BEHAVIOR RECOGNITION WITH LSTM DEEP LEARNING MODEL AND MEDIAPIPE

Tran Song Toan; Minh-Hai Le; Huu-Phuc Dang

Authors

Tran Song Toan Tra Vinh University, Vietnam
Minh-Hai Le Tra Vinh University, Vietnam
Huu-Phuc Dang Tra Vinh University, Vietnam

Keywords:

deep learning, human behavior recognition, long short-term memory (LSTM), MediaPipe

Abstract

Human behavior recognition is crucial for assisting and monitoring the activities of patients, particularly, the elderly or young children. With the advancement of technology, modern methods based on computer vision have been developing. Deep learning is one of the
prominent methods for dealing with problems related to behavior recognition. In this study, a long short-term memory deep learning model is used for identifying abnormal behaviors. The MediaPipe library is used to collect body points and consecutive frames to generate training data and recognition. The behaviors considered in this paper include headache, stomachache, fall down, and others. With the dataset collected from videos and self-recorded, the experimental results show that the proposed long short-term memory
network model achieves 94.54% accuracy in behavior recognition. This result demonstrates the feasibility of the proposed model for the task of behavior recognition.