Vehicle trajectory prediction for safe navigation of autonomous vehicles

Abstract

Trajectory prediction of the other road users in the vicinity of an autonomous vehicle is important for safe navigation in dense traffic. Once an autonomous vehicle anticipates how the other road actors will react in the near future, path planning is a lot more simpler and safer. Moreover, the knowledge of future movement of other road actors allows control of sudden jerks in the planned ego vehicle’s path and thus makes travel smoother. This trajectory prediction stage can be used at any level, from restricted driver assistance to full vehicle autonomy. In this thesis two novel trajectory prediction models have been developed. In the first model, the spatio-temporal features that form the basis of behaviour prediction were captured using a Convolutional Long Short Term Memory (Conv-LSTM) neural network architecture consisting of three modules: 1) Interaction Learning to capture the motion of and interaction with surrounding cars, 2) Temporal Learning to identify the dependency on past movements and 3) Motion Learning to convert the extracted features from these two modules into future positions. In addition, a novel feedback scheme was introduced in which the current predicted positions of each car are leveraged to update future motion, encapsulating the effect of the surrounding cars. In the second model a conventional Long Short Term Memory (LSTM) cell based encoder-decoder architecture was developed which uses not only the historical observations but also the associated map features. Moreover, unlike existing architectures, the proposed method incorporates and updates the surrounding vehicle information in both the encoder and decoder, making use of dynamically predicted new data for accurate prediction in longer time horizons. This seamlessly performs four tasks: first, it encodes a feature given the past observations, second, it estimates future maneuvers given the encoded state, third, it predicts the future motion given the estimated maneuvers and the initially encoded states, and fourth, it estimates future trajectory given the encoded state and the predicted maneuvers and motions. Both the developed models were evaluated extensively on two publicly available datasets which include both multi-lane highway and signalled intersections, to benchmark the prediction accuracy with the state-of-the-art models. Later, the conventional encoder-decoder model was also evaluated with a newly collected “Radiate” dataset which includes two intersections, the Kingussie T-junction and the Edinburgh four-way junction, both without traffic signals. The accuracy of the predicted trajectories on the benchmark datasets are comparable with state-of-the-art methods. Moreover, evaluation on the latter dataset (“Radiate”) made it possible to understand better the effect of inter-vehicle interactions on future motion without any influence from mandatory traffic signals.