Person Tracking Using AI and Surveillance Cameras

Abstract

The rapid development of urban infrastructure and the growing demand for stricter safety measures have made video surveillance an essential part of modern safety management. But a basic problem with traditional surveillance systems is the need for human operators to watch, assess and follow people in large camera networks. This project “Person Tracking Using AI and Surveillance Cameras” is aimed at solving these major in-efficiencies by building a complete automated system for real-time person tracking in multiple non-overlapping camera views. The main purpose of this work is to keep “Identity Continuity” i.e. to identify that a person entered in the field of view of one camera is the same person who entered in the field of view of another camera moments ago. The core of the proposed approach is a complex AI pipeline combining deep learning based Re-Identification (Re-ID) with state-of-the-art object identification. For the detection part, we used the YOLOv10 (You Only Look Once version 10) architecture. Unlike previous versions YOLOv5 and YOLOv8 which have computational overhead caused by Non-Maximum Suppression (NMS), YOLOv10 uses NMS-free training. Therefore, our system can achieve high speed real-time person detection with low latency even on resource-limited hardware. This is to ensure that the system can handle multiple HD streams at the same time, without frame drops, at high security environments. The most innovative feature of this project is the Re-ID Manager which uses a ResNet18 backbone to overcome the “Tracking Gap” in tra-ditional surveillance. When a person is detected, the system extracts a 128- dimensional feature embedding (a digital signature) based on visual characteristics such as clothing patterns, color, height, and body shape. These embeddings are stored and compared with Cosine Similarity measures. This takes away the privacy-intrusive nature of facial recognition and allows for cross-camera tracking. The technology generates a contin-uous record of movement across various physical areas by automatically assigning the previous ID to the person when the similarity score betweenv a new detection and a saved embedding is above a calibrated threshold. We developed a full-scale Multi-Camera Dashboard with the PySide6 framework that allows the security staff to use this technology. The interface is multi-threaded, so that the heavy-duty AI processing is done in the background and the user interface stays responsive. The interface displays up to four video feeds simultaneously and tracks individuals with persistent ID boxes. We also added a database management system, SQLite3, to log all tracking events. This data base stores the unique PersonID, CameraID and timestamps to allow a searchable historical record, which may be used for forensic research or post-event inquiry. A variety of indoor settings, such as hallway transitions and different lighting conditions, were used to test the implementation. Compared to conventional tracking-by-detection techniques, preliminary results show a high level of identity retention stability and a notable decrease in IDswitching mistakes. The project successfully connects the dots between unprocessed video footage and intelligent, useful information. Future research will concentrate on optimizing the system for densely populated areas and incorporating 2D path reconstruction, which entails mapping these digital IDs onto a top-down floor plan. This technology improves the overall accuracy and dependability of monitoring equipment in establishments like shopping centers, hospitals, and educational institutions while greatly reducing the workload for security teams by automating the tracking process.