Human action and posture identification are two of the most difficult challenges in computerized systems. Categorization of these tasks is particularly relevant to the health and robotics fields. It is feasible to define recurrent physical movements and postures by using technologies based on artificial intelligence. The kinect activity recognition dataset (KARD) and the microsoft research (MSR) Pairs dataset were utilized for this research. Classifying videos with convolutional neural networks (CNNs) can introduce some difficulties. Training performance of video-based data with CNN models is generally lower than image data. This study makes a contribution of different perspective to the detection of motions by using frames extracted from video-based datasets of CNN models. It has brought human poses to the fore in order to prevent unnecessary images in the obtained frames. The images of the videos were extracted sequentially at a certain frequency. Then, the poses of the human activities in each frame were extracted in accordance with the order of their occurrence and all the frames in each video were combined into a single image format with respect to their order. Input data was trained by CNN models and classified according to motion types. In the classification process, the performance of all CNN models was increased and the training time of the models was saved. The proposed approach achieved 100 % success in classifying the two datasets according to the activity types.