A Generative AI-Driven Architecture for Intelligent Transportation Systems

The rapid acceleration of urbanization underscores the urgent need for developing intelligent transportation systems (ITS) to enhance the efficiency, safety, and sustainability of urban mobility. Within this context, accurately predicting vehicle trajectories is paramount for facilitating superior traffic management and control. To this end, the paper presents an innovative architecture that combines a Long Short-Term Memory (LSTM) module with a generative artificial intelligence (Gen-AI) component, specifically the RoBERTa Transformer model. By leveraging these sophisticated architecture, the LSTM network with a recursive decoder outperforms the teacher forcing decoder on clean datasets, showing higher robustness in time-series predictions. When video data was partially missing, performance decreased, but using the RoBERTa model to reconstruct the missing data significantly improved results for both decoders (from $\mathbf{3 7 \%}$ up to $\mathbf{9 2 \%}$). The reconstructed data notably enhanced the performance of the LSTM models, particularly when larger portions of the video data were absent. These findings highlight the effectiveness of data reconstruction techniques in mitigating the challenges posed by uncontrollable events (common in real ITS scenarios) which can bear to incomplete information.