High Quality Real-Time Panorama on Mobile Devices

This paper proposes a real-time stitching algorithm for constructing panoramic images as large as 70 mega pixels (MP) on mobile devices. The conventional methods use feature matching, warping and multi-band blending, which pose processing bottleneck on mobile devices. For real-time operation, the proposed method applies low resolution tracking followed by novel methods of weighted patch alignment and dynamic Blending. As input frames are captured at a high rate (20fps) and contain large overlap, the camera motion is approximated to a translational model, thereby avoiding time-consuming operations such as warping. According to the experiments done on ARM Cortex-A53 based mobile device, the proposed method is able to generate more than 70MP panorama instantly, achieving 15 times faster performance compared to existing methods. The results show improvement in visual quality for both subjective and objective evaluation.