Motorcyclists are at high risk in road traffic collisions, however motorcycle usage is growing in popularity thanks to their affordability and their peculiarity to easily move in traffic congestions. The major cause of death and disability to motorcyclists is Traumatic Brain Injuries (TBIs), and the only possibility to protect the head is wearing a certified helmet. Despite the helmet usage, the head can be subjected to a serious impact in road crashes, and early diagnosis and treatment of potential TBIs is the only way to save riders’ lives. Recently, the European Community developed an automatic emergency call system (eCall) to enable a rapid and focused treatment in the event of a road crash. However, the adoption of such a device on Powered Two Wheelers (PTWs) is hampered by difficulties in defining exactly when an emergency call needs to be made.
The goal of this study was to develop a methodology for a real time estimation of TBIs in road crashes involving motorcyclists. TBI prediction can be used as activation criterion for the eCall enabling this system for PTWs. The methodology developed in this work consists of 5 single-axis accelerometers attached to the inner surface of the helmet outer shell and an Artificial Intelligence (AI) module, which consists of two sub-modules: the kinematics sub-module and the injury sub-module.
The first part of this study involved the development of the kinematics submodule. The aim of this sub-module was to estimate the head linear and rotational accelerations. It is based on Long Short Term Memory (LSTM) Artificial Neural Networks (ANNs) fed with the accelerations acquired by the accelerometers embedded into the helmet. ANNs were trained using numerical data obtained reproducing in a virtual environment several helmeted head impact simulations. The model accuracy was investigated with experimental tests reproducing oblique impacts carried out in collaboration with AGV company and the Imperial College of London. Unlike past technologies, the methodology described in this thesis provides an estimation of the head acceleration time patterns without the need of a direct contact between sensors and head, ensuring the possibility of a practical application in the motorcycling field.
The second part of this study focused on the development of the injury submodule. Its aim was to predict the TBI risk in terms of brain strain and strain rate in 12 specific Regions of Interest (ROI). It is based on Convolutional Neural Networks (CNNs) fed with the components of the linear and rotational head acceleration outputted by the kinematics sub-module. CNNs were trained with numerical data extracted from a finite element model of brain. Experimental oblique impact tests were used to assess the model performances. Results showed that, for specific applications, a detailed 3D computational model of brain used to estimate TBIs can be replaced by ANNs, addressing the main limitations of the computational models: high computation time and complex computational software. In addition, the results proved that ANNs can be trained using only numerical data and then used with real-world data, drastically reducing the amount of experimental tests needed to develop such complex models.