Hey, I'm Michael

Program Profile

Interdisciplinary Master’s program in Robotics, Cognition, and Intelligence, combining Informatics with Electrical and Mechanical Engineering. The program focuses on intelligent robotic systems that perceive, learn, and act autonomously, with strong emphasis on AI-driven perception, learning-based decision making, and human-robot interaction in real-world scenarios such as autonomous driving.

Core Expertise and Skills Acquired

I developed a solid theoretical and practical foundation in robotics and intelligent systems, with a clear specialization in perception and learning:

• Computer Vision & Perception: Autonomous driving perception, multi-view geometry and 3D reconstruction, object detection, segmentation, and tracking.
• Sensor Data Processing: Sensor fusion and image-based perception for autonomous systems.
• Machine Learning & AI: Mathematical foundations of machine learning, including optimization, probability, and statistical learning theory, along with supervised and unsupervised methods.
• Real-Time Systems: Timing constraints, scheduling, and reliability in safety-critical systems, complemented by parallel and concurrent programming concepts for performance-critical applications.
• Robotics & Control: Classical robotics, kinematics, sensing, and system integration.

Professional Orientation

Prepared for research and industry roles in robotics, autonomous driving, and intelligent systems, with applications across automotive, aerospace, medical technology, and consumer electronics.

Program Profile

Bachelor’s program in Informatics with a strong emphasis on software engineering and the design of robust, scalable software systems. The curriculum focused on systematic software development, clean code principles, and architectural design, preparing for professional software engineering roles in complex technical environments.

Core Expertise and Skills Acquired

I developed a strong foundation in modern software engineering practices, covering the full software lifecycle from design to deployment:

• Software Architecture & Design: Clean code principles, modular system design, architectural patterns, and maintainable codebases.
• Software Quality & Testing: Systematic testing strategies, software quality assurance, and reliability-focused development.
• Project & Process Management: Structured software project management, collaboration, and requirement-driven development.
• Distributed & Cloud Systems: Cloud-based architectures, distributed systems, and scalable backend services.
• Embedded & Mobile Development: Software development for embedded systems and mobile platforms, with focus on resource constraints and system integration.

Professional Orientation

Prepared for professional roles in software engineering, backend and distributed systems, and embedded software development, with strong emphasis on code quality, architectural thinking, and engineering best practices.

Multimodal Pseudo-Labeling for 3D Object Detection in Autonomous Driving

Designed and implemented a pipeline to improve 3D object detection under limited or noisy LiDAR annotations by leveraging camera information. The thesis focuses on camera–LiDAR fusion, pseudo-label generation, and domain adaptation across datasets, targeting scalable training of 3D detectors with reduced labeling effort.

Key Contributions and Achievements

• Developed a camera-supported pseudo-labeling pipeline combining monocular depth estimation, 2D detection/segmentation, and LiDAR-based 3D detection.
• Generated pseudo-LiDAR point clouds from monocular images using state-of-the-art depth estimation and depth correction, enabling LiDAR-free or LiDAR-augmented training.
• Implemented projection- and segmentation-based label refinement, updating 3D classes using 2D semantic evidence with robust matching strategies.
• Integrated and extended OpenPCDet and nuScenes, including custom evaluation (e.g., range-based metrics) and efficient annotation processing.
• Analyzed domain gaps across datasets and evaluated semi-supervised and unsupervised adaptation strategies.
• Achieved measurable improvements in detection quality and label consistency, demonstrating the effectiveness of multimodal supervision with minimal manual labeling.

Strong emphasis on clean, modular research code, reproducible experiments, and systematic evaluation aligned with real-world autonomous driving constraints.

Focus Areas

The role involved developing and integrating modular C++/Python components for sensor and image processing systems running on Linux. ROS2 was used to structure and manage the distributed data flow between sensors and processing algorithms.

Simultaneously, I focused on specialized on tiny object detection, managing the full lifecycle from training and data annotation to deploying optimized State-of-the-Art models on target hardware.

Key Contributions and Achievements

• Object Detection Pipeline: Designed, trained, and deployed optimized State-of-the-Art Tiny Object Detection models (e.g., for aircraft recognition) tailored for specific sensor data and domain constraints.
• Data and Annotation: Managed, annotated, and augmented custom datasets to support the training and validation of computer vision models.
• Sensor Processing: Developed modular C++ and Python components using ROS2 applications for efficient processing and analysis of various sensor data streams.
• Refactoring and Optimization: Conducted major refactoring of existing software modules to significantly improve performance, reduce latency, and increase long-term maintainability.
• Cross-Platform Deployment: Integrated and tested software components on diverse target systems, including ARM (Jetson), x86, and custom hardware, managing hardware connection and interfacing.
• Quality Assurance: Implemented robust and automated testing frameworks using GoogleTest (C++) and PyTest (Python) to ensure code reliability and system stability.

Focus Areas

The role centered on benchmarking and efficiently deploying deep neural network based object detectors on limited hardware, such as edge devices. This involved designing a generic evaluation framework, optimizing state-of-the-art models, and performing cross-platform comparison of vendor-specific hardware and software optimization pipelines.

Bachelor Thesis & Publication

• Thesis Title: “Evaluation and Optimization of Deep-Learning based Object Detection on Embedded Inference Hardware from Xilinx and Nvidia”. Grade: 1.3
• Publication: A Framework for Benchmarking Real-Time Embedded Object Detection, DAGM GCPR 2022.

Key Contributions and Achievements

• Generic Benchmarking Framework: Designed and implemented a novel, lightweight framework to connect a host evaluation PC with multiple embedded target devices.
• Comprehensive Evaluation: Separated data distribution and evaluation to simultaneously measure accuracy, runtime, and power consumption without influencing the target device's processing.
• Optimization and Cross-Platform Deployment: Utilized and compared vendor-specific optimization pipelines (Nvidia TensorRT for Jetson AGX Xavier and Xilinx Vitis AI for ZCU104) and applied techniques like quantizatioon and pruning to analyze accuracy vs. speed trade-offs of deployed models.
• Embedded Systems Implementation: Implemented the target C++ software and configured Embedded Linux environments for the direct deployment and verification of Deep Learning models on operational FPGA/GPU platforms.