Application of KNN for Fall Detection on Qualcomm SoCs

Abstract

A fall of an elderly person often leads to serious injuries, even death. Many falls occur in the home environment, and hence, a reliable fall detection system that can raise alarms immediately is a necessity. Wrist-worn accelerometer-based fall detection systems have been developed, and there are various data sets available, but the accuracy and precision have not been standardized or compared; even where comparison does exist, it has been run on GPUs. No analysis of the workability of the models and the data sets on SoCs has been previously attempted. Though over the last few years, ML and DL algorithms have been increasingly used in fall detection, and there have also been some suggestions for the use of compressed modelling, there are no concrete statistics available to form this conclusion. In this paper, we attempt to understand why ML algorithms cannot run as-is on existing SoCs; We are using Snapdragon 410c to do our analytics as it is primarily used in Biomedical and IoT applications, has low power consumption and small form factor making it ideal for wearables. In this paper, we have used KNN to prove that ML cannot be used directly on SoCs. We are using KNN as it does not have any pre-training period, and it is a very simple algorithm that gives good accuracy. In this paper, we establish the need for model and data compression for fall detection if we must use ML or DL algorithms on SOCs. We have done this with statistical analysis across different data sets.