<p>Human Activity Recognition (HAR) is a challenging task in computer vision that involves analyzing and detecting human actions for various applications such as healthcare and security. Indeed, HAR can be divided into two types: sensor-based and video-based. Sensor-based involves the use of sensors and machine learning algorithms to recognize human activities based on data collected from wearable devices or other sources such as accelerometers and magnetometers or video-based HAR refers to the process of recognizing human activities from video data using machine learning and computer vision techniques. There are different approaches for video-based HAR, but one common method is to use machine learning algorithms to classify the activity based on features extracted from the video frames. In general, HAR consists of two phases: data collection and processing and activity classification. In this paper, a new method is presented for recognizing human activities using body articulations, which are connections between bones in the skeletal system, such as the hand and shoulder. These joints allow for various degrees and types of movement. As contributions we used the MediaPipe algorithm in the first phase to extract key-point coordinates from human skeleton, and in the second phase which is activity classification transfer learning concept was employed to classify the extracted coordinates. The newly developed method was applied on the KTH, Weizmann, and Olympic Sports datasets and demonstrated higher performance in recognizing human activities, surpassing previous approaches.</p>