Towards a Real-time Overlapping Multi-camera Single-Target Tracking System with Feature-based Camera Selection

Multi-camera systems have gained prominence in computer vision for the potential to further expand the range of applications to larger and more complex real-world scenes. Commonly used in indoor CCTV networks, Overlapping Camera Systems (OCSs) offer comprehensive coverage for indoor monitoring and surveillance, however selecting the optimal camera for single target tracking remains an open problem. To address this, the study implements a system that is capable of extracting both visual and spatial features of a target from each camera view and utilizing a Multi-Layer Perceptron (MLP) model for automated camera selection. A real-world video dataset encompassing varying occlusion and target density levels was collected. The implemented object detection model, YOLOv8, achieved approximately 98% performance across all cameras, even under frequent occlusion. However, limitations in the re-identification model, OSNet, were identified, particularly for handling similar target appearances and extreme occlusions. The MLP camera selection model demonstrated promising results with 73% ac-curacy, providing an analytical approach to camera selection based on its target features. The contributions of this research to multi-camera tracking include: (1) developing a feature-driven camera selection approach for single target tracking in OCS, (2) demonstrating the effectiveness of a dedicated model for camera selection in dynamic indoor environments, and (3) designing a modular framework for target tracking and camera selection within an indoor calibrated space which can be configured to the needs of the end user.