Hybrid CPU-GPU acceleration of a multithreaded image stitching algorithm

Real-time image stitching is critical, especially in un-manned aerial vehicles, and its acceleration has received attention in recent years. This paper describes an image stitching acceleration scheme for heterogeneous (CPU+GPU) devices. Acceleration is attempted with both multithreading and multiprocessing. Most time-critical functions in the algorithm are offloaded on to the GPU. We crafted a 3-buffer ping-pong mechanism for synchro-nization and data transfer among threads/processes in order to maximize CPU utilization. We carried out our experiments on Nvidia Jetson AGX Xavier. Results show that more than 3x acceleration is achieved.