Prediction of Vehicles’ Movement based on Kalman-filter and LSTM Algorithms

Abstract

The development of autonomous vehicles is advancing at a very high pace and could radically transform the future of the transportation system. If autonomous vehicles have to start driving safely and efficiently on public roads, they would have to know how the surrounding vehicles are behaving and act accordingly. This ability to predict and adapt is essential to make autonomous vehicles safe to use. In this thesis, the problem is considered as a sequence-to-sequence trajectory prediction problem. The proposed solution is a combination of a Kalman filter with an LSTM encoderdecoder framework. An attention mechanism is introduced to the LSTM framework for the interaction between the vehicles i.e. context attention and lane attention. Context vectors are chosen which helps in improving the accuracy. The model is trained and validated on the HighD Dataset and evaluated with the root mean squared error metric. The experiment shows good improvement in the accuracy when the attention mechanism is applied to the LSTM framework.