The number of the world's population aged 65 or over (elderly) is projected to increase to almost 1.5 billion by 2050. Elderly is vulnerable to various risks on their daily activities, so it is necessary to recognize their actions with machine vision technology automatically. One of the methods to do action recognition is using Convolutional Neural Network (CNN). However, using CNN without preprocessing will result in poor classification accuracy. The preprocessing methods affect the performance of the resulting model. Therefore, it is necessary to research various image preprocessing methods on CNN input to get the optimal model. In this study, various preprocessing methods, namely resizing, enhancement, creation of binary and gradient images, and data augmentation, are compared. After that, the obtained models are evaluated using action recognition dataset. In the validation results, it is found that the best preprocessing method is 64×64 grayscale image preprocessing with sharpening and augmentation in the form of the horizontal flip, which achieves an accuracy of 0.852. Meanwhile, in the testing results, the preprocessing method that produces the best accuracy is the 64×64 grayscale image preprocessing with sharpening, with an accuracy of 0.660.