Real-Time Crowd Analysis for Resource Management Using YOLOv8

Abstract

This paper presents a actual-time, drone-enabled crowd tracking machine designed for efficient resource management and protection in temple environments in India. The machine leverages YOLOv8 for robust object detection, enhanced with a grid-based location-clever method to improve accuracy in crowded and overlapping scenes. To reduce manual labeling, a vulnerable supervision technique is adopted, in which “silver” labels are generated the use of multiple labeling capabilities (e.g., motion detection, heuristics, and small classifiers). these noisy labels are combined using a label aggregator to create a training dataset, extensively lowering annotation prices. For green runtime overall performance, we implemented 2 - stage inference pipeline is implemented: a quick, light-weight classifier (stage 1) monitors frames to determine while and where to deploy the heavy YOLOv8 detector (stage 2), boosting effective frames per 2d (FPS) and scalability. The machine further contains active getting to know, periodically selecting uncertain or disagreeing frames for manual annotation to iteratively improve both the classifier and detector. OpenCV processes stay video streams from drone-hooked up or ground-based cameras, logging crowd statistics every 30 seconds in CSV layout for real-time analysis and emergency response planning. A Tkinter-based totally GUI allows picture-based crowd estimation, making the gadget versatile for both static and dynamic surveillance settings. The method demonstrates progressed scalability, speed, and accuracy as compared to standard methods, offering a practical answer for smart crowd management in temple and public areas.