Home-Based Real-Time Abnormal Movement Detection System Deployed on On-Device Artificial Intelligence

Abstract

In response to the aging trend in society and to Human Augmentation beings for home-based activities, this paper proposes an Abnormal Movement Detection system, using the common at-home movements of standing up and hand tremors while picking up items for abnormal movement verification. This can be easily applied in ordinary homes or long-term care institutions; for those living alone with limited resources, there is no longer any need to purchase expensive monitoring equipment to achieve improved quality of life. Therefore, this research collected and built the own dataset as the first important step of the study. The proposed Abnormal Movement Detection system is implemented by designing a deep learning network. Several issues, including the network architecture, the novel method of data augmentation and the scoring method of expanding the intervals between abnormality levels, are studied. For achieving the home-based real-time detection, there are four main contributions of this paper. The first is that a training dataset was collected and established: From this, the pathognomonic movement categories are easy to observe in home activities and geometric data augmentation can be used to improve the related home activity video collection. The second is the abnormal behavior detection architecture: This architecture has several important function blocks including detecting object, detecting action, inspecting abnormal movement and reminding event, using Convolutional Neural Network combined with Long Short-Term Memory ([Formula: see text]) as the core network for abnormal motion detection. With movement abnormality evaluation based on different levels, it can judge abnormal behaviors and conduct model training, performance evaluation and architecture optimization with both public domain datasets and the movement dataset collected in this research project. The third is the proliferation of new attributes in the videos: New attributes are added to the original videos through a Generative Adversarial Network (GAN), producing new training videos; the effectiveness of two different generation methods is evaluated. Finally, the algorithms developed in this paper are deployed on resource-constrained On-device Artificial Intelligence (AI). Activity videos from a total of 20 people were collected; in all, 53 videos of StandUp and 60 videos of PickUpItems were obtained to establish the training dataset. When CNN and LSTM network were added to Batch Normalization (BN), and Global Average Pooling (GAP) replaced Fully Connected (FC) layers, the accuracy rate reached 98.4%. In terms of data augmentation, geometric transformations and GAN were used to estimate the performance. The experimental results showed that the geometric transformation using brightness adjustment had the highest accuracy rate of 98.6%. Finally, the Softmax layer using Phi-Softmax–tan(⋅) function was shown to be the best method to expand the intervals between abnormality levels.