Detect Human Falls in an End-Edge-Cloud Orchestrated Architecture.

Activities Daily Living (ADLs) are diverse and complicated to define, but they can be classified as either static or non-static activities. Static activities involve less postural movement, such as standing and sitting, while non-static activities include falls and other dynamic movements. Falls can have detrimental effects, especially for old people, resulting in physical and emotional repercussions. To address this issue, we propose an end-edge-cloud architecture that accurately detect human falls. The end layer consists of a prototype based on an Arduino nano microcontroller board, which is worn around the waist. Additionally, it includes an MPU6050 sensor to collect data and a Bluetooth Low-Energy module for transmitting data to the edge layer. The edge layer uses an iOS-based application for buffer data storage and predicting activities. The cloud layer handles bulk data storage and updates. We evaluated the architecture’s performance using a publicly available dataset and our own measurements. The recorded ADLs encompass falling, jogging, lying down, sitting, standing, and walking. We employed algorithms, such as long short-term memory (LSTM), gated recurrent units (GRU), and convolutional neural networks (CNN), using binary and multiclass techniques. In the binary class, ADLs are categorized as either falling or not falling. In the multi-class, all six ADLs are used as inputs. Various metrics such as false alarm rate (FAR), accuracy, sensitivity, precision, and Fl score were evaluated for each algorithm. Our evaluation results showed that LSTM and GRU achieved the best performance in the multi-class, with FARs of 0.l% and 0.05% respectively, and other metrics ranging between 95% and 99.9%. CNN performed well in the binary class, achieving a FAR of 2.34% and other metrics ranging between 82% and 98.6%. Furthermore, we designed a CoreML framework to integrate models of the algorithms into the edge layer. The proposed architecture had an average delay of 110 milliseconds. Our study demonstrates the feasibility of accurately detecting falls using the proposed architecture.