REAL TIME OBJECT DETECTION SYSTEM

Abstract

This paper presents a modular Real Time Object Detection System that combines Ultralytics YOLOv8, OpenCV, a reusable Python inference class, and a FastAPI web service. The system accepts webcam frames, static images, or browser uploads; predicts multiple object classes; applies confidence filtering and non-maximum suppression; and returns colour-coded bounding boxes, labels, confidence scores, object counts, and real-time throughput. In addition to the 80 categories available in the pretrained COCO model, the project provides a custom-class pipeline that collects images through DuckDuckGo search, removes low-resolution and duplicate samples, and automatically produces YOLO annotations. Known categories are localized by the pretrained detector, while genuinely new categories are verified with CLIP before weak bounding-box assignment. YOLOv8n is then finetuned through transfer learning and evaluated on a held-out split. The implemented CLI and FastAPI interfaces share the same ObjectDetector core, preventing duplicated inference logic. Documented results include detection confidence above 85% for clearly visible objects, 45-60 FPS on an NVIDIA GPU, and 8-12 FPS on CPU-only execution. The resulting architecture offers a practical foundation for surveillance, retail analytics, inventory recognition, and assistive computer-vision applications. Keywords— YOLOv8, object detection, real-time inference, OpenCV, FastAPI, CLIP, transfer learning, custom-object training.