VIDEO COLOUR VARIATION DETECTION AND MOTION MAGNIFICATION TO OBSERVE SUBTLE CHANGES

Abstract: Our thesis work is based on revealing minor informative variations in a video which are hard to perceive, that can be further exaggerated to extract hidden variations of color and motions in a video. In our thesis we apply dierent techniques of a video decomposition like Laplacian, Steerable and Gaussian pyramids to observe the improvement in performance of the videos. We start with a standard input video to decompose it in dierent spatial pool of frequencies, the temporal ltering process is applied to the frames to extract hidden signals. The resultant signals from the temporal processing are then amplied by a given factor to reveal hidden information in the videos. These amplied signals are added back to the original signals and then a pyramid is collapsed to generate an output video. Performance of Gaussian and Steerable pyramids for video decomposition is evaluated over Eulerian motion magnication. The output videos from all pyramids decomposition is computationally analyzed and compared with each other through SSIM and PSNR graphs. The video processing time is used to compare decomposition methods. It is observed that Eulerian motion magnication with Steerable pyramid decomposition has potential of revealing hidden motions more than Laplacian and Gaussian pyramids, precisely in monitoring and diagnostic applications. Steerable pyramid decomposition method performs better than the other methods when input video is noisy.