Pedestrian Trajectory Pediction using Deep Learning and Social Force Models

Abstract

Pedestrian trajectory prediction is vital for intelligent transportation and human-aware autonomous systems, as it ensures safety and supports context-aware navigation. While LSTM- based deep learning models with attention mechanisms effectively capture temporal dependencies, they often lack the ability to simulate pedestrian motion dynamics, group behavior, and social interactions in crowded urban environments. On the other hand, Social Force Models (SFMs) provide an interpretable, physics-inspired framework capable of modeling social interactions and collective pedestrian behavior but are limited in learning complex temporal patterns from data. This work presents a hybrid framework that integrates socially-aware behavioral features extracted via a Social Force Model (SFM) with deep learning models such as LSTM, BiLSTM, and GRU. Each hybrid variant combines SFM with a recurrent architecture (LSTM+SFM, BiLSTM+SFM, GRU+SFM) to capture both social interactions and sequential motion patterns. Using the JAAD dataset, we derived features such as social interaction forces, temporal transitions, and motion-related attributes like velocity and acceleration to support context-aware trajectory prediction. Evaluation on clean data showed that the proposed hybrid models outperform traditional baselines, offering accurate and semantically rich predictions by combining the interpretability and social reasoning of SFM with the sequential learning capabilities of recurrent neural networks. Notably, the BiLSTM+SFM hybrid model demonstrated the highest effectiveness, achieving the lowest error metrics with (ADE=17.04, FDE=24.65, MSE=213.3, in 0.5s, CFMSE=450.74, CMSE=280.87), significantly outperforming the standalone BiLSTM model, which yielded (ADE =31.81, FDE = 89.89, MSE=576.82, in 0.5s, CFMSE=1195.63, CMSE=803.87). The results highlight that the hybrid model outperforms single-method approaches (using either BiLSTM or SFM alone), particularly in complex multi-agent scenarios where accurate prediction requires modeling crowd influence and interaction patterns. This demonstrates the model’s ability to represent fine-grained temporal dynamics and social responses of pedestrians, making it especially suitable for real-time applications such as autonomous driving systems, intelligent surveillance, and robotic path planning.