Incorporating Vehicle Detection Algorithms via Edge Computing on a Campus Digital Twin Model

Abstract

A digital twin involves a realistic digital model, sensing of real-world parameters, depiction of those parameters in a digital environment, and simulation of the virtual system. While various methods exist to create digital models of geographic locations and their road networks, little work has been done to automatically track vehicles from video and generate virtual vehicles in the digital model. This paper presents one pipeline for performing vehicle detection and tracking on an edge device, and the results are visualized in a university campus digital model running on a workstation. Challenges stemming from camera lens effects, irregularities in the digital model, and object tracking uncertainties are discussed. The precise transformation of 2D vehicle detection to 3D representation is the main challenge addressed in this work. The approach presented focuses on calibrating different parts of the scene separately. The focus of future work will be to make this calibration process semi-automated.