Pedestrian Trajectory Prediction with Pose Representation and Latent Space Variables

Pedestrian trajectory is challenging to predict due to its nonlinear nature, and uncertainty about pedestrian movement intentions. At the same time, it is an important information to ensure safety in an environment shared by humans and intelligent vehicles. Many algorithms focus on pedestrian detection to predict its position in a nearby future, combined with external features such as head orientation, visual context from scenarios, social iteration, vehicles information, etc. Nevertheless, the features are usually generated by a human annotator, an expensive task that does not necessarily represent the data quality obtained by a real on-board vision system. Moreover, the use of some variables, such as sequence of images, request more computational processing. We propose a model for predicting pedestrian’s future trajectories, which uses the pedestrian’s key points that represent its pose and position. These key points are used as inputs in a recurrent neural network. We also explore the advantages of using auto extracted features from the latent space of a recurrent autoencoder. Our final approach decreases the mean squared error of the future trajectories of pedestrians and it does not significantly increase computational processing. In addition, it automatically extracts the features without needing a human annotator.