Introduction
Predicting pedestrian movement is a cornerstone of safe autonomous navigation. To be effective, models must account for both individual intent and social context—how people avoid collisions or move in groups. However, existing state-of-the-art (SOTA) solutions, often relying on Recurrent Neural Networks (RNNs) or Generative Adversarial Networks (GANs), suffer from high latency. This computational burden limits their deployment on edge devices, such as delivery robots, where real-time decision-making is critical.
Methodology: CARPe Posterum
The proposed framework, CARPe Posterum, replaces sequential processing with parallelizable convolutional operations to achieve high-speed inference.
Social Graph Module
To model social context, pedestrians are represented as nodes in a fully connected graph. The system employs Graph Isomorphism Networks (GIN), which offer superior discriminative power compared to standard graph attention mechanisms. This allows the model to effectively capture complex social dynamics using fewer layers and parameters.
Convolutional Prediction Module
Departing from traditional LSTMs, CARPe Posterum utilizes a Convolutional Neural Network (CNN) for trajectory forecasting. By mapping temporal data into a 2D domain, the CNN can analyze past positions across multiple time scales simultaneously. This architecture leverages the parallel processing capabilities of modern GPUs and CPUs, significantly reducing inference time compared to step-by-step RNN processing.
Experimental Evaluation
The model was benchmarked on the ETH and UCY datasets, which feature diverse real-world pedestrian trajectories in environments ranging from university walkways to busy streets.
Performance and Efficiency
CARPe Posterum achieves a remarkable 38x speedup over previous SOTA models, reaching 762 FPS on an Nvidia P100 GPU. Furthermore, it utilizes 97% fewer parameters and floating-point operations (FLOPs) than leading LSTM-based approaches, making it ideal for hardware like the Nvidia Jetson Nano.
Accuracy
Despite its efficiency, the model maintains high precision, yielding an Average Displacement Error (ADE) of 0.54m and a Final Displacement Error (FDE) of 1.05m. These results are competitive with, and in some cases superior to, significantly more complex and slower models.
Discussion and Future Work
The success of CARPe Posterum demonstrates that long-term recurrence is not strictly necessary for short-term path prediction (3–5 seconds). While the model excels at social modeling, future iterations could benefit from integrating visual features, such as body pose or scene segmentation, to better anticipate sudden changes in individual intent that social norms alone cannot predict.
Summary updated