Construction of a Crowd Density Analysis and Model Deployment Pipeline Based on Jetson Orin

I established a development workflow spanning from the training and comparison of crowd density analysis models to on-device deployment and execution validation. Along with the model's analytical performance, I implemented an operational architecture capable of applying the improved model to field devices and continuously modifying it.

CLIP-EBC · MLflow · Kubernetes · On-device AI

I developed a model that analyzes the number of people and crowd density from videos, and built a model deployment pipeline that can be executed and replaced on NVIDIA Jetson Orin. Participating as an undergraduate research assistant in a corporate project, I was responsible for AI model development and MLflow-based pipeline configuration.

Problem to Solve

The process of managing training results and applying them to actual devices was repeated every time the model was improved. A structure that makes it easy to reflect the tuned model into the actual system was necessary by connecting the experimental environment for comparing model performance with the on-device execution environment.

My Contributions

• Tuning and comparing crowd density analysis models
  I conducted training and performance comparison utilizing the ShanghaiTech A and B, JHU-CROWD++, UCF-QNRF, and NWPU-Crowd datasets. I selected CLIP-EBC, which showed the best performance among the compared models, as the final model.

• Building an MLflow-based model management and deployment pipeline
  I configured a pipeline utilizing MLflow to facilitate model modification and redeployment. I included an automatic data saving function and established a flow that connects the training results to the model used in the actual device.

• Application and validation of the Jetson Orin execution environment
  I optimized the model to run on Jetson Orin and verified the inference results and operations on the actual device. I validated the feasibility of execution in real-world usage environments by repeating model tuning and device application.