Date of Award

6-10-2025

Document Type

Thesis

Publisher

Santa Clara : Santa Clara University, 2025

Department

Computer Science and Engineering

First Advisor

David Anastasiu

Abstract

Multi-camera people tracking is the process of tracking persons and their paths, continuously across different camera fields of view. It can help track suspects across large areas, and assist individuals in emergency situations. The current state-of-the-art (SOTA) method involves using geometric-consistent constraints, information on the appearance of subjects, and pose estimation for dealing with occlusion issues. This current SOTA works well, but is still lacking in its ability to handle occlusion and perform well in real-time applications on a network. With occlusion, the IDs assigned to persons can be accidentally swapped in high density areas, or places where there are drastically different camera angles. In the case of real-time deployment of these methods, the models can suffer from the fact that they are not yet well optimized, and their performance and accuracy can be diminished at the cost of output quality. Using pose estimation models like RTMPose or HRNet could aid in reducing occlusion issues, and using lightweight models that are better suited for scaling while simultaneously maintaining accuracy, like OSNet, can help address performance issues and reduce compute requirements.

Download

Share

COinS