Hybrid Indirect-Direct Visual SLAM for Embedded Platforms (2025)
Developed GLidE-SLAM, a hybrid indirect-direct visual SLAM framework for embedded and severely resource-constrained platforms, featuring custom photometric image alignment and a vendor-agnostic GPU compute-shader pipeline with no CUDA dependency. The system achieved up to 4× speedup over a CPU baseline and enabled online SLAM on low-power commodity devices. The project is currently under review for IROS 2026; code, supplementary material, additional information, and benchmarks will be released after the review process.
Early VR SLAM Prototyping (2023)
An early prototyping stage of the work that later became Mesh2SLAM in VR, a fast geometry-based SLAM framework for rapid prototyping in virtual reality applications. This project focused on building and testing the core tracking and mapping pipeline in C++ and OpenGL from the ground up, including early GPU compute-shader experiments for SLAM acceleration.
Bayesian Multi-Object Tracking for ADAS (2020-2022)
Developed from scratch a classical, learning-free Bayesian Multi-Object Tracking (MOT) module for an Advanced Driver Assistance Systems (ADAS) LiDAR perception pipeline. The system used Kalman filtering, probabilistic data association, gating, and covariance propagation to track objects in real time, maintain identities through occlusions, and generalize robustly across varied driving conditions. The work included integration and testing under operational constraints, and was later extended toward radar-based and defense-related perception applications.
Object classes are color-coded; bounding boxes are shown only for moving objects.
Reimplemented a simpler version of the tracker using UKF (Unscented Kalman Filter) and some mosquitoes from a video from outside my office window. Video was shot at 60 FPS which, despite issues with auto-focus, tracking was relatively robust. All done in C++ and OpenCV.
Simulation and Training Systems (2019)
Developed VR-based simulation and training software for defense applications using real-time 3D engines. Due to NDA constraints, project details are limited, but the work involved real-time 3D training applications, custom interaction logic, and simulation behavior. I also contributed to experimental camera calibration, camera–LiDAR fusion, and 3D mapping work.
Best thesis award (2018)
My Bachelor’s thesis was my first substantial work with Artificial Intelligence (AI), focused on classifying LCD glass surface defects using Convolutional Neural Networks (CNNs) and multi-spectral/channel industrial images. The project required a full pipeline: inspecting and preparing a real imbalanced dataset, evaluating different channel combinations, applying sampling and augmentation strategies, training CNN classifiers, and analyzing performance through precision, recall, F1-score, and confusion matrices. While application-specific to industrial inspection, it gave me a thorough practical foundation in deep learning, data quality, and real-world computer vision.
Invention: Pinch/Grab (2011–2013)
While at Softkinetic Systems, I co-invented and helped develop the pinch/grab interaction concept for depth and RGB-D camera systems. This work was later patented and became a widely adopted interaction paradigm in VR, AR, and smart glasses.