Pose classification for assistive unmanned vehicles with deep learning at the edge

Technology in the last decades has revolutionized our way of life, making more and more powerful devices available to everyone. The development of such technologies has been exploited also to conduct researches in several scientific fields, i.e. medical area, biotechnologies, computer science etc. Nowadays, thanks to the improvement made in Robotics and Artificial Intelligence, there can be found virtual assistance devices in many households and public places. Service robotics and artificial intelligence allows to help elder and disabled people, on different daily and basic tasks, in a noninvasive manner. The idea of this thesis was born at PIC4SeR ( PoliTO Interdepartmental center of service robotics ) with the purpose of implementing a complex service robotics application to support elder and disabled people in the house environment. The goal of this project is the developing of an algorithm able to detect the pose of a person and recognize the static action done. The actions that can be recognized, in this first evolution of the work, are limited to three: standing, sitting and laying. Different data sets have been created to build several neural network architectures. These architectures have been tested to find the best trade-off between accuracy and computational cost to satisfy a real-time result, which is one of the main requirements of the system. The data sets have been created using different cameras, the pose estimation and action prediction are done using a stereo-camera, and an edge TPU coprocessor. At the end of the experimentation, good results have been achieved giving an accurate prediction of the person's pose in the house environment.