Patch-based Learning for Radar-based Fall Event Detection Using Gramian Angular Field

This paper proposes a Gramian Angular Field (GAF) for detection of fall events using radars. These GAF representations are directly obtained using range-time information and utilize temporal dependency among time samples for enhanced fall event detection. This work compares the performance of a patch-based learning model with attention (improved vision transformer (I-ViT)) and two patch-based learning models without attention (Multi-Layer Perceptron-Mixer (M-mixer) and Convolutional-Mixer (C-Mixer)) for fall detection with GAF plots as input with those obtained with range-time plots as input. Models are evaluated using a publicly available dataset with location-wise testing strategy. Patch-based learning model using GAF plots as input perform significantly better as compared to patch-based learning models using range-time plots at two out of three testing locations.