A Dynamic and Static Context-Aware Attention Network for Trajectory Prediction

Abstract

Forecasting the motion of surrounding vehicles is necessary for an autonomous driving system applied in complex traffic. Trajectory prediction helps vehicles make more sensible decisions, which provides vehicles with foresight. However, traditional models consider the trajectory prediction as a simple sequence prediction task. The ignorance of inter-vehicle interaction and environment influence degrades these models in real-world datasets. To address this issue, we propose a novel Dynamic and Static Context-aware Attention Network named DSCAN in this paper. The DSCAN utilizes an attention mechanism to dynamically decide which surrounding vehicles are more important at the moment. We also equip the DSCAN with a constraint network to consider the static environment information. We conducted a series of experiments on a real-world dataset, and the experimental results demonstrated the effectiveness of our model. Moreover, the present study suggests that the attention mechanism and static constraints enhance the prediction results.

MALS-Net: A Multi-Head Attention-Based LSTM Sequence-to-Sequence Network for Socio-Temporal Interaction Modelling and Trajectory Prediction

“…DSCAN [ 46 ]: This method uses a constraint network and models attention between vehicles to extract the weights to make future predictions.…”

A Review of Deep Learning-Based Vehicle Motion Prediction for Autonomous Driving

“…[31,76,79] first incorporate the map information into each vehicle by applying cross-attention and then use the attention operation among the vehicles to obtain interaction features. Yu et al [77] first learn spatial proximity information based on the encoding of the grid map, and then the influence of different grid cells on the future motion of the target vehicle is obtained by using the attention mechanism. To fully extract the interaction information, some works [32,73] use the multi-headed attention mechanism to consider the interaction in different dimensions.…”

“…[76] 2021 1D-CNN, GRU, AM [79] 2022 1D-CNN, GRU, GCN, AM [31] 2021 LSTM, AM [32] 2021 LSTM, AM [73] 2020 LSTM, AM [77] 2021 LSTM, CNN, AM [60] 2022 LSTM, MLP [80] 2022 Transformer, MLP [81] 2023 GAN, GCN, FPN, AM Encode the correlation features between each surrounding vehicle and the target vehicle based on a fully connected layer. [46] 2020 GNN, AM, MLP Construct a fully connected undirected graph, and then extract the interaction feature based on graph attention network.…”

“…Considering environmental elements such as roads and weather conditions, Yu et al [77] design a constraint net to model environmental constraints. The authors embed the discrete environment elements into dense continuous vectors with the same dimension, which are then concatenated together to form the environment vector.…”

A Review on Intention-aware and Interaction-aware Trajectory Prediction for Autonomous Vehicles

“…Kim, Kum and Choi [160] presented a recursive prediction framework, where the predicted trajectory is also used as input feature in an LSTM-based encoder-decoder architecture. Similar to other approaches, the attention mechanism appears between the encoder and decoder components to selectively highlight specific features of the context vector before the decoder takes it as input [101], [161]. However, an attention mechanism can also extract features from specific contexts, such as in Zhang et al [162] that proposed a module for social features only, which extract spacial and temporal interaction features that are concatenated with other context vectors from other encoders before being fed to the decoder.…”