Peng-Chen Chen

Graduate Research Assistant at University of Michigan

Volumetric Flow Prediction & Sparse Sensor Optimization: Constructed a 3D flow dataset by interpolating Ansys CFD outputs into structured arrays for deep learning. Adapted the 3DShape2VecSet architecture to encode 2D flow fields into latent vector sets, utilizing a Transformer-based VAE and Fourier feature embeddings to capture complex fluid dynamics and reconstruct flow fields with partial samples. Designed a sparse sensor placement strategy using PCA analysis and QR pivoting to achieve ~75% flow field reconstruction accuracy with only 15 sensors in a factory environment.

Computer Vision Intern - AR/VR at Samsung Research America

Developed a real-time spatial mapping framework by integrating feature-based visual SLAM for precise camera tracking with a custom dense point cloud generation using Open3D. Built a depth prediction pipeline on Android devices using TensorFlow Lite and deployed the Depth-Anything-V2 model for on-device depth inference. Created an absolute depth estimation pipeline for monocular camera systems using RGB images and positional information, achieving robust performance with ~0.10 m MSE.

Research Assistant at C. Galban Lab, University of Michigan

Designed a novel deep learning method for pulmonary artery-vein segmentation in CT images, aiming to enhance diagnostic accuracy and streamline clinical workflow. Implemented and evaluated 3D U-Net segmentation methods on COPDGene datasets and applied post-processing (skeletonization, connected component analysis, and branch-based voxel majority voting) to generate pseudo ground truth data for model retraining.

Graduate Researcher - HybridDyn-VSLAM at University of Michigan

HybridDyn - VSLAM: Real-Time Visual SLAM for Dynamic Environments. Integrated YOLOv11n, FastSAM, and optical flow to filter dynamic features, achieving 83–97% improvements in Absolute Trajectory Error (ATE) compared to ORB-SLAM3 on dynamic scenes across TUM RGB-D and Bonn RGB-D datasets, and reduced inference time from 195 ms to 28 ms (~86%), enabling real-time dynamic SLAM.

Graduate Researcher - NeRF-based SLAM at University of Michigan

NeRF-based SLAM: Integrated Neural Radiance Fields into a SLAM system using implicit neural representations for real-time 3D mapping and camera pose tracking from RGB-D image inputs. Developed a custom MLP with positional embeddings and tuned photometric and geometric losses to improve scene rendering accuracy.