Light-Weight Learning Model with Patch Embeddings for Radar-based Fall Event Classification: A Multi-domain Decision Fusion Approach

With the increasing morbidity and mortality rate in older adults above 65 years of age due to accidental fall, privacy-preserving radar-based fall event detection is becoming crucial. Deep learning algorithm like vision transformers (ViT) for human fall-event detection using different radar domain representation have shown excellent fall-detection accuracy. However, such techniques are computationally very expensive and unsuitable when training datasets are small. Patch-based learning models such as Multi-Layer Perceptron-Mixer (MLP-Mixer) and Convolutional-Mixer (ConvMixer) models have been developed as alternatives to ViT. In this work, the decision outputs of light-weight ConvMixer models with different domain representations of radar returns as inputs are fused for classifying the events as fall or non-fall. This proposed approach of event classification utilizes supplementary information present in different domains for enhancing the classification accuracy. Evaluation done on publicly available dataset shows an improved performance of the multi-domain ConvMixer model over ViT and MLP-Mixer. This further justifies the choice of light weight ConvMixer as a preferred learnable model when only limited training dataset is available.

I. Introduction

Fall is the leading cause of increased morbidity, disability, and mortality rate in older adults over 65 years of age worldwide. More than 33% of the elderly fall each year globally and about 10% of them experience multiple falls in a year [1]. Falls lead to hospitalization and loss of confidence to live independently. Therefore, developing remote continuous monitoring systems for fall event detection in older adults is crucial. Classifying fall events from non-fall events using remote sensors like RGB and depth camera fail to preserve privacy of individuals being monitored. Radars serve as privacy-preserving contactless sensor that can be used to detect a fall event in an independent or assisted living environment for older adults. Radar-based fall event classification may be reduced to an image classification task, for instance, by converting 1-D received radar signals into 2D-spectrograms and non-linearly transforming obtained spectrograms into RGB images. Research in this field has progressed from classification using hand-crafted features for traditional machine learning to automatic feature extraction through deep learning. Convolutional neural network (CNN), long-short term memory (LSTM) and auto-encoders are widely used deep-learning algorithms for radar-based human-fall detection [2]–[4].