Residual Learning for Mixed Traffic Prediction in Shared Space

In  recent  years,  with  the  increased availability of computational power and large-scale datasets, data--driving approaches, especially Deep Learning approaches, have  been  largely  used  for  trajectory modeling. Nevertheless, predicting mixed traffic trajectories in shared space is not trivial. The decision of an agent (road user) for the next step not only depends on its own motion pattern (past trajectory and destination), but is also impacted by other agents in the vicinity (interaction).  Helbing et al. call these factors repulsive and attractive effects. In order to capture these effects, Social LSTM (Alahi et al., 2016) and DESIRE (Lee et al., 2017) use a pooling layer to parse the interactions between the target agent and other neighborhood agents. However, the pooling layer is only based on the existence of the neighborhood agents within in a predefined interactive zone. According to the pooling mechanism, all the neighborhood agents within the zone give the same impact to the target agent. This might not be the case in reality, especially when those neighborhood agents interact with the target agent with different speed (acceleration), distance, and position. On the other and, they either predict the distribution of the coordinates or directly predict the coordinates of a position at each time step, which is more difficult than predicting the displacement (residual) between two consecutive steps. To this end, Xu et al. propose to use an LSTM-based model for predicting the relative positions at each time step for a trajectory. Because a relative position can be regarded as speed. The LSTM-based model can extract speed and acceleration information directly from relative positions. This approach achieves better performance with less model complexity for pedestrian trajectory prediction (Xu et al., 2018).

The objective of this master thesis is to implement this model and extend it for mixed trajectories---not only pedestrians, but also cyclists and vehicles. The model will be trained using real--world data extracted from shared spaces with different space layout. Its performance will be evaluated by measuring the displacement errors between the predicted trajectories and true trajectories regarding agent's transport mode.