Don't use PCA for dimentionality reduction or feature selection in case of videos or images. Instead use some feature extraction techniques in images like HOG features, sift features, optical flow(for videos), etc. Since the datapoints are very less(assuming 25 datapoints), you could use some data augementation techniques for videos.
For example, crop some random sequence from the video for temporal augmentation then use Image augmentation for each frame.
For Image augmentation, you can refer to documentation here: https://www.mathworks.com/help/deeplearning/ref/imagedataaugmenter.html
Hope it helps !!! :)
