Abstract:
Trajectory prediction of road users is a very important task in autonomous vehicles for road safety. It becomes even more important when road users are pedestrians being more vulnerable. Recently, machine learning based approaches have
been used for trajectory prediction of pedestrians including deep learning using
deep neural networks. It is a question that what are the best features to include
in order to predict the future trajectory of pedestrians more accurately. Moreover, the recurrent neural networks (RNNs) based decoder in enocoder-decoder
architecture suffer from the accumulated error that degrades long term predictions. In this work, feature selection for the trajectory prediction of pedestrians
has been done and a novel bi-directional RNN decoder has been proposed. JAAD
and PIE datasets have been used which focus on pedestrians. Different groups
of available features in the datasets were selected, and the exploration analysis
was done. Furthermore, in order to address the issue of accumulated error in unidirectional RNN decoders, a bi-directional decoder which has a forward RNN and
a backward RNN, was proposed and compared with the uni-directional decoder.
Results show that attributes features when added with spatial features improve
the accuracy of the trajectory prediction and the proposed bi-directional decoder
improves the accuracy of long term trajectory prediction.
