Knowledge Transfer for Human Trajectory Prediction

Human trajectory prediction task aims to analyze human future movements given their past status, which is an important topic in several application domains such as socially-aware robots, intelligent tracking systems and self-driving cars. The goal of this work is to transfer knowledge from already seen scenes from the training dataset to the test set. In order to do that we create a module that, for each human in the scene, extracts local features from a patch around the agent’s current position. To test its ability the proposed model was placed next to both the SAR and the Goal SAR architectures. To this end, the resulting feature descriptors coming from our approach are concatenated to the lightweight attention-based recurrent backbone that acts solely on past observed positions for both the aforementioned architectures. We conducted extensive experiments on training the model on the SDD dataset and tested it in the ETH dataset to show our approach performances compared to the baseline.