Motion Prediction Using LSTM Network in a Vehicular Environment

Autonomous driving has been rapidly increasing in demand and interest over the past several years. It is critical to understand if a pedestrian/cyclist is about to cross the road or if a car is in the middle of parallel parking or making a right turn. There are hundreds of possible actions a road user can take and it is imperative to accurately predict the behaviour of other road users which makes this a challenging problem in autonomous driving. Correctly predicting and evaluating other road users' movements is essential to being able to reduce and avoid crashes thus improving the safety of autonomous vehicles. This MRP solves a part of the Waymo motion prediction challenge using the Waymo motion dataset. The problem is to predict 8 seconds into the future given 1 second of history. The work in this MRP uses (LSTM) Long Short Term Memory networks which is a special type of (RNN) Recurrent Neural Network, which have been widely used for time series prediction such as stock price prediction and electric load forecasting. This problem has been approached by many researchers using different machine learning algorithms such as (CNN) Convolutional Neural Network, LSTM and encoders/decoders. We found that LSTM networks can outperform several state-of-the-art approaches without the use of all available features. Surprisingly, we found that only using the agents' location, speed and angle was sufficient to get state-of-the-art results with a simple LSTM model and data preprocessing. The results obtained from the proposed solution outperformed current LSTM-based methods. The proposed method is implemented in python using the Tensorflow/Keras framework and is available in GitHub https://github.com/wfrei020/motion-prediction-v1. The MRP presents results from eight different road scenarios.