An Enhanced Multiview Person Tracking Model with YOLOv5s Combined Hungarian Matching Operations

A rising demand for strong multi-camera person tracking in surveillance, retail analytics, and smart cities is being emphasized with real-time accurate and robust identity association across views affected in the process.Most of the current solutions suffer high computation costs or low generalization to appearance changes, and many of them have a limited scalability to MultiView configurations.Furthermore, many methodologies adopt complicated architectures that are unsuitable for real-time use or for the edge environment.In concern to these, present work lightweight, modularize the pipeline for cross-view person detection, re-identification, and tracking with front-top synchronized videos.The system comprises four key modules optimized for speed, accuracy, and interpretability: at First, YOLOv5s trains per frame detection for fast, compact object identifier trained on the COCO dataset.Speedaccuracy trade-off enables real-time inference.Active feature-extracting vector cores, OSNet x0.25 extract 512-dimension feature vectors, while input is using resized person crops of 128256.Efficiently cross-view appearance representation is maintained by the model while this section extracts scaled feature vectors.Next, identity association is done using cosine similarity of feature vector closeness and then optimally hooking globally by Hungarian algorithm.An empirical similarity threshold (0.7-0.75) filters out spurious matches, while achieving a mean matching accuracy of ~88%.Centroid-based tracking algorithms measure movements per ID in the top view, with dynamic distance thresholds that preserve temporal continuity and minimize ID switches.The system achieves >92% trajectory completeness with real-time operations, making it suited for indoor surveillance, retail behavior analysis, and crosscamera tracking.Modular design allows deployment on edge devices and generalization to diverse environments with minimal adaptations.