Source-linked AI summary
SGCN:Sparse Graph Convolution Network for Pedestrian Trajectory Prediction
Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, Gang Hua
TL;DR
Pedestrian trajectory prediction is difficult because complex interactions can make dense undirected models introduce superfluous relations while neglecting motion tendency. SGCN learns sparse directed spatial and temporal graphs and uses them to predict trajectories, outperforming comparative methods on ETH and UCY. Its evaluation follows assumptions about trajectory coordinates and temporal graph structure.
Problem
Complex pedestrian interactions challenge trajectory prediction, while dense undirected models can introduce superfluous interactions and neglect trajectory motion tendency.
Method
SGCN learns sparse directed spatial and temporal graphs for adaptive interactions and motion tendency, then uses graph convolutions and a TCN to estimate bi-Gaussian trajectory distributions.
Results
SGCN outperforms comparative state-of-the-art methods on the ETH and UCY datasets.
Takeaways & Limitations
The method can predict trajectories more accurately in complex scenes and identify sparse directed interactions and motion tendencies.
Takeaways & Limitations
The method assumes temporal dependency through an upper-triangular initial temporal adjacency matrix and assumes trajectory coordinates follow the Social-LSTM formulation.
Abstract
from arXiv · showhide
Pedestrian trajectory prediction is a key technology in autopilot, which remains to be very challenging due to complex interactions between pedestrians. However, previous works based on dense undirected interaction suffer from modeling superfluous interactions and neglect of trajectory motion tendency, and thus inevitably result in a considerable deviance from the reality. To cope with these issues, we present a Sparse Graph Convolution Network~(SGCN) for pedestrian trajectory prediction. Specifically, the SGCN explicitly models the sparse directed interaction with a sparse directed spatial graph to capture adaptive interaction pedestrians. Meanwhile, we use a sparse directed temporal graph to model the motion tendency, thus to facilitate the prediction based on the observed direction. Finally, parameters of a bi-Gaussian distribution for trajectory prediction are estimated by fusing the above two sparse graphs. We evaluate our proposed method on the ETH and UCY datasets, and the experimental results show our method outperforms comparative state-of-the-art methods by 9% in Average Displacement Error(ADE) and 13% in Final Displacement Error(FDE). Notably, visualizations indicate that our method can capture adaptive interactions between pedestrians and their effective motion tendencies.
1. Introduction
Pedestrian trajectory prediction is challenging because pedestrians interact in complex, selective ways and exhibit motion tendencies that dense undirected models may miss. SGCN addresses these issues with sparse directed spatial and temporal graphs, achieving stronger benchmark performance.
- Motivation: Pedestrian trajectory prediction remains challenging because complex interactions can disturb individual pedestrian motion.The task supports applications including autonomous driving, video surveillance, and visual recognition.
- Limitations of prior work: Dense interaction models assume each pedestrian interacts with all others, introducing superfluous interactions.Undirected distance-based models additionally treat pairwise influence as identical in both directions.
- Motion tendency: Motion tendency captures short-term trajectory direction, helping distinguish straight motion from collision-avoidance detours.The paper argues that effective intermediate trajectory points can facilitate prediction.
- Proposed approach: SGCN jointly learns sparse directed spatial and temporal graphs to model adaptive interactions and motion tendencies.Self-attention and asymmetric convolution networks produce sparse directed adjacency structures for subsequent graph convolution.
- Results: SGCN outperforms comparative state-of-the-art methods on the ETH and UCY pedestrian trajectory datasets.The introduction reports extensive experiments supporting this comparison.
2. Related Works
Prior trajectory-prediction methods use recurrent, adversarial, attention, graph, or Transformer-based models, but often represent interactions within fixed neighborhoods or across all pedestrians. SGCN instead dynamically learns sparse directed adjacency and uses a parameter-efficient single attention layer.
- Pedestrian trajectory prediction: Social-LSTM uses recurrent networks and pooled hidden states to model interactions within a certain radius.
- Graph-based approaches: Graph-based methods represent pedestrian interactions with graph attention, relative-distance edges, manually defined groups, or Transformer-based spatial and temporal modeling.Examples include Social-BiGAT, Social-STGCNN, RSGB, and STAR.
- Motivation for SGCN: Previous methods generally model either fixed-range neighborhoods or all pedestrians, which may introduce superfluous interactions.SGCN seeks adaptive pedestrians involved in each interaction and captures effective motion tendency.
- Graph convolution networks: SGCN aggregates nodes using a learned sparse adjacency matrix whose aggregated node set is dynamically determined.
- Self-attention mechanism: Unlike Transformer-block approaches, SGCN uses a single attention layer and is described as parameter-efficient while achieving better performance.
3. Our Method
SGCN learns sparse directed spatial and temporal graphs from trajectory inputs, then combines them through graph convolutions to represent interactions and motion tendencies for prediction.
- Graph Inputs: SGCN constructs spatial graphs over pedestrians and temporal graphs over each pedestrian’s trajectory from observed coordinates.The spatial graph represents pedestrian locations at each time step, while the temporal graph represents each pedestrian’s trajectory.
- Sparse Directed Spatial Graph: Self-attention computes asymmetric dense spatial interactions, which are fused across time before asymmetric convolutions derive high-level interaction features.The resulting feature map has size Tobs × N × N and distinguishes directed influence between nodes.
- Sparse Directed Spatial Graph: Thresholding sigmoid-activated interaction features creates a sparse mask; adding self-connections and multiplying by dense interactions yields the sparse spatial adjacency matrix.The threshold ξ determines which directed interactions are retained.
- Sparse Graph Learning: Zero-Softmax normalizes adjacency matrices while preserving sparsity, avoiding the dense reconnection caused by ordinary Softmax normalization.Ordinary Softmax assigns non-zero values to zero inputs, forcing otherwise non-interacting pedestrians to interact again.
- Trajectory Representation and Prediction: A temporal graph models motion tendency, and two graph-convolution branches process spatial and temporal graphs in opposite orders before summing their outputs.The branches produce interaction-tendency and tendency-interaction features whose sum forms the trajectory representation.
4. Experiments and Analysis
Experiments on ETH and UCY show that SGCN outperforms competing methods, while ablations and visualizations support the roles of sparse directed interaction and motion tendency.
- Comparison with State-of-the-Arts: SGCN significantly outperforms nine competing methods on both ETH and UCY, surpassing STAR by 9% in ADE and Social-STGCNN by 13% in FDE.The comparisons use ADE and FDE on the ETH and UCY datasets.
- Comparison with State-of-the-Arts: On the dense-crowd UNIV sequence, SGCN outperforms dense-interaction methods including SGAN, Sophie, GAT, Social-BiGAT, Social-STGCNN, and STAR.The authors associate this result with removing superfluous interactions through Sparse Directed Interaction.
- Ablation Study: Removing Motion Tendency causes 67% ADE and 83% FDE degradation, while removing Sparse Directed Interaction causes 78% ADE and 96% FDE degradation.Both components therefore contribute substantially to the reported prediction performance.
- Ablation Study: Overall performance peaks at ξ = 0.5; sparse directed interactions outperform dense interactions, while eliminating interactions produces the lowest performance.The threshold study compares no, very sparse, relatively dense, and dense directed interaction variants.
- Visualization: In visualized parallel and meeting scenarios, SGCN produces less-overlapping distributions with better ground-truth tendency than Social-STGCNN and SGAN.For a stationary pedestrian, SGCN also produces smaller variance, while its distributions match multi-pedestrian scenes more closely.
- Visualization: Sparse Directed Interaction captures dynamically changing effective interaction objects, including cases where one node is influenced by only part of the other nodes.The visualized graphs also include scenes where a node interacts with all marked nodes.
5. Conclusion
The proposed sparse graph convolution network leverages sparse directed interactions and motion tendency to improve pedestrian trajectory prediction, including in complex scenes.
- SGCN leverages Sparse Directed Interaction and Motion Tendency for trajectory prediction.
- The method achieves better performance than previous methods according to extensive experimental evaluations.
- SGCN predicts trajectories more accurately in complex scenes, including groups of pedestrians walking in parallel.