Pedestrian Path Prediction based on Body Language and Action Classification

Safety-related driver assistance systems are becoming mainstream and nowadays many automobile manufacturers include them as standard equipment. For example, pedestrian protection systems are already available in a number of commercial vehicles. However, there is still work to do in the improvement of the accuracy of these systems since the difference between an effective and a non-effective intervention can depend on a few centimeters or on a fraction of a second. In this paper, we use the 3D pedestrian body language in order to perform accurate pedestrian path prediction by means of action classification. To carry out the prediction, we propose the use of GPDM (Gaussian Process Dynamical Models) that reduces the high dimensionality of the input vector in the 3D pose space and learns the pedestrian dynamics in a latent space. Instead of combining a reduced number of subjects in a single model that will have to deal with the stylistic variations, we propose a much more scalable approach where all the subjects are separately trained in individual models. These models will be then hierarchically separated according to their action (walking, starting, standing, stopping) and direction of the motion. Finally, for a test sequence, the appropiate model will be selected by means of an action classification system based on the similarity of the 3D poses transitions and the joints velocities. The estimated action will constrain the models to use for the prediction, taking into account only the ones trained for that action. Experimental results show that the system has the potential to provide accurate path predictions with mean errors of 7 cm, for walking trajectories, 20 cm, for stopping trajectories and 14 cm for starting trajectories, at a time horizon of 1 s.