Relational Recurrent Neural Networks For Vehicle Trajectory Prediction

Abstract

Scene understanding and future motion prediction of surrounding vehicles are crucial to achieve safe and reliable decision-making and motion planning for autonomous driving in a highway environment. This is a challenging task considering the correlation between the drivers behaviors. Knowing the performance of Long Short Term Memories (LSTMs) in sequence modeling and the power of attention mechanism to capture long range dependencies, we bring relational recurrent neural networks (RRNNs) to tackle the vehicle motion prediction problem. We propose an RRNNs based encoder-decoder architecture where the encoder analyzes the patterns underlying in the past trajectories and the decoder generates the future trajectory sequence. The originality of this network is that it combines the advantages of the LSTM blocks in representing the temporal evolution of trajectories and the attention mechanism to model the relative interactions between vehicles. This paper compares the proposed approach with the LSTM encoder decoder using the new large scaled naturalistic driving highD dataset. The proposed method outperforms LSTM encoder decoder in terms of RMSE values of the predicted trajectories. It outputs an estimate of future trajectories over 5s time horizon for longitudinal and lateral prediction RMSE of about 3.34m and 0.48m, respectively.

Attention Based Vehicle Trajectory Prediction

“…Lenz et al [3] use as input to their model only the current state of a set of neighboring vehicles in order to achieve the Markov Property. Other existing studies [4], [5], [6], [7], [1], [8] use a sequence of past features to benefit from extra temporal information in the prediction task. Environment features.…”

“…3) Prediction: Vehicle intent prediction is divided into two main aspects: maneuver [4], [23] and trajectory prediction [5], [24], [8]. The former generates a high-level representation of the motion such as lane changing and lane keeping.…”

“…Different forms of outputs are used in the motion prediction task. In [5], [8], the exact future positions are predicted. Others [7], [25], [6] deploy a multimodal solution using Gaussian mixture models over predicted states.…”

“…Multi-head attention mechanism was introduced by Vaswani et al [34] for natural language processing purposes. A relational recurrent network based on attention mechanism was deployed in [8] for trajectory prediction. In [1], an attention-based nonlocal vehicle dependencies model that represents vehicles' interactions based on their importance to the target vehicle is introduced.…”

Trajectory prediction for intelligent vehicles using spatial‐attention mechanism

“…The encoder–decoder network based on LSTM and beam‐search method is proposed on [23], the model is evaluated on the data sets of aerial trajectory photos in Seoul. In [24], two methods are proposed to capture the spatial–temporal dependencies between the input tracks and test it on the highD data set, these two methods are RRNN with scene embedding (Sc‐RRNN) and RRNN with per lane embedding (L‐RRNN).…”