Design and evaluation of neural network models for radar-based human-gait recognition in an open-set environment

In this era of technological innovation, a growing number of areas are implementing applications based on Machine Learning (ML) models and Neural Network (NN) architectures for human sensing using Radio Frequency (RF) signals. Among these, person identification using biometric gait features extracted from the reflected signals of millimeter-wave radar devices are attracting a lot of interest. The appeal of this approach lies in its contactless nature, that does not require placing sensors directly on the subjects (e.g., wearables), and in their capability to function under any lighting or weather condition. In this project, a ML model for human-gait recognition is presented, using for its training a dataset consisting of a series of point clouds describing the movement patterns of 16 subjects collected by a millimeter-wave radar. The design is based on an adversarial autoencoder with a semi-supervised learning paradigm, and innovative techniques such as softpooling have been applied with the aim of achieving an optimal handling of the data features. In addition, we have extended the scope of the project to a booming field that is lacking solutions, developing the model in an open-set environment. In this way we have not only evaluated the behaviour of the system in terms of classifying known subjects during the training process, but also analysed its performance against unknown subjects that have not been introduced to the model until the testing phase. Despite the challenging task of tackling the open-set problem and the implementation of softpooling in an unprecedented environment, this project achieves promising results in unknown subject recognition and point cloud reconstruction tasks, providing inspiration for future lines of research.