Source-linked AI summary
GANet: Goal Area Network for Motion Forecasting
Mingkun Wang, Xinge Zhu, Changqian Yu, Wei Li, Yuexin Ma, Ruochun Jin, Xiaoguang Ren, Dongchun Ren, Mingxu Wang, Wenjing Yang
TL;DR
Motion forecasting is difficult because road participants’ futures are uncertain, while exact goal coordinates are hard to predict and use limited road context. GANet predicts goal areas and aggregates their lane features with GoICrop to condition trajectory forecasting. It reports state-of-the-art performance on Argoverse benchmarks, including reductions against LaneGCN on four metrics.
Problem
Exact goal coordinates are difficult to predict and evaluate, and their point representation limits the road context available to goal-conditioned motion forecasting.
Source: §I · p. 1Method
GANet predicts potential goal areas as conditions for motion forecasting and uses GoICrop to aggregate semantic lane features and model future actor interactions.
Source: §I · pp. 1–2, C. GoICrop · p. 4Results
GANet reports state-of-the-art benchmark performance; on the public challenge leaderboard it beats LaneGCN by 28% in MR6, 15% in minFDE6, 13% in brier-minFDE6, and 9% in minFDE1.
Source: B. Comparison with State-of-the-art · p. 5, §V · p. 6Takeaways & Limitations
Goal-area map features provide a data-driven way to represent map–trajectory interactions and model future interactions between actors.
Source: §V · p. 6
Abstract
from arXiv · showhide
Predicting the future motion of road participants is crucial for autonomous driving but is extremely challenging due to staggering motion uncertainty. Recently, most motion forecasting methods resort to the goal-based strategy, i.e., predicting endpoints of motion trajectories as conditions to regress the entire trajectories, so that the search space of solution can be reduced. However, accurate goal coordinates are hard to predict and evaluate. In addition, the point representation of the destination limits the utilization of a rich road context, leading to inaccurate prediction results in many cases. Goal area, i.e., the possible destination area, rather than goal coordinate, could provide a more soft constraint for searching potential trajectories by involving more tolerance and guidance. In view of this, we propose a new goal area-based framework, named Goal Area Network (GANet), for motion forecasting, which models goal areas rather than exact goal coordinates as preconditions for trajectory prediction, performing more robustly and accurately. Specifically, we propose a GoICrop (Goal Area of Interest) operator to effectively extract semantic lane features in goal areas and model actors' future interactions, which benefits a lot for future trajectory estimations. GANet ranks the 1st on the leaderboard of Argoverse Challenge among all public literature (till the paper submission), and its source codes will be released.
I. INTRODUCTION
Motion forecasting must combine map context, actor history, and interactions to predict plausible future trajectories. GANet replaces isolated goal coordinates with goal areas and uses their map features to condition forecasting.
I. INTRODUCTION
- The paper positions motion forecasting as a core challenge because road participants’ trajectories depend on HD-map context and interactions with other actors.Forecasts support subsequent planning and control in autonomous driving.Source: §I · p. 1
- Prior goal-based methods condition trajectories on limited isolated coordinates and depend on carefully designed candidate-goal spaces that may not match real-world destinations.The cited examples sample vehicle goals from lane centerlines and pedestrian goals from a virtual grid.Source: §I · p. 1
- Goal areas offer richer road context and a softer constraint than exact coordinates, while actor history helps localize them.A fast-moving vehicle’s area may be far away, whereas a stationary vehicle’s may remain around itself.Source: §I · p. 1
- GANet predicts potential goal areas as trajectory conditions and uses GoICrop to aggregate lane features and model future actor interactions.The framework encodes motion history and scene context before goal prediction and final trajectory forecasting.Source: §I · pp. 1–2
II. RELATED WORK
- GANet’s GoICrop extracts goal-area map features that capture map–trajectory interactions and provide context for modeling future actor interactions.The paper states that these features can constrain trajectories to driving rules and map topology in a data-driven manner.Source: §II · p. 2, §I · p. 2
III. METHOD
GANet encodes motion and scene context, predicts possible goals, and uses goal areas to condition multi-feasible trajectory forecasts. Its stages are presented as pluggable modules.
III. METHOD
- In the goal-based formulation described here, the candidate-goal space depends on the driving context, and the paper notes that this space requires careful design.The formulation also conditions future motion on map context, the actor’s history, and other actors’ observed states.Source: §III · p. 2
- GANet’s three-stage pipeline encodes context, predicts possible goals, then forecasts multiple feasible trajectories with confidence scores.GoICrop retrieves and aggregates goal-area map features and models actors’ future interactions.Source: §III · p. 2, Figure 1 · p. 3
- Rather than relying only on local map information, GANet incorporates map features from goal areas to support trajectory prediction.The method predicts possible goals from actor features and uses them as anchors to retrieve and aggregate goal-area map elements.Source: §III · p. 3
A. Motion history and scene context encoding
The context encoder combines actor-motion and vectorized-map features to form actor and lane-node representations.
A. Motion history and scene context encoding
- The encoder uses a 1D CNN with an FPN for actor histories, multi-scale LaneConv for lane graphs, and fusion across actors and lane nodes.It produces actor feature matrix X and lane-node feature matrix Y; the authors also report implementing VectorNet++ as an alternative encoder.Source: A. Motion history and scene context… · p. 3
B. Goal prediction
The goal-prediction stage estimates multiple endpoint goals and their confidence scores, with an optional middle-position goal to aggregate additional map features. Supervision combines goal classification and regression.
B. Goal prediction
- The goal header predicts E possible endpoint goals and confidence scores, supporting multiple possible driving intents.The paper gives stopping, going straight, and turning at an intersection as examples of multimodal intent.Source: B. Goal prediction · p. 3
- The closest predicted endpoint goal to the ground-truth endpoint is designated positive, and the stage combines classification and regression losses.The classification margin encourages each goal to capture a specific pattern, while smooth L1 regression is applied to positive goals.Source: B. Goal prediction · p. 3
- An optional middle-position goal regresses the actor’s trajectory midpoint and aggregates map features to assist endpoint-goal and full-trajectory prediction.The passage identifies the supervised midpoint as the trajectory’s ground-truth middle position.Source: B. Goal prediction · p. 3
- The GANet M 3 overview combines a middle-position goal module with three goal-area predictions before producing K trajectory predictions and confidence scores.The three-goal module aggregates goal-area features and models future actor interactions.Source: Figure 2 · p. 4
C. GoICrop
GoICrop gathers lane context around predicted goal anchors and uses it to update actor features. The framework also applies the operator to predicted anchors to model actors’ future interactions.
C. GoICrop
- A 6-meter crop around the highest-confidence predicted goal anchor forms a more tolerant goal area for trajectory prediction under motion uncertainty.The actual endpoint is described as more likely to fall within candidate areas than among scattered endpoint predictions.Source: C. GoICrop · p. 4
- GoICrop uses actor features as a query and anchor-to-lane-node relative distances to aggregate goal-area lane features back into actor features.The operator is described as spatial distance-based attention; it concatenates the query, relative-distance feature, and lane-node feature before aggregation.Source: C. GoICrop · p. 4
- Using predicted anchors, GoICrop models future actor interactions by considering other actors whose future anchors are within 100 meters.In this use, the attended features and relative distances are those of other actors and their predicted anchors.Source: C. GoICrop · p. 4
D. Motion estimation and scoring
The third stage predicts multiple future trajectories and confidence scores from updated actor features. It trains trajectory regression and mode classification with separate losses, including extra emphasis on endpoint accuracy.
D. Motion estimation and scoring
- The third stage predicts K future trajectories and corresponding confidence scores for each actor from the updated actor features.A two-branch header uses regression for trajectories and classification for scores.Source: D. Motion estimation and scoring · p. 4
- The classification branch uses a margin loss, while trajectory regression applies smooth L1 loss to all steps of positive trajectories.The positive mode is selected by the smallest Euclidean distance between its endpoint and the ground-truth endpoint.Source: D. Motion estimation and scoring · p. 4
- An additional loss term emphasizes the endpoint penalty for predicted positive trajectories.The provided passages identify the ground-truth and predicted positive-trajectory endpoints as the endpoint quantities involved.Source: D. Motion estimation and scoring · p. 4
E. Training
The training objective combines the differentiable modules’ losses, with parameters chosen to balance training. The accompanying table reports test-set results for Argoverse 1 and Argoverse 2.
E. Training
- The model is trained with a loss function because all modules are differentiable.The passage does not specify the full combined loss in the supplied text.Source: E. Training · p. 4
- The training parameters are chosen to balance the training process.No specific balancing parameters are given in the supplied passage.Source: E. Training · p. 4
- Table I covers test-dataset results for Argoverse 1 and Argoverse 2; a dash marks results not reported in the cited papers.The supplied table passage gives its scope and dash convention, but no cell values.Source: Table I · p. 5
A. Experimental settings
The experiments use Argoverse 1 and Argoverse 2 motion-forecasting datasets with different observation and prediction horizons. The reported implementation trains for 42 epochs on two A100 GPUs.
A. Experimental settings
- Argoverse 1 contains over 30K real-world sequences, and the task predicts an agent’s next 3 seconds from an initial 2-second observation.Training and validation sequences are 5 seconds long; test sequences provide 2 seconds, with 3 seconds withheld for leaderboard evaluation.Source: A. Experimental settings · p. 5
- Argoverse 2 uses 11-second scenarios, with five seconds observed and six seconds predicted for leaderboard evaluation.Its scenarios span six geographically diverse cities and are approximately twice as long and more diverse than Argoverse 1 scenarios.Source: A. Experimental settings · p. 5
- Training uses two A100 GPUs, batch size 128, Adam, and 42 epochs, with the learning rate decaying from 1 x 10^-3 to 1 x 10^-4 at epoch 32.The initial and decayed learning rates are reported in the implementation passage.Source: A. Experimental settings · p. 5
B. Comparison with State-of-the-art
GANet outperforms goal-based methods and LaneGCN on reported Argoverse comparisons; ablations examine its components, goal counts, and backbone generality.
B. Comparison with State-of-the-art
- On the public leaderboard, GANet beats LaneGCN by 28% in MR6, 15% in minFDE6, 13% in brier-minFDE6, and 9% in minFDE1.The paper also reports that GANet outperforms TNT, LaneRCNN, and DenseTNT, and wins the CVPR 2022 Argoverse 2 Motion Forecasting Challenge.Source: B. Comparison with State-of-the-art · p. 5
C. Ablation studies
- LaneGCN++ improves ADE1 and FDE1, but the enhanced network shows little improvement in multimodal prediction.The ablation study uses LaneGCN as its baseline and adds components progressively.Source: C. Ablation studies · pp. 5–6
- GANet 1 improves minFDE6 by more than 10% over LaneGCN++; GANet 3 performs better, and GANet M 3 improves further.The number of goals has only marginal effects on overall performance, and the GANet models improve all metrics compared with LaneGCN++.Source: C. Ablation studies · p. 6
- With a VectorNet++ backbone, GANet improves minFDE6 by 9.9% and minADE6 by 5.2%, demonstrating generality across scene-context encoders.Source: C. Ablation studies · p. 6
D. Qualitative results
Qualitative examples show GANet predictions following lane topology in challenging cases, while the conclusion describes GoICrop’s role in aggregating lane features and modeling interactions.
D. Qualitative results
- In the distant lane-change example, LaneGCN predicts six straight paths while GANet follows the lane topology; at a complex right turn, LaneGCN diverges from traffic rules while GANet follows the lanes.GANet M 3 produces smoother trajectories than GANet 1 in the lane-change case.Source: D. Qualitative results · p. 6
V. CONCLUSION
- GoICrop extracts and aggregates semantic lane features in goal areas, implicitly modeling map–trajectory and future actor interactions.Experiments on the Argoverse motion forecasting benchmark demonstrate GANet’s effectiveness.Source: §V · p. 6