Abstract
Trajectory prediction in urban mixed-traffic zones (a.k.a. shared spaces) is critical for many intelligent transportation systems, such as intent detection for autonomous driving. However, there are many challenges to predict the trajectories of heterogeneous road agents (pedestrians, cyclists and vehicles) at a microscopical level. For example, an agent might be able to choose multiple plausible paths in complex interactions with other agents in varying environments. To this end, we propose an approach named Multi-Context Encoder Network (MCENET) that is trained by encoding both past and future scene context, interaction context and motion information to capture the patterns and variations of the future trajectories using a set of stochastic latent variables. In inference time, we combine the past context and motion information of the target agent with samplings of the latent variables to predict multiple realistic trajectories in the future. Through experiments on several datasets of varying scenes, our method outperforms some of the recent state-of-the-art methods for mixed traffic trajectory prediction by a large margin and more robust in a very challenging environment. The impact of each context is justified via ablation studies.
Introduction
Predicting the movement of diverse agents in urban "shared spaces"—areas lacking clear lane markings or traffic signals—is a fundamental challenge for autonomous systems. These environments are characterized by high heterogeneity (cars, cyclists, and pedestrians), multi-modality (multiple socially acceptable paths), and complex contextual influences. Existing models often struggle to integrate social dynamics and environmental constraints into a unified framework that generalizes across different urban layouts.
Methodology
The Multi-Context Encoder Network (MCENET) addresses these challenges by leveraging a Conditional Variational Autoencoder (CVAE) architecture. The model integrates three primary types of context: motion history, social grouping logic, and environmental scene context.
Contextual Encoding
- Grouping Context: Unlike models that treat all neighbors as obstacles, MCENET uses the DBSCAN clustering algorithm to identify "friends" moving together as a group. This allows the model to differentiate between group-following behavior and obstacle avoidance.
- Triple Scene Context: The model extracts features from three environmental sources: statistical Motion Heat Maps of common paths, Aerial Photographs for global visual context, and Segmented Maps (binary masks) defining accessible areas like roads versus sidewalks.
Network Architecture
The system employs a dual-encoder setup during training. The X-Encoder processes past information (motion history, agent type, and environment), while the Y-Encoder processes future ground-truth trajectories to help the model learn the underlying distribution of possible paths.
During inference, the model samples from a learned latent space to generate $N$ plausible future trajectories, which are then ranked using a bivariate Gaussian distribution to identify the most likely path.
Experimental Evaluation
MCENET was evaluated on mixed-traffic datasets (Stanford Drone Dataset's Gates3 roundabout, HBS, and HC) and pedestrian-only benchmarks (ETH and UCY).
Key Results
The model consistently outperformed state-of-the-art baselines like S-LSTM and S-GAN.
- Mixed Traffic Accuracy: On the challenging Gates3 dataset, MCENET reduced the Average Displacement Error (ADE) to 1.48m, a significant improvement over S-LSTM's 3.63m.
- Pedestrian Benchmarks: On the "Hotel" dataset, MCENET achieved an ADE of 0.23m, nearly doubling the performance of Social-GAN (0.48m).
- Ablation Studies: Results confirmed that the full model (Motion + Grouping + Scene Context) outperforms any subset, justifying the multi-context approach.
Discussion and Conclusion
The success of MCENET highlights several key insights. Distinguishing social "friends" from strangers is vital for realistic path prediction, as agents synchronize speed with group members but maintain repulsive distances from others. Furthermore, motion heat maps act as a powerful "scene memory," often providing more predictive value than raw aerial imagery. While the model requires high-quality environmental data and faces increasing uncertainty over longer horizons, its ability to be fine-tuned for new environments makes it a robust candidate for real-world intelligent transportation systems.
Summary updated