Human Behaviors Classification Using Deep Learning Technique

Abstract

AbstractHuman behaviors is an action performed by human. There are various types of human behaviors such as running, walking, jumping, sitting and the others complex movement. In this paper, human behaviors video-based classification using Long Short Term Memory (LSTM) model with multiple layers were proposed to classify the human behaviors. A pre-trained pose estimation model, OpenPose was used to extract the body key points from the Berkeley Multimodal Human Action Database, MHAD database. Six activities, jumping, jumping jacks, punching, waving with two hands, waving with right hand and clapping hands of MHAD database were used for the training and testing. The individual frame of MHAD database will group into 32 window width. Dataset had been increased by creating the 26 of 32 frame overlapping. The performance of 2 layers LSTM model, 3 layers LSTM model, 4 layers LSTM model without dropout layers and 4 layers LSTM model with dropout layers were evaluated and compared. Result shows that 4 layers LSTM model with dropout layers had better performance as compared to 2 layers LSTM model, 3 layers LSTM model and 4 layers LSTM model without dropout layers reached the testing accuracy of 95.86%. With adding of dropout layers in the LSTM model with 4 layers, generalization performance in training process had been increased.KeywordsLSTMMachine learningPose estimationHuman behaviorsAction classification