Anh-Quan Pham

Improved DMSTrack, a differentiable multi-sensor Kalman filter for 3D cooperative tracking, by modifying its learnable covariance module to enforce nonnegative uncertainty estimates. Our changes, ReLU-clamped residuals and reduced initial covariance scale, resulted in faster convergence and improved tracking performance on the V2V4Real benchmark. Implemented in PyTorch with Open3D visualization, trained end-to-end on real-world multi-vehicle LiDAR sequences using a custom differentiable Kalman filter pipeline.

Enhanced diffusion policy frameworks for robotic manipulation tasks by introducing transformer-based observation and action encoders to capture relationships across the whole trajectory. The algorithm is evaluated in a PyMunk-based PushT simulation environment, with comprehensive evaluation metrics.

Designed and implemented planning and control algorithms for the Franka Emika Panda 7-DOF robot to efficiently pick, place, and stack static and dynamic blocks for maximum height. Utilizing ROS and Gazebo for simulation.

This project implements a NeRF-based framework that represents a 3D scene in a compact way and render novel views given a number of 2D views of the same static scene. It integrates positional encoding, stratified sampling, and volumetric rendering.

Designed and built a custom quadcopter drone from scratch for environmental data collection and autonomous flight research. The system integrates mechanical, electrical, communication, and control components including a laser-cut mica frame, BR2204 brushless motors, Mamba F405 MK2 flight controller, and multiple sensors (DHT22 for humidity, BMP280 for atmospheric pressure, GP2Y1010AU0F for air quality). Implemented PID control algorithms for stable flight and used Betaflight firmware for sensor calibration and motor configuration.

Developed a real-time face verification system in C++ using OpenCV, integrating yunet.onnx for detection and sface.onnx for feature extraction and matching. Optimized for edge devices, ensuring efficient performance with minimal latency and low power consumption, suitable for IoT deployments.

A real-time emotion recognition system was built using Python, leveraging Mediapipe holistic solution to detect and track face and hand landmarks. The data was normalized to enhance model performance, ensuring it was position-invariant. Computer vision techniques were applied for real-time object detection and tracking, and the emotion recognition model was implemented with Keras. Videos were captured via OpenCV, processed with Mediapipe, and fed into the trained model for real-time mood prediction, with a Streamlit-based web interface and WebRTC for live video streaming.

A virtual machine that simulates the LC-3 architecture, providing a simplified CPU model. It reads, writes data, performs arithmetic, and interacts with I/O devices, making it ideal for learning system-level programming. Similar to the Java Virtual Machine, it abstracts hardware, allowing programs to run across devices without modification. Written in C, this LC-3 virtual machine was tested with 2048, hangman, and rogue game files (also uploaded in the repo).