Session: 02-05-01: Artificial Intelligence and Data-Driven Design 2

Paper Number: 193458

Real-Time Multi-Camera Vehicle Tracking on AI Edge Device for Intelligent Traffic Surveillance 

The increasing demand for real-time traffic monitoring and intelligent transportation systems has created a need for advanced vehicle tracking technologies capable of operating efficiently in large-scale urban environments. Multi-camera tracking systems are widely used to monitor traffic flow, detect incidents, and support intelligent traffic management. However, many traditional systems rely heavily on centralized cloud infrastructures, where raw video data from multiple cameras are continuously transmitted to remote servers for processing. This centralized architecture introduces several limitations, including high network latency, bandwidth congestion, and potential privacy risks associated with transmitting large volumes of video data. These limitations can significantly hinder the scalability and responsiveness of traffic monitoring systems, particularly in dense urban areas with numerous cameras and heavy traffic activity. To address these challenges, this study investigates the development of a multi-camera vehicle tracking system implemented on an AI edge device to enable real-time traffic surveillance while reducing network dependency.

The main contribution of this research lies in the design and implementation of an edge-based multi-camera tracking (MCT) system that integrates vehicle detection, vehicle re-identification, and spatial-temporal information for robust tracking across multiple camera views. By performing computational tasks directly on an AI edge device equipped with GPU capabilities, the proposed system minimizes the need for continuous data transmission to centralized servers. This architecture improves real-time responsiveness, reduces network congestion, and enhances data privacy by processing video streams locally. Furthermore, the proposed system improves vehicle tracking performance across both overlapping and non-overlapping camera views by combining appearance-based vehicle descriptors with spatial-temporal constraints.

The methodology of the proposed system consists of several key components. First, vehicle detection is performed using a retrained YOLO model that identifies vehicles in each camera frame with high accuracy and computational efficiency. Second, vehicle re-identification is conducted using a modified appearance descriptor derived from the DeepSORT framework, enabling the system to match vehicles across different camera views based on visual characteristics. Third, spatial-temporal information is incorporated to improve tracking reliability by considering vehicle motion patterns and transitions between camera zones. A global coordinate framework is also established to maintain consistent vehicle trajectories across multiple cameras. These components are integrated into an edge-based processing pipeline that performs detection, feature extraction, data association, and cross-camera matching directly on the AI edge hardware.

Experimental evaluations were conducted to assess the performance of the proposed system in terms of tracking accuracy, processing speed, and resource utilization on the AI edge device. The results demonstrate that the proposed edge-based MCT system achieves improved tracking accuracy compared with commonly used algorithms such as SORT and DeepSORT. In addition to improved accuracy, the system operates in real time by processing video streams locally, thereby reducing latency and eliminating the need for high-bandwidth communication with centralized servers. The ability to perform detection and tracking directly on edge hardware also enhances scalability, making the system suitable for deployment in large urban traffic networks.

Presenting Author: Hwan-Sik Yoon The University of Alabama

Presenting Author Biography: Dr. Hwan-Sik Yoon is an Associate Professor in the Department of Mechanical Engineering at The University of Alabama. His research interests include application of Machine Learning and Artificial Intelligence to Automotive and Manufacturing systems.

Authors:

Shenglin Li The University of Alabama
Hwan-Sik Yoon The University of Alabama