence in partial or fully autonomous traffic environment to overcome the limitations of a single agent planning in a decentralized fashion. Specifically, our proposed methods enable the 3GPP C-V2X communication technology to handle thousands of vehicles in heavily congested environments.

System Design & Modeling

We leverage Model-Based Communication (MBC) and propose a solution that enables cooperative control of vehicle platoons under non-ideal communication scenarios.

We model the inter-vehicle links in a platoon with a first-order Markov model to capture the prevalent temporal correlations for each link.

We propose a combination of control-aware communication and model-based communication. Our proposed solution reduces communication overhead by ~47% while maintaining nearly the same level of efficiency (less than 1% speed deviation) in cooperative adaptive cruise control.

We propose a modular architecture with separate subsystems for application and perception with a novel non-parametric Bayesian inference-based prediction method. We validate the architecture in conjunction with the prediction mechanism with real environment using Denso On-Board-Unit (OBU). The proposed system shows enhanced immunity to communication loss in V2X channel.

We introduce a real-world maneuver-based driving dataset that is collected during our urban driving data collection campaign.

Patent

We developed a novel 1-shot transmission scheme for Semi-Persistent Scheduling (SPS) that improves the latency and reliability of connected messages, independent of SPS-allocated resources.

Miscellanea

EEL 6938 Special Topics, Spring 2025, University of Central Florida

EEE 6712 Modeling and Analysis of Networked Cyber-Physical Systems, Spring 2024, University of Central Florida

Spring 2023, University of Central Florida

Originally stolen from Jon Barron's amazing website (source code). Feel free to repurpose.