Evaluation and analysis of open-set radar-based human gait recognition performance with an adapted radar dataset.

Radar-based Human Gait Recognition is an innovative technology that identifies users, analysing their walking patterns. This recent technology has a wide variety of applications in different fields, such as security and surveillance. The main advantage of this recognition method is that it does not require the cooperation of users or physical interaction with them, thus making it a non-intrusive method for identifying them in real-life situations. In this thesis, we have analysed and compared the behaviour of a deep learning model in an open-set environment, that recognises human gait by processing sparse point clouds sequences, which are captured by an mmWave radar sensor. These sequences contain information of users walking in different conditions from two different datasets. This semi-supervised model, which is an adversarial autoencoder, proposes a new way of processing point clouds by ordering point cloud features according to their activations, following the guidelines of an innovative technique which is called soft pooling. The main goal has been to assess and compare the performance of this model in three different tasks with both datasets. First, we explore the ability to classify users, based on extracted features from the point clouds, through detailed experiments which compare the model performance with different configurations. Besides, we assess the task of reconstructing the sparse point clouds, analysing how the model deals with the task of reconstructing them from a latent space. Finally, we have conducted experiments to see how the model faces the open-set problem, which involves the detection of unseen users. This problem is present in real-world scenarios where the model must identify users who do not belong to the training classes.