[Paper Review] Maneuver-Aware Pooling for Vehicle Trajectory Prediction

Mohamed Hasan, Albert Solernou+4 more|arXiv (Cornell University)|Apr 29, 2021

Autonomous Vehicle Technology and Safety18 references17 citations

TL;DR

This paper proposes a maneuver-aware pooling mechanism for vehicle trajectory prediction that improves accuracy during lane changes and highway merging by incorporating polar trajectory representation, vehicle orientation, and radial velocity. The method enhances interaction modeling in an encoder-decoder generative model, outperforming state-of-the-art pooling strategies—especially in maneuver-specific evaluations—on the NGSIM dataset with a publicly available codebase.

ABSTRACT

Autonomous vehicles should be able to predict the future states of its environment and respond appropriately. Specifically, predicting the behavior of surrounding human drivers is vital for such platforms to share the same road with humans. Behavior of each of the surrounding vehicles is governed by the motion of its neighbor vehicles. This paper focuses on predicting the behavior of the surrounding vehicles of an autonomous vehicle on highways. We are motivated by improving the prediction accuracy when a surrounding vehicle performs lane change and highway merging maneuvers. We propose a novel pooling strategy to capture the inter-dependencies between the neighbor vehicles. Depending solely on Euclidean trajectory representation, the existing pooling strategies do not model the context information of the maneuvers intended by a surrounding vehicle. In contrast, our pooling mechanism employs polar trajectory representation, vehicles orientation and radial velocity. This results in an implicitly maneuver-aware pooling operation. We incorporated the proposed pooling mechanism into a generative encoder-decoder model, and evaluated our method on the public NGSIM dataset. The results of maneuver-based trajectory predictions demonstrate the effectiveness of the proposed method compared with the state-of-the-art approaches. Our "Pooling Toolbox" code is available at https://github.com/m-hasan-n/pooling.

Motivation & Objective

• To improve trajectory prediction accuracy for autonomous vehicles during complex maneuvers like lane changes and highway merging.
• To address the limitation of existing pooling methods that rely solely on Euclidean trajectory representations and fail to capture maneuver intent.
• To model inter-vehicle dependencies more effectively by incorporating higher-order dynamics such as vehicle orientation and radial velocity.
• To develop a pooling mechanism that implicitly encodes maneuver awareness without requiring explicit maneuver labeling during training.
• To evaluate the method not only overall but also on specific maneuvers, revealing performance gains in critical driving scenarios.

Proposed method

• The proposed pooling mechanism uses polar coordinates (distance r and angle φ) instead of Cartesian positions to represent relative vehicle trajectories.
• Vehicle orientation is incorporated to enhance maneuver awareness, enabling better modeling of directional intent during lane changes and merges.
• Radial velocity (Vr) is added to the representation to capture motion trends toward or away from the ego vehicle, improving dynamic interaction modeling.
• The pooling operation concatenates LSTM-encoded states of neighboring vehicles with their polar trajectory features, forming a maneuver-aware context vector.
• The method is integrated into a generative encoder-decoder model that outputs multi-modal trajectory distributions conditioned on semantic maneuvers.
• A maneuver recognition module is used in the full model to predict maneuver classes (e.g., merge, left lane-change) based on LSTM-encoded vehicle dynamics and radial velocity.

Experimental results

Research questions

• RQ1 Can incorporating vehicle orientation and radial velocity into the pooling mechanism improve trajectory prediction accuracy during lane changes and highway merging?
• RQ2 How does polar representation of relative trajectories compare to Euclidean representation in modeling inter-vehicle interactions?
• RQ3 Does a pooling strategy that implicitly encodes maneuver intent lead to better performance than conventional pooling methods in both overall and maneuver-specific evaluations?
• RQ4 To what extent does the inclusion of radial velocity enhance the model's ability to predict high-risk maneuvers like lane changes?
• RQ5 Can a maneuver-aware pooling mechanism improve learning efficiency without requiring explicit maneuver annotations during training?

Key findings

• The proposed Polar- Vr pooling strategy significantly outperforms S-LSTM, CSP, and S-GAN baselines in both overall and maneuver-based trajectory prediction accuracy.
• The use of polar coordinates (Polar) alone already improves performance over Euclidean baselines, especially for lane change and merging maneuvers.
• Incorporating radial velocity (Polar- Vr) further enhances prediction accuracy, demonstrating the importance of higher-order dynamics in interaction modeling.
• The maneuver recognition module did not improve RMSE when applied to baseline models, indicating that the key improvement comes from the pooling mechanism itself rather than maneuver classification.
• The model achieves superior results in maneuver-specific evaluations—particularly for left and right lane changes and merging—despite the data imbalance favoring keep-lane behavior.
• The results confirm that modeling vehicle orientation and radial velocity leads to a more efficient learning process and better generalization in complex traffic scenarios.