Efficient classification of physical exercises enables patients to evaluate their levels of physical activity and functional capacity to maintain their physical fitness and decrease the risk of chronic diseases. This paper presents the use of automated deep-learning classification for physical exercise based on the type of images. The MobileNet is used as a classifier for the collected dataset from the IntelRealsense camera to classify the physical exercise as correct or incorrect. The classifier is applied for three groups of images, colored images, images with human poses, and skeleton images. The result shows the highest classification accuracy (100%) for colored images than the images with pose (96.34%) and skeleton images (99.89%). The time required for classification when using human pose images is also calculated, as the time for each epoch was (35 s.), while the time required for the colored images and skeleton images was (115 s.), (and 120 s.) respectively.