Real-Time Multi-Camera Panorama Image Registration-Based Object Detection and Tracking System

Panoramic image stitching integrates multiple images into a wide field-of-view representation and is widely used in fields such as autonomous driving and surveillance. As the need for reliable object detection and tracking across wide scenes increases, multi-camera systems have become common; however, independently detecting and tracking objects on each camera and merging the results introduces structural limitations. To address this issue, this study proposes a system that stitches multi-camera inputs into a panoramic image and performs object processing within a single unified view. The system employs cylindrical projection, SuperPoint feature extraction, BruteForce matching, RANSAC-based outlier removal, affine transformation, and alpha blending to generate a seamless panoramic image. Object detection is conducted using TensorRT-optimized YOLOv5, and tracking is performed with DeepSORT. Experiments on the NVIDIA Jetson AGX Orin demonstrate accurate object detection and stable tracking performance within the panoramic environment.