Real-Time Video Stabilization Using a Modified Optical Flow Algorithm

Camera shake degrades video recorded from vehicles, drones and handheld devices. This paper describes a real-time digital video stabilizer built on a modified optical flow algorithm, written in Python with OpenCV and run on a Raspberry Pi. To test it under controlled conditions we built a vibration rig from an Arduino UNO and an SG90 servo motor. Two cameras were mounted on the rig together. One fed the stabilizer, and the other fed a second Raspberry Pi that only recorded, which gave a shaky reference of the same scene. We recorded twelve pairs of 640 × 480 videos at 50 frames per second, each pair at a different vibration frequency, and compared them using average peak signal-to-noise ratio (PSNR) and video quality metric (VQM). Stabilized video had a higher average PSNR than shaky video at eleven of the twelve frequencies and an equal one at the twelfth. The mean gain was 0.4 dB and the largest was 1.1 dB, at 1.5 Hz. Side-by-side viewing of the paired videos also showed less visible jitter.