StarNet: Global Interaction Network for Pedestrian Trajectory Prediction

This article presents the research and implementation of StarNet, a neural‑network‑based algorithm for predicting pedestrian trajectories in large‑scale delivery scenarios. The work originates from Meituan’s autonomous delivery team and was published at IROS 2019.

Background : Improving the overall efficiency of massive delivery networks requires accurate prediction of surrounding pedestrians for safe and smooth operation of autonomous vehicles. Existing methods mainly model pairwise interactions using LSTM or traditional filters, which are computationally expensive and ignore global context.

Significance of pedestrian trajectory prediction : Accurate forecasts enable autonomous vehicles to plan collision‑free paths and reduce unnecessary trajectory fluctuations.

Challenges : (1) Pedestrians exhibit highly flexible motion, making it difficult to define a deterministic dynamics model. (2) Interactions among multiple pedestrians are complex and abstract, often modeled only by relative positions, orientations, or velocities.

Related work : Traditional approaches rely on Kalman filters, HMM, Gaussian Processes, or early LSTM‑based models such as Social‑LSTM and Social‑GAN. Recent methods incorporate appearance, skeleton, scene layout, and graph‑based representations (GCN, GNN, Message Passing).

StarNet Overview : StarNet introduces a global dynamic map that aggregates the positions of all obstacles at each timestep, forming a spatio‑temporal feature map. Two subnetworks are employed:

Hub Network – a global interaction module that encodes the dynamic map using a fully‑connected layer, max‑pooling, and an LSTM to produce a shared interaction vector.

Host Network – an LSTM‑based trajectory predictor that queries the global interaction vector for each target pedestrian and combines it with the pedestrian’s own state to forecast the next position.

The interaction computation uses simple dot‑product operations (similar to attention) rather than exhaustive pairwise calculations, reducing complexity from O(N²) to O(N).

Experimental Evaluation : StarNet was compared with four classic baselines on the UCY and ETH datasets (ZARA‑1, ZARA‑2, UNIV, ETH, HOTEL). Metrics include Average Displacement Error (ADE), Final Displacement Error (FDE), inference time, and model size. Results show that StarNet outperforms baselines in ~80% of scenarios while maintaining a fast inference time (LSTM inference ≈ 0.029 s).

Conclusions : StarNet achieves higher accuracy by modeling global interactions via a shared dynamic map and improves efficiency through the Hub‑Host architecture.

Future Work : Plans include exploring more sophisticated model structures, enhancing interpretability of interaction modeling, and incorporating explicit obstacle tracking across timesteps.