Real-time smart surveillance using YOLO-Faster R-CNN hybrid

Abstract

This research introduces a hybrid object detection framework integrating YOLOv5 and Faster R-CNN aimed at improving the accuracy and reliability of high-speed smart surveillance systems. YOLOv5, being a fast detector, works as the primary detector for multi-class detection of persons, backpacks, handbags, books, guitars, and cell phones. However, YOLO's deliberate design for speed can generate false positives or incorrectly placed bounding boxes from time to time. To tackle this, the system refines selected classes of persons, backpacks, and handbags with Faster R-CNN, a more accurate but slower-based model. First, YOLOv5 detects possible objects, and detections are filtered by class and confidence threshold. For select classes, RoIs are cropped and passed to Faster R-CNN for refined bounding box predictions and confidence evaluation. That finally outputs the detections with color-coded annotations and labels for each class to maintain clarity and robustness. This approach is an elegant balance of speed and precision, making it extremely competent for real-time surveillance in a dynamic environment. The modular design facilitates scalability and adaptation to other application domains where high detection confidence is required without grossly compromising performance.