Hi, I'm Hoa Huynh

End-to-end in-cabin monitoring system built on a single NIR fisheye camera, compliant with Euro NCAP standards, and complemented with proprietary advanced features. Integrates up to eight deep learning computer vision models (object detection, image classification, 2D/3D body keypoints, facial landmarks, gaze estimation) into a C/C++ pipeline deployed on System-on-Chip devices (Texas Instruments SK-TDA4VM/TDA4AEN, Samsung Exynos V920). As Team Lead, owned the technical roadmap, task allocation, milestone tracking, and client demonstrations.

Demo confidential under client NDA — figures above are approved for public disclosure.

Real-time head pose estimation pipeline using classical computer vision (PnP solver, ZYX Euler angles). Recovers full rotation matrix and translation vector for pitch, yaw, and roll from monocular images, driven by a custom 2D facial-landmark model and stabilized with an Extended Kalman Filter to smooth motion and handle occlusion.

Camera-based LiDAR replacement using absolute depth estimation and camera calibration to generate pseudo point clouds from image pixels. Rendered in real-time 3D using OpenGL and Open3D. Designed a lightweight YOLOv8-based encoder-decoder depth model with accuracy competitive against SOTA methods (NeWCRFs, PixelFormer) at a fraction of the parameter count.

Detect and estimate worker positions using a single thermal (Boson LWIR) camera. Combined YOLOv5 object detection with NeWCRFs monocular depth estimation. As the first developer on the project, built the full calibration pipeline: camera calibration, data collection, and Camera-LiDAR calibration using an Ouster OS0 LiDAR. The resulting dataset and depth model were submitted as Team DepthSquad to CVPRW 2023's Second Monocular Depth Estimation Challenge — 1 of 8 submissions worldwide (out of 28 total) accepted for outperforming the SOTA baseline, beating it on image-based accuracy metrics (MAE, RMSE, AbsRel).

Proof-of-concept multi-camera real-time stitching system for ADAS, built as sole end-to-end developer. Implemented two stitching algorithms (simple and advanced) covering input streaming, cache-parameter registration, and panorama composition. Requires ≥ 30% FoV IoU overlap between each camera pair for reliable feature matching.

Detects whether a driver is looking at their phone by combining an object detection model with a gaze estimation model. Uses a cone-based FOV intersection algorithm with configurable confidence and count thresholds. The underlying intersection geometry has since been generalized into a standalone toolkit (linked).

Official implementation of the stereo-vision nighttime vehicle-to-vehicle positioning algorithm from my MDPI publication: gradient-boosted stereo matching for taillight-region correspondence, paired with a neural-network-based matching stage to associate left/right taillights under urban traffic conditions.

From-scratch simulations spanning the recursive-vs-batch spectrum of robotics state estimation on the SO(3)/SE(3) manifolds: IMU integration/pre-integration, point-cloud pose tracking (EKF vs. Invariant EKF vs. batch Gauss-Newton), pose-graph optimization with loop closure, and full bundle adjustment (joint camera-pose + landmark refinement, Umeyama-aligned to ground truth). Every method implemented twice — hand-rolled NumPy Lie-algebra math, and via the manif Lie-theory library — backed by a growing set of written theory notes.

A curated collection of pre-trained vision encoders (backbones) from various detection and classification models, ready to use as feature extractors for downstream tasks such as classification, detection, keypoint estimation, segmentation, and depth estimation.

A growing personal toolkit reused across multiple detection and calibration projects: fisheye camera calibration and undistortion, chessboard corner detection, YOLO↔Pascal VOC label conversion, and video frame extraction for dataset preparation.

A Python/OpenCV toolkit for estimating camera intrinsics and removing lens distortion, supporting both fisheye and pinhole models across RGB and thermal sensors. Handles checkerboard detection from live streams or batch images, with configurable quality thresholds and reprojection-error reporting.

A collection of PyTorch, TensorFlow, and R tutorial notebooks plus CSC321 (Univ. of Toronto) coursework, covering model building and training, transfer learning, RNNs/LSTMs, autograd, and deploying models via a REST API.

A face detection and blurring tool for video files. C++ backend (FFMPEG + inference) with a C# WinForms UI supporting drag-and-drop input.

A five-step pipeline for detecting ego-vehicle lanes from LiDAR point clouds: ground point extraction, intensity histogram peak detection, sliding-window lane search, and polynomial curve fitting to output left/right lane coefficients.

A C++ point cloud registration solution that aligns and merges multiple 3D point clouds to a master reference via notch region detection and correspondence matching, through preprocessing, detection, registration, and merging stages.

A C# reimplementation of the classic Minesweeper game — a fun exercise in GUI programming and game logic.