Radar-Based Fall Event Detection Using Histogram of Oriented Gradients of Binary-Encoded Radar Signatures

Existing deep learning techniques for radar-based fall event detection are computationally expensive and data hungry. This work proposes a novel methodology for radar-based fall event detection using computationally inexpensive traditional machine learning algorithms with the histogram of oriented gradients (HOG) of binary-encoded radar signatures. Radar signatures (binary) encoded using nonparametric local transforms, such as census transform (CT) and local binary pattern (LBP), contain enhanced gradient information and lead to distinguishable HOG features for better fall event detection. A comparative analysis of five different types of machine learning algorithms using HOG features from three different types of radar domain representations that are binary encoded in two different ways (CT and LBP) is presented. All evaluations are done on a publicly available dataset of human radar signatures. K-nearest neighbor algorithm using HOG features of binary-encoded spectrograms achieves an F1-score of 0.995.