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

Junhong Xiang, Jingmin Zhang, Zhixiong Nan

Abstract

Trajectory prediction is a critical component of autonomous driving, yet most high-performing models rely heavily on high-definition (HD) maps, which are costly to produce and maintain. Furthermore, many existing architectures sacrifice computational efficiency for accuracy, hindering real-time deployment. This paper introduces a novel, map-free trajectory prediction model that utilizes a two-stage architecture to decouple spatial and temporal features. By employing a combination of Long Short-Term Memory (LSTM) networks, attention mechanisms, Graph Convolutional Networks (GCN), and temporal transformers, the model captures complex agent dynamics and interactions. Evaluated on the Argoverse dataset, the proposed model achieves state-of-the-art performance among map-free methods and remains competitive with map-based models while significantly reducing inference latency to 28ms.

Introduction

Predicting the future paths of vehicles and pedestrians is essential for safe autonomous navigation. Traditionally, researchers have relied on HD maps to provide environmental context. However, the reliance on HD maps introduces significant bottlenecks: they are expensive to build, difficult to update in real-time, and often unavailable in rural or rapidly changing urban environments.

Beyond the map dependency, computational complexity is a major hurdle. Many state-of-the-art models utilize Transformer architectures where the complexity scales at $O(S^2 \times T^2)$, where $S$ represents the number of agents and $T$ is the sequence length. This quadratic growth makes real-time processing difficult on standard vehicle hardware. This research addresses these challenges by proposing a map-free model that prioritizes both predictive accuracy and computational efficiency through dimension decoupling.

Methodology

The proposed model follows a two-stage architecture designed to extract rich features without the need for explicit map data.

Stage 1: Single Agent Encoder

The first stage focuses on individual agent dynamics. Rather than processing all agents and time steps simultaneously, the model treats them independently at this stage. It uses LSTMs paired with Multi-Head Attention to extract spatial features ($S_A$) and temporal features ($T_A$). By decoupling these dimensions, the computational complexity is reduced from a multiplicative $O(S^2 \times T^2)$ to an additive $O(S^2) + O(T^2)$, facilitating much faster inference.

Stage 2: Multiple Agents Interaction

After encoding individual features, the model explores social interactions. A Graph Convolutional Network (GCN) is employed to model spatial relationships between agents (e.g., proximity and relative positioning). This is followed by a Temporal Transformer that captures how these social interactions evolve over time. This hierarchical approach allows the model to learn the "rules of the road" implicitly from the movement patterns of surrounding traffic.

Multimodal Trajectory Prediction

To account for the inherent uncertainty of the future, the model employs a Laplace Mixture Density Network (MDN) as its decoder. This allows the system to generate multiple potential future trajectories (multimodality) and assign a confidence score to each, providing the autonomous system with a probabilistic understanding of the environment.

Experimental Results

The model was rigorously tested using the Argoverse motion forecasting benchmark, which includes over 245,000 sequences. The task involved predicting 3 seconds of future motion based on a 2-second historical window.

Performance and Accuracy

The model achieved a Minimum Average Displacement Error (minADE) of 0.93m on the Argoverse test set. This result represents the highest performance among map-free methods evaluated. Remarkably, the model also outperformed several map-based state-of-the-art models, such as MTPLA and PRIME, demonstrating that sophisticated interaction modeling can effectively compensate for the absence of HD maps.

Computational Efficiency

Efficiency is a primary contribution of this work. The model recorded an inference speed of 28ms. In comparison, the map-free CRAT-Pred requires 178ms, and the map-based mmTransformer takes 59ms. The proposed model is approximately twice as fast as the map-free version of HiVT-64 while maintaining superior accuracy.

Discussion and Limitations

The success of this model highlights that dimension decoupling is a powerful strategy for reducing the computational overhead of trajectory prediction. By combining LSTMs for sequential dynamics with attention mechanisms for feature weighting, the model achieves a synergy that rivals more complex Transformer-only architectures.

However, certain limitations remain. As a map-free model relying solely on trajectory coordinates, it lacks visual context. It may struggle in scenarios where behavior is dictated by non-agent factors, such as traffic lights or temporary construction barriers. Additionally, the model does not explicitly account for static obstacles (e.g., walls or parked cars) unless their presence is reflected in the movement patterns of other agents.

Conclusion

This paper presents an efficient, map-free trajectory prediction model that bridges the gap between accuracy and real-time performance. By decoupling spatial and temporal encoding and utilizing a two-stage interaction framework, the model provides a viable path for autonomous driving in environments where HD maps are unavailable or too costly to maintain. Future work will focus on integrating lightweight sensor data to improve performance in complex edge cases without sacrificing the model's speed advantages.

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},
}
Figure