Human Fall Detection Based on Unidirectional Multi-Scale Spatial Encoder and Bidirectional Temporal Decoder Neural Network

Abstract

According to a statistic report from World Health Organization (WHO), fall is the second leading cause of death. They even add that residents with a disability who live in the nursing home or chronic care facility have a higher risk for falls. Approximately 30 to 50% of them had encountered fall every year, about 40% of them had more than once. Based on a statistic report released by the National Health Insurance Administration of Taiwan in 2018, fall is the second leading cause of death for people above 65 years old. Based on the population record, Taiwan will turn Super-aged society in 2025. The increasing demand and recent success of deep learning technology in the various research fields, especially computer vision, encouraged us to develop a deep learning based fall detecting system. Through this system, we aim to reduce the detection time of falls. A Unidirectional Multi-Scale Spatial Encoder and Bidirectional Temporal Decoder neural network is proposed to enhance performance for human fall detection. Additionally, we also utilize an auxiliary dataset "UCF50" to enhance the model's spatial recognition ability. The proposed method was able to achieve accuracy of 99.7% and F1-score of 99.97 at URFD.