Nguyen Mau Dung

Nguyen Mau Dzung loves developing AI-based real-time and accurate applications. Currently, he is a Machine Learning Engineer at Axon. Prior to joining Axon, he was a Senior AI Engineer at FPT Software and VinAI Research company in Viet Nam. In September 2020, he worked for Seoul Robotics (Top 100 startups in South Korea) as a Research Engineer who developed 3D perception algorithms by using LiDAR sensors.

He graduated Master's degree with a major in HCI & Robotics at Korea Institute of Science and Technology (KIST) (Top 6 innovative research institution in the world). During that time, he got the best paper award at the IEEE International Conference on Consumer Electronics (IEEE ICCE 2020) which was held in Las Vegas, USA. In 2011, he entered an honor program designed for talented engineers at Hanoi University of Science and Technology (Top 1 university in the field of science and technology in Vietnam). He graduated from the university with a distinction degree and at the top 2% of students in the school (Top 7 out of 428 students). In 2008, he enrolled in a high school for gifted students, specializing in mathematics. During that time, he got two third prizes in Vietnam national Physics contests for high school seniors (in 2010 and 2011).

EXPERIENCE

Machine Learning Engineer Axon (Vietnam Office) Nov 2022 – Present

Write code, save lives.

Solution Architect 5 (Senior AI Engineer) FPT Software, Viet Nam Aug 2021 – Oct 2022

Developed a lightweight parking slot detector running on a Jetson AGX Xavier at a frame rate of 25 FPS and with maximum errors of 3 centimeters of locations and 2 degrees of orientation;
Developed 2D/3D surrounding views, guidelines, and auto-calibration modules for the SVM system. The SVM system runs with real-time performance and impresses users with high-tech features and beautiful views.

Senior AI Engineer VinAI Research, Viet Nam Apr 2021 – Aug 2021

Developed perception module for the AutoPilot project;
Developed the localization module for the first self-driving car in Viet Nam (video link);

Research Engineer Seoul Robotics, Korea Sep 2020 – Mar 2021

Developed the lightweight classifier for LiDAR clusters. The classifier achieved an F1_score of 0.94 over the three classes such as cars, pedestrians, and cyclists. The model could classify 1000 clusters within 5ms on an Intel Core i5 processor (CPU only, no GPU needed);
Improved the LiDAR-based 3D perception algorithms.

Artificial Intelligence Researcher KIST, Seoul, Korea Mar 2018 – Aug 2020

Implemented the fast and accurate 3D object detection algorithm based on 3D LiDAR point clouds (source code) The Resnet-based Keypoint Feature Pyramid Network takes RGB-maps that are encoded by height, intensity, and density of 3D point clouds as the inputs. The model could detect well cars, pedestrians, and cyclists in a range of ±50m with a speed of 95 frames per second on a single GTX 1080Ti;
Implemented the real-time temporal and spatial video analysis system of table tennis using deep CNN networks from scratch. The model could run real-time with around 130fps on a GTX 2080Ti, detect ball positions with an error of 4 pixels in the full HD resolution, recognize events spotting (including bounce and net hit events) with an accuracy of 97%, and segment the human, table, and scoreboards with an IoU score of 0.96;
Estimated 3D human joint angles from multi-view images/videos. We built a new dataset that includes the images captured by 28 digital cameras and the ground truth of 3D human joint angles extracted from the Motion Capture System. The hip joint angles, knee joint angles, and ankle joint angles could be estimated with a small range of errors around 2.34 degrees;
Classified three kinds of foot groups (normal, abnormal, and athlete) by using spectrograms transformed from raw IMU sensor data as the input of deep CNN networks. Seven IMU sensors were attached to the lower human body while 29 semi-professional athletes, 19 normal participants, and 21 participants with foot abnormalities walked on a 20-m straight path. The three-foot groups could be classified with accuracies of 97.58%, 98.19%, and 99.40% when using 1 sensor (only Pelvis sensor), 3 sensors (the combination of Pelvis and 2 feet), and all 7 sensors respectively;
Detected dangerous situations of elderly people in the bedroom based on 2D human body key points in images extracted by using the OpenPose algorithm. The system could raise a warning bell and send messages to their family members within 2-3 seconds;

Embedded Software Engineer HUMAX VINA, Hanoi, Vietnam Jul 2016 – Apr 2017

Developed set-top-box products. I was a member of the middleware team.

Research Assistant ESRC Lab/HUST, Hanoi, Vietnam Sep 2013 – Jun 2016

Developed algorithms to effectively fetch videos from servers via the HTTP protocol in order to improve users’ experience;
Monitored the Lab with 50 people in 6 research groups.

EDUCATION

Nanodegree Udacity Self-Driving Car Engineer Apr 2020 – May 2020

Master’s Korea Institute of Science and Technology (KIST) Mar 2018 – Aug 2020

Major: Human-Computer Interaction & Robotics GPA: 4.27/4.50
Thesis: Image-based 3D Human Joint Angle Estimation for Daily Activities

Bachelor Hanoi University of Science and Technology (HUST), Vietnam Sep 2011 – Jun 2016

Honor Program for Talented Engineers

Major: Electronics and Telecommunications GPA: 3.58/4.00
Grade: Distinction Rank: 07/428 (top 2%)

High School Hatinh High School for Gifted Students (Specializing in mathematics) Sep 2008 – Jun 2011

SELECTED PRODUCTS

360-degree Surrounding View Monitoring System (SVM System) Sep 2021 - Present

A real-time 2D/3D views of the area surrounding cars;

Auto Parking Sep 2021 - Mar 2022

An autonomous car-maneuvering system that drives a vehicle into a parking slot by performing parallel, perpendicular, or angle parking;

LiDARs-based Smart 3D Perception Engine Sep 2020 - Mar 2021

3D perception software platform for Lidar sensors: Fuse, calibrate, and process LiDARs with ease;

The AI-based athletic performance analysis system Nov 2018 - Aug 2020

3D Human joint angle and joint position estimation based on multi-view images/videos;

SKILLS

Programming languages: Python, C++, SQL, Cuda;
Languages: Native Vietnamese, Proficient English, Basic Korean;
Deep learning frameworks: Pytorch, Tensorflow, Keras;
Version control systems: Git;
Operating systems: Linux, Windows, MacOS.

AWARDS

Top performance member of the year at NextG, FPT Software (2021)
The best paper award at the 2020 IEEE International Conference on Consumer Electronics (Las Vegas, USA) (2020);
The excellent student from the Dean of Hanoi University of Science and Technology (Vietnam) (2013, 2014);
Two third prizes in Vietnam national Physics contests for high school seniors (in 2010 and 2011).

PUBLICATIONS

[1] “Multiple Classification of Gait Using Time-Frequency Representations and Deep Convolutional Neural Networks”, in IEEE Transactions on Neural Systems and Rehabilitation Engineering, doi: 10.1109/TNSRE.2020.2977049;
[2] “IMU-based Spectrogram Approach with Deep Convolutional Neural Networks for Gait Classification”, 2020 ICCE (Las Vegas, USA);
[3] “Walking-in-Place Characteristics-Based Geriatric Health Assessment Using Deep Convolutional Neural Networks”, 2020 EMBC (Canada);
[4] “Deep Neural Network-based gait classification using wearable inertial sensor data”, 2019 EMBC, in Berlin, Germany;
[5] “An evaluation of segment duration effects in HTTP adaptive streaming over mobile networks”, 2015 NICS (Vietnam);
[6] “Quality-delay tradeoff optimization in multi-bitrate adaptive streaming”, 2015 ICCE (Las Vegas, USA).

PATENTS

[1] “Smart Walk Simulator for Elderly Healthcare”, Patent Registration No. 2019-0145553, South Korea;
[2] “Method and system for evaluating elderly health status using stepping characteristics”, Patent Registration No. 2019-0128445, South Korea;
[3] “3D human body joint angle prediction method and system through the image recognition”, Patent Registration No. 2019-0060025, South Korea.

PERSONAL TRAINING CERTIFICATES

Self-driving Car Engineer

Udacity (Nanodegree – Sebastian Thrun)

The Nanodegree has two parts: Computer Vision, Deep Learning, and Sensor Fusion; and Localization, Path Planning, Control, and System Integration. During the course, I did the end-to-end self-driving car projects, such as finding lane lines; traffic signs classification; behavior cloning; extended Kalman filters; kidnapped vehicles using particle filters; path planning on the highway; PID controller; and programming a real self-driving car using ROS.

AI for Medicine

Deeplearning.ai (Coursera Specialization – Pranav Rajpurkar & Andrew Ng)

The specialization has 3 courses: AI for Medical Diagnosis, AI for Medical Prognosis, and AI for Medical Treatment. These courses covered how to diagnose chest x-rays and brain scans, evaluate models, handle missing data, estimate the effect of treatments, and efficiently automate the task of labeling medical datasets using natural language entity extraction and question-answering methods.

Deep Learning

Deeplearning.ai (Coursera Specialization – Andrew Ng)

The five courses in the specialization are Neural networks and deep learning; Improving deep neural networks: hyper-parameter tuning, regularization, and optimization; Structuring machine learning projects; Convolutional neural networks; and Sequence models.

Machine Learning

Standford Online (Coursera - Andrew Ng)

The course covered (i) Supervised learning (parametric/non-parametric algorithms, support vector machines, kernels, neural networks); (ii) Unsupervised learning (clustering, dimensionality reduction, recommender systems, deep learning); (iii) Best practices in machine learning (bias/variance theory, innovation process in machine learning and AI).