The general overview of the proposed action recognition framework is shown as follows. It can be regarded as two stages. First, i project each channel signal as a square matrix using modified RP(recurence plot) respectively and combine them into a color image. After normalization, we implement a tiny ResNet to do the classification task end to end.
Dataset consist of 42 persons, 35 persons to training and 7 persons to testing. Each of person have 12 gestures. Graph.py get data from csv file and generate graph.
Matrix.py get data from csv file and generate image, dataset has 4,379 Files. For example at a single measurement of 1 person with 12 gestures.
Finally, divide the above dataset into 2 parts and put it into Resnet-18.
This is my result: 76%. Not pre-trained.
