Pedestrian Trajectory Prediction Using Deep Learning and Rule Based 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 deep learning models such as LSTM, GRU, and BiLSTM with attention mechanisms effectively model temporal dependencies, they often lack interpretability and require extensive annotated data. In contrast, rule-based models offer semantic clarity but struggle with flexibility in complex scenarios. This work presents a hybrid framework that integrates interpretable behavioral features extracted via a rule-based module with deep learning models. Using the JAAD dataset, we derived indicators like motion states and temporal transitions to support context-aware trajectory prediction. The proposed framework was evaluated by comparing three deep learning models (LSTM, GRU, BiLSTM) and their hybrid counterparts that incorporate rule- based components. Results on clean data demonstrate that the hybrid RBM-BiLSTM model achieves lower prediction errors (ADE = 24.36, CMSE = 437.74) compared to the pure BiLSTM model (ADE = 28.94, CMSE = 642.84), highlighting the benefit of integrating semantic cues for enhanced prediction accuracy.