VISION BASED HUMAN ACTIVITY DETECTION USING POSE ESTIMATION

Abstract

Human Activity Recognition (HAR) plays a vital role in various applications such as intelligent surveillance, healthcare monitoring, human–computer interaction, and sports analytics. This project presents a Vision-Based Human Activity Detection System using Pose Estimation to accurately recognize human actions from video data. The proposed system utilizes the MoveNet pose estimation model to extract human skeletal keypoints from video frames, which represent the positions of different body joints. From these keypoints, additional motion-based features such as joint angles and interframe velocity are computed to effectively capture body movements. Temporal sequences consisting of 40 consecutive frames are constructed to model dynamic motion patterns over time. These sequences are then used as input to a Long ShortTerm Memory (LSTM) neural network, which is designed to learn temporal dependencies and classify human activities. The model architecture includes stacked LSTM layers followed by dropout layers to prevent overfitting and dense layers for final classification. A Softmax output layer is used to classify activities such as walking, jogging, running, and hand-waving. The system also supports real-time activity recognition using webcam input, where predictions are stabilized using a sliding window voting mechanism to improve consistency and accuracy.