A Spatio-temporal Graph Network Allowing Incomplete Trajectory Input for Pedestrian Trajectory Prediction

TL;DR

STGN-IT tackles pedestrian trajectory prediction under incomplete historical data by modeling pedestrians and obstacles as a spatio-temporal graph and encoding observation states to distinguish missing data. The method uses a two-stage prediction pipeline: an initial pass without environment context, followed by obstacle augmentation from occupancy grid maps and a second, refined prediction; DBSCAN-based node ordering and clustering enhance interaction modeling. Empirical results on STCrowd show STGN-IT achieves state-of-the-art accuracy (lower ADE and FDE) with modest performance degradation under incomplete inputs and demonstrates improved safety when evaluated in pad mode for robot navigation. The work advances practical pedestrian forecasting for mobile robots by robustly handling incomplete trajectories and incorporating environmental constraints, with potential impact on real-time navigation and collision avoidance.

Abstract

Pedestrian trajectory prediction is important in the research of mobile robot navigation in environments with pedestrians. Most pedestrian trajectory prediction algorithms require the input historical trajectories to be complete. If a pedestrian is unobservable in any frame in the past, then its historical trajectory become incomplete, the algorithm will not predict its future trajectory. To address this limitation, we propose the STGN-IT, a spatio-temporal graph network allowing incomplete trajectory input, which can predict the future trajectories of pedestrians with incomplete historical trajectories. STGN-IT uses the spatio-temporal graph with an additional encoding method to represent the historical trajectories and observation states of pedestrians. Moreover, STGN-IT introduces static obstacles in the environment that may affect the future trajectories as nodes to further improve the prediction accuracy. A clustering algorithm is also applied in the construction of spatio-temporal graphs. Experiments on public datasets show that STGN-IT outperforms state of the art algorithms on these metrics.

Figures (5)

• Figure 1: Comparison of the STCrowd dataset and the ETH dataset. Green boxes indicate pedestrians are observable and orange boxes indicate pedestrians are obscured. Pedestrians are more likely to be obscured in the egocentric view than in the top-down view.
• Figure 2: Label and Prediction results for filtration mode, pad mode, and encoding mode with incomplete trajectories. Incomplete trajectories are not predicted in filtration mode. Incomplete trajectories are predicted in pad mode with unobservable positions set to 0. The encoding method encodes the observation state of positions.
• Figure 3: The scene picture (top), point cloud map (bottom left), and occupancy grid map (bottom right) of a scene in the STCrowd dataset. The occupancy grid map can be automatically generated from the point cloud map.
• Figure 4: The STGN-IT algorithm includes two predictions. The first prediction searches for obstacles in the environment and adds them to the spatio-temporal graph to provide more information for the second prediction.
• Figure 5: The prediction results of STGN-IT and some state of the art algorithms. GraphTERN makes predictions in condition "f-f". STGN-IT, SGCN, STIGCN make predictions in condition "p-p". GraphTERN does not make predictions in scene B, F, G, indicating that the "f-f" prediction condition is more likely to cause robot collisions. The trajectories predicted by the STGN-IT are more reasonable, as the trajectories avoid static obstacles in scenes A, B, C and avoid other pedestrians in scenes D, E, F. With incomplete trajectory input in scene G, predictions of STGN-IT are smoother and more reasonable.