Lane-Attention: Predicting Vehicles’ Moving Trajectories by Learning Their Attention Over Lanes

Abstract

Accurately forecasting the future movements of surrounding vehicles is essential for safe and efficient operations of autonomous driving cars. This task is difficult because a vehicle's moving trajectory is greatly determined by its driver's intention, which is often hard to estimate. By leveraging attention mechanisms along with long short-term memory (LSTM) networks, this work learns the relation between a driver's intention and the vehicle's changing positions relative to road infrastructures, and uses it to guide the prediction. Different from other state-of-the-art solutions, our work treats the on-road lanes as non-Euclidean structures, unfolds the vehicle's moving history to form a spatio-temporal graph, and uses methods from Graph Neural Networks to solve the problem. Not only is our approach a pioneering attempt in using non-Euclidean methods to process static environmental features around a predicted object, our model also outperforms other state-of-the-art models in several metrics. The practicability and interpretability analysis of the model shows great potential for large-scale deployment in various autonomous driving systems in addition to our own.

StarNet: Joint Action-Space Prediction with Star Graphs and Implicit Global Frame Self-Attention

“…Consequently, they are able to capture a longer range longitudinally, as well as account for the different semantic meaning of lateral lanes. Another class of works is [29] and [19], which model attention to specific lanes or sections along a reference polyline, constructed by concatenating individual polyline points. In the case of [19], this enables placement of hypothetical goals along the polyline to condition the trajectory prediction.…”

“…We task a Graph Attention Network (GAT) [31] with the answer; the attention mechanism learns to determine the most relevant vectors without artificially limiting the receptive field. Furthermore, our map representation is simpler to pre-process than [10], [28] since we use standard graph convolutions, as well as [19], [29], since we do not manually select the reference polyline and allow other map element types to be attended to as well.…”

LaPred: Lane-Aware Prediction of Multi-Modal Future Trajectories of Dynamic Agents

“…However, since trajectory data tend to be represented by the sequence of points in spatial coordinates, it is not trivial to reason the relationship between the trajectories and scene context (e.g., lanes) drawn on a 2D image. Hence, in [9,14,17,18,20,33], the instance-level representation of the scene context is used and described in the same coordinate domain as the trajectory. Vectornet [9], WIMP [14], LaneGCN [17], LA [20], and TNT [33] obtain the scene context from graph structures constructed from the instance-level semantic map information.…”

“…Hence, in [9,14,17,18,20,33], the instance-level representation of the scene context is used and described in the same coordinate domain as the trajectory. Vectornet [9], WIMP [14], LaneGCN [17], LA [20], and TNT [33] obtain the scene context from graph structures constructed from the instance-level semantic map information. MTPLA [18] estimates the most correlated lane using instance-level lane information.…”

Jointly Learning Agent and Lane Information for Multimodal Trajectory Prediction

“…Typically, the regression-based [3], [15] and fixed anchor-based methods [5], [6] either concentrate on the main mode or restricted for a certain scene. The existing lane-based methods [7], [8], [9], [10] either consider a single lane or fuse all lanes. They do not explicitly take lanes as goals and are hard to generate map-adaptive trajectories.…”