FMI-PTP: Pedestrian trajectory prediction based on field of view and motion intent

Pedestrian trajectory prediction is a critical component for optimizing autonomous driving decision-making and enhancing road safety, with its core challenges stemming from the complexity of social interactions and the obscurity of motion intentions. Existing methods often struggle to address these challenges: first, they tend to dissociate social interactions from motion intentions, failing to capture the intrinsic nature of pedestrian motion; second, deep learning models are mostly "black-box" in structure, lacking interpretability and underutilizing the value of visual information. To address these issues, this paper proposes a Field-of-View and Motion Intention-based Pedestrian Trajectory Prediction model (FMI-PTP). First, the model constructs a sector-shaped field of view for pedestrians and incorporates a two-stage filtering mechanism to accurately capture visually dominated social interactions, while introducing affine transformations to achieve data augmentation with interaction-invariant properties. Second, it employs Gaussian Mixture Models to perform multimodal clustering on historical trajectories, quantifying latent motion intentions into interpretable modality labels. On this basis, a Transformer encoder-decoder is adopted to fuse multi-source features, including interaction and intention features, and a joint loss function combining Huber loss and smoothed cross-entropy is applied to simultaneously optimize trajectory accuracy and modality distribution, enabling parallel generation of multimodal trajectories. Experiments on the ETH, UCY, and SDD datasets demonstrate that the proposed method outperforms existing mainstream models across multiple evaluation metrics, while maintaining low parameter counts and high inference efficiency. The framework thus provides a trajectory prediction solution for autonomous driving systems that balances accuracy, interpretability, and real-time performance.