ABi-LSTM: Attention-Based Bidirectional-LSTM for Multiclass Traffic Agents Trajectory Prediction

This study proposes the Attention-Based Bidirectional Long Short-Term Memory (ABi-LSTM) model for multimodal trajectory prediction of multiclass traffic agents in cross-perspective scenarios. Unlike prior models that primarily focus on pedestrian agents or static viewpoints, ABi-LSTM incorporates dynamic traffic contexts by integrating agent location, speed, motion status, atomic actions, and ego-vehicle odometry data. First, a bidirectional LSTM architecture captures temporal dependencies from both past and future contexts, enabling richer sequential representation. Second, an attention mechanism is integrated to enhance the features extraction of relevant dynamic traffic contexts. We employ shared network trained jointly on both egocentric and top-down data, leveraging cross-perspective information producing view-specific trajectory predictions. ABi-LSTM demonstrates superior performance compared to previous benchmarks while maintaining the computational complexity on two datasets: the egocentric TITAN and top-down SDD benchmarks. Both datasets represent highly interactive urban traffic scenarios, showcasing the robustness of our model.