JEN-HAO YANG

Utilize object detection models and object tracking algorithms to perform real-time people and traffic flow statistics on camera streams.
Independently design the system pipeline architecture.
Started using Deepstream and employed C and C++ as development languages to maximize hardware performance and maximize recognition speed on edge devices.
Real-time object intrusion detection:
Set up and maintain an CVAT image annotation platform, transitioning the company from standalone annotation software to an online collaborative platform that allows multiple users to work together on annotations.
Developed a CVAT plugin for reading Darknet weights, enabling model-assisted annotation and significantly accelerating the efficiency of student workers' annotations.
Wrote Python, Bash shell scripts, and Docker files to standardize and automate the company's Darknet training process as much as possible.
Real-time people and traffic flow calculation:
Utilize AI to detect object intrusion in camera image streams and send alerts to the ELK system.
Employ deep learning models for object detection and design AI image detection systems.
Designed a modular system to maintain system flexibility and incorporated multi-threading (multiprocessing) to meet varying throughput requirements.
Lung nodule detection project:

Responsible for the company's AI medical imaging recognition - lung CT scan nodule detection Proof of Concept (POC) project, optimizing the Deep Learning Pipeline to locate suspicious nodules in lung images.
Performed data cleaning starting from raw CT files and doctor annotations, then proceeded to train models by preprocessing and feeding images into the model, using TensorBoard to monitor training progress.
Previous work experience:
In past work experiences, I designed objects using the Strategy pattern, creating a common interface for objects but with different implementations.
Introduced the Observer pattern to establish a signal messaging mechanism between objects.
Introduced the State machine to define different behaviors for objects in different states.
By incorporating design patterns, the project achieved significant decoupling and vastly improved scalability.
Skills:

Nvidia Deepstream development.
Proficient in Python, C++.
Used Yolo object detection models and trained Yolo models using Darknet.
Utilized Fastapi to wrap computational models into WebAPIs for general user use.
Ubuntu development environment.
Use Docker and Docker Compose for service deployment.
Used git and GitLab for version control.
Use Shell Script for handling repetitive tasks.