MediaPipe's Pose-Based Human Activity Recognition With LSTM

It is difficult to comprehend and categorize human movements in Human Activity Recognition. Despite being widely utilized, convolutional neural networks (CNN) might not adequately represent the intricacy of motions. This work investigates an alternative method for classifying human motion: the LSTM model combined with MediaPipe. Sequential processing is used by the LSTM model for keypoint sequences, which improves its accuracy in identifying complex human movement features. Although OpenPose provides precise body mapping, MediaPipe is used in this research because of its computational efficiency and ease of integration. Using CV2 libraries, preprocessed video data is read at 15 frames per second. MediaPipe software is then used to extract 33 keypoints coordinates from each frame. Then, in order to make these keypoints compatible with the LSTM model, they are compressed into single-dimensional arrays. The keypoints from each picture frame are taken out and combined with other frames for a total of 15 frames into a single sequence. The aim of this study is to evaluate the usefulness of the LSTM model in human action classification by contrasting its performance with more conventional sequence modeling methods. Following evaluation, we employed the Recognition of Human Action dataset from Kungliga Tekniska Hogskolan (KTH) University, which is used by LSTM models. Over 99% of the metricsAccuracy, Precision, Recall, and F1-scorewere attained by the model.