A Fast and Map-Free Model for Trajectory Prediction in Traffics

Junhong Xiang, Jingmin Zhang, Zhixiong Nan

TL;DR

Researchers have developed a new trajectory prediction model for autonomous driving that eliminates the need for expensive high-definition (HD) maps while significantly boosting processing speed. By decoupling spatial and temporal features into a two-stage architecture, the model achieves state-of-the-art performance among "map-free" methods and rivals the accuracy of complex map-based systems at a fraction of the computational cost.

What Problem Does This Paper Address?

Predicting where a car or pedestrian will move next is a cornerstone of autonomous driving. Most current high-performing models rely on High-Definition (HD) maps to understand the environment. However, HD maps are expensive to build, difficult to maintain in real-time, and often unavailable in rural or rapidly changing urban areas.

Furthermore, many existing models prioritize accuracy at the expense of speed. Many use Transformer architectures, which are powerful but suffer from high computational complexity—specifically $O(S^2 \times T^2)$, where $S$ is the number of agents and $T$ is the time sequence length. This makes real-time deployment on vehicle hardware challenging. The authors aim to bridge this gap by creating a model that is both map-free and computationally efficient.

Methods

The authors propose a two-stage architecture that breaks down the prediction task into manageable pieces:

  1. Single Agent Encoder: Instead of processing everything at once, the model first looks at each agent individually. It uses a combination of LSTMs and Multi-Head Attention to extract spatial features ($S_A$) and temporal features ($T_A$). Crucially, it processes these dimensions separately, reducing computational complexity from a multiplicative $O(S^2 \times T^2)$ to an additive $O(S^2) + O(T^2)$.
  2. Multiple Agents Interaction: Once individual features are extracted, the model looks at how agents interact. It uses a Graph Convolutional Network (GCN) to handle spatial interactions (e.g., "how close is that car to me?") and a Temporal Transformer to capture how those interactions evolve over time.
  3. Multimodal Decoder: Because the future is uncertain, the model doesn't just predict one path. It uses a Laplace Mixture Density Network (MDN) to generate multiple possible future trajectories (multimodality) along with confidence scores for each.

Datasets

The model was trained and evaluated on the Argoverse dataset, a standard benchmark for motion forecasting.

  • Training/Validation: Over 245,000 sequences.
  • Task: Predict 3 seconds of future motion based on 2 seconds of history.
  • Comparison: The authors compared their results against both map-free baselines (like CRAT-Pred) and map-based heavyweights (like HiVT and TNT).

Key Insights and Ideas

  • Dimension Decoupling is Key: The "aha moment" here is that you don't need to attend to every time step of every agent simultaneously. By decoupling spatial and temporal encoding, the model maintains high accuracy while drastically reducing the math required for each inference.
  • Map-Free Doesn't Mean Accuracy-Poor: The results suggest that rich interaction modeling (GCN + Temporal Transformer) can compensate for the lack of explicit map data. The model learns the "rules of the road" implicitly through the movement patterns of other agents.
  • LSTM + Attention Synergy: While Transformers are the current "it" architecture, the authors found that LSTMs are still highly effective at capturing the sequential dynamics of individual trajectories when paired with an attention mechanism.

Key Results

  • Speed: The model achieves an inference speed of 28ms, making it significantly faster than competitors like CRAT-Pred (178ms) and even map-based models like mmTransformer (59ms).
  • Accuracy: On the Argoverse test set, the model achieved a minADE (Minimum Average Displacement Error) of 0.93m. This is the best performance recorded for a map-free method in the study and is competitive with several established map-based models (e.g., it outperforms the map-based MTPLA and PRIME models).
  • Efficiency: The model is roughly twice as fast as the map-free version of HiVT-64 while providing better prediction accuracy.

Limitations and Future Work

  • Lack of Visual Context: Since the model is map-free and relies solely on trajectory coordinates, it may struggle in complex scenarios where visual cues (like traffic lights or temporary road construction) are the primary drivers of behavior.
  • Static vs. Dynamic: While the model handles dynamic interactions well, it doesn't explicitly account for static obstacles that aren't "agents" (like a parked car or a wall) unless they are represented in the trajectory data.
  • Future Directions: The authors suggest that further optimizing the attention mechanisms or integrating lightweight sensor data (without full HD maps) could further refine the model's performance in edge cases.

Summary updated

Abstract

To handle the two shortcomings of existing methods, (i) nearly all models rely on high-definition (HD) maps, yet the map information is not always available in real traffic scenes and HD map-building is expensive and time-consuming and (ii) existing models usually focus on improving prediction accuracy at the expense of reducing computing efficiency, yet the efficiency is crucial for various real applications, this paper proposes an efficient trajectory prediction model that is not dependent on traffic maps. The core idea of our model is encoding single-agent's spatial-temporal information in the first stage and exploring multi-agents' spatial-temporal interactions in the second stage. By comprehensively utilizing attention mechanism, LSTM, graph convolution network and temporal transformer in the two stages, our model is able to learn rich dynamic and interaction information of all agents. Our model achieves the highest performance when comparing with existing map-free methods and also exceeds most map-based state-of-the-art methods on the Argoverse dataset. In addition, our model also exhibits a faster inference speed than the baseline methods.
BibTeX
@inproceedings{xiang2023fast,
  author = {Junhong Xiang and Jingmin Zhang and Zhixiong Nan},
  title = {{A Fast and Map-Free Model for Trajectory Prediction in Traffics}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  url = {https://openreview.net/forum?id=GedmoWQwjt},
}

Loading related content...

Figure