Trajectory prediction for intelligent vehicles using spatial‐attention mechanism

Abstract

It is of great interest for autonomous vehicles to predict the trajectory of other vehicles when planning a safe trajectory. To accurately predict the trajectory of the target vehicle, the interaction between vehicles must be considered. Interaction aware prediction methods track the previous trajectories of both the target vehicle and its surrounding vehicles. In this study, the authors consider trajectory prediction as a sequence‐to‐sequence prediction problem. They tackle this problem with an LSTM encoder–decoder framework. Moreover, they propose two spatial‐attention mechanisms to account for the interaction between vehicles, i.e. context attention and lane attention. Spatial‐attention mechanisms adopt the selective‐attention mechanism of human drivers. They choose context vectors to help the model understand the surrounding environment better and thus improve its prediction accuracy. They evaluate the authors’ methods on the highD data set recorded in German highways with root mean squared error metric. Their experimental results show superior performance to other state‐of‐the‐art methods. Code is available at https://github.com/momo1986/Spatial-attention.

Trajectory prediction for autonomous driving based on multiscale spatial‐temporal graph

“…Thus, we further aggregated a traffic scene within 100 m (±50 m) where only vehicles within this region would be processed together, and other vehicles would be discarded. highD datasets: highD is a large‐scale naturalistic vehicle trajectory dataset from German highways captured at 25 Hz, consisting of 16.5 h of measurements from six locations with 110,000 vehicles, a total driven distance of 45,000 km and 5600 recorded complete lane changes. Like [48], we set the data sampling rate to 5 Hz and extracted total 178,539 driving sequences. We split the trajectories in highD into segments of 8 s, using 3 s of track history to predict the trajectories in the next 5 s. ApolloScape trajectory dataset: ApolloScape is a trajectory prediction dataset for urban streets during rush hours [14], which uses the data from the first 3 s to predict trajectories for the next 3 s. Compared to NGSIM, the most significant difference of ApolloScape is that it was collected for the trajectory prediction of heterogeneous traffic agents that contained highly complicated traffic flows mixed with vehicles, riders, and pedestrians.…”

“…highD datasets: highD is a large‐scale naturalistic vehicle trajectory dataset from German highways captured at 25 Hz, consisting of 16.5 h of measurements from six locations with 110,000 vehicles, a total driven distance of 45,000 km and 5600 recorded complete lane changes. Like [48], we set the data sampling rate to 5 Hz and extracted total 178,539 driving sequences. We split the trajectories in highD into segments of 8 s, using 3 s of track history to predict the trajectories in the next 5 s.…”

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

“…This approach shares an encoder module for both intention estimation and trajectory prediction, which takes only historical trajectory as input of the model. Yan et al [154] explored an architecture with two different types of self-attention mechanisms, one for the driving context and one for the driving lane. The attention mechanism for the driving context is responsible for differentiating the importance of each vehicle present in the scenario.…”

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

“…There are many applications of attention for trajectory prediction. For instance, it can learn the most relevant temporal and spatial features for prediction [134], [154], [155], or differentiate the contribution of each surrounding vehicle for the target vehicle trajectory [24], [51], [156]. In intention-aware frameworks, it can also differentiate the importance of features for each maneuver intention.…”