This is an interesting idea but have you conducted any entropy tests by generating a class of passwords from a family of videos and seeing how different they really are (e.g. character freq)? Without any hard numbers I can pretty confidently assume that the colors (8 bit unsigned perhaps mean R, G, and B pixel values) in a video are not uniformly/randomly distributed in the color space and subsequent frames of the video are also highly correlated. Not to mention you specifically throw away any non ascii character so a large portion of the UTF-16 space is not even allowed). Unless by chance, those you throw away happen to also be the super common int values, I feel that it's likely the entropy of these passwords is going to be surprisingly low.
What do you think? The entropy test would make a great blog post.