Vehicle speed prediction with LSTM from trajectories

Abstract

This project was undertaken with the objective of predicting a vehicle’s velocity using Trajectory data on an LSTM. The average velocity of a vehicle was calculated based on distance covered in time, and this velocity was interpolated into a continuous form, to predict continuous data.

This project uses a Neural Networks based model, which takes in continuous data on a vehicle’s velocity and outputs its next velocity within every 10 second interval. This was implemented using TensorFlow, Python, and a Stacked LSTM (Long Short Term Memory) model. The data was from a Navigation system dataset. The time span is July 3-9, 2017, which contains five workdays (Jul 03-07) and two weekends (Jul 08-09).

The pre-processing of the dataset involved cleaning the data, then restructuring it into a continuous format with 1 second intervals. Following this, the data was split into a 9 input and 1 output format.

The data was then clustered using the K-Means method, and this was used to train and test the model. This paper will focus on the design and implementation of the data, the model, as well as the challenges encountered during this process.