Human trajectory predictor for indoor mobile robot applications

The present dissertation aims at implementing a C++ real time human trajectory predictor for autonomous mobile robots in indoor unstructured environments. The main objectives of the prediction system are the optimization of the robot path planning process and the improvement of the robot ability to safely coexist with the surrounding human beings. To make the robot aware of the future behaviour of the people, the proposed software package first generates a prediction taking into account the human-human, the human-scene and the human-robot interactions and then makes the obtained information available to the robot planner by inserting proper cost areas into those locations of the navigation costmap that are believed to be crossed by the pedestrians over the next 2.4 s; thanks to these additional costs, the robot is induced to avoid the crossed areas, and no trajectory superposition takes place. The forecasting algorithm generates the predictions exploiting three main types of input: the current and past pedestrians’ states (i.e. positions and velocities), the environment static map and the current robot planned path. Given the inputs, the system produces an energy cost function from the minimization of which it obtains the pedestrians’ future trajectories. Each trajectory is expressed as a set of 6 discrete bidimensional points, where each point is temporally spaced from the previous one and from the next one by a Δt of 0.4 s. Beside the prediction block, two supporting software tools have been developed: a human tracker, that identifies all the pedestrians in a stereo camera field of view extracting their positions and velocities, and a prediction costmap layer, that acquires the forecasted bidimensional points from the predictor, converts them into the corresponding cost areas and integrates these costs into the navigation costmap. The overall system, designed as a ROS 2 package, has been initially evaluated in two different environments in the Gazebo simulator and then on a real robot. The tests have led to both quantitative and qualitative considerations that have allowed to assess the numerical precision of the predictions and the positive effects of such predictions on the robot navigation.