Evaluating face identification and body pose estimation models for intelligent vehicle applications

Currently, the emergence of industrial revolution in the future vehicle sector is experiencing rapid growth. In the vehicles of tomorrow, in addition to Automated Driving, importance is also given to advanced technologies that involve many other aspects, such as new biometric solutions to access the vehicle based, e.g., on recognition of the face or the body pose of the vehicle owner. Face and body pose are considered important components for future intelligent vehicle applications, such as determining whether a person is authorized to access and operate the vehicle. The challenge is to build a fast and accurate system able to detect, recognize and verify the driver’s identity with the constraint introduced operating in the wild (operating in an environment external to the vehicle, characterized by largely variable conditions). This master’s thesis work is part of a larger research, whose aim is to explore the use of face identification models and body pose estimation models for intelligent vehicle applications. The research includes the design of a framework to combine face and body pose (in particular, gait) information, which is addressed in a parallel thesis work, and the evaluation of various algorithms to be leveraged by the above framework in addition to the baseline ones, which is the main subject of this thesis work. In particular, in the context of this thesis work, a comprehensive and detailed investigation was undertaken to guide the choice of the most appropriate algorithms for extracting face and gait features. This investigation spanned a broad range of algorithms, from those that are widely recognized to those that are less known. Simultaneously, effort was devoted to exploring algorithms for the extraction of silhouettes, a critical component in gait analysis. This process entailed the examination of a variety of techniques, each with its own set of strengths and weaknesses. Following the testing of a multitude of algorithms on the CASIA-B dataset, those that exhibited the highest performance were chosen. This selection was influenced by a meticulous assessment of the individual performance of each algorithm, with consideration given to the characteristics of the dataset used. The performance was evaluated using Rank-1 accuracy and the analysis of the Area Under the Curve (AUC) and Receiver Operating Characteristic (ROC) curves. These methodologies, along with various other metrics, allowed for a comprehensive and accurate comparison of the performance of the various algorithms. The results obtained are very promising, indicating a bright future for this line of research. It is anticipated that progress in this field can lead to innovative solutions in the mobility domain. Through collaboration with the Centro Ricerche Fiat (CRF) / Stellantis, the thesis aims to advance driver recognition technology, contributing to the creation of safer and more reliable intelligent vehicles.