A trustworthy AI system must be well perceived by the humans interacting with it. One of the REXASI-PRO components that is more impacted by this requirement is the wheelchair. In fact, having an autonomous wheelchair that drives in the crowd without hitting obstacles and harming human beings is not enough to consider it trustworthy. The system must also drive in a way that is considered well perceived or at least acceptable both, by the person using the wheelchair and by the person surrounding it.
In order to achieve this result, the neural network that controls the wheelchair motor will be trained to compute trajectories without abrupt changes both in terms of acceleration and direction. Moreover, the computed trajectory will also be influenced by the presence and trajectories of the surrounding people. A key component to give the network this capability is to design a robust people tracker to recognize the trajectories of the people in front of the wheelchair.
In the attached video you can see the first version of a people tracker that using multiple sensors (2 RBGD cameras and a 2D LIDAR) recognizes the presence of a person and project its 2D position on the ground plane. The information about the skeleton key points of the tracked person is also valuable information that can be used for example to know where the sight of the person is directed, an information that the neural network or the safety layer of the wheelchair can use to know if the person has seen the wheelchair and to estimate its future walking direction.
