TAG-BiLSTM: A Temporal Attention-guided BiLSTM for Real-time Rear-end Collision Prediction in Intelligent Transportation Systems

Ebenezer Penoo, Akilan Thangarajah
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:308-318, 2026.

Abstract

Rear-end collision prediction is a critical component of modern Intelligent Transportation Systems (ITS), enabling the early detection of hazardous driving conditions and supporting advanced driver assistance functions. Although recent deep learning (DL) approaches have shown strong predictive capability, many architectures, especially those relying on heavy attention or graph-based computations, suffer from high inference cost, limiting their suitability for real-time deployment on edge devices. To address this challenge, we propose a light-weight temporal attention-guided bidirectional long short-term memory (TAG-BiLSTM) network that efficiently models historical vehicle-kinematic sequences while capturing long-range temporal dependencies. The BiLSTM backbone encodes forward and backward motion context, while the attention gate selectively emphasizes the most safety-critical frames associated with rapid deceleration, headway compression, or abrupt motion changes. Experimental evaluation on benchmark open-source datasets, including the Next Generation Simulation (NGSIM)1 and highD2, demonstrates the robustness of the proposed model, achieving a mean performance exceeding 94%.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-penoo26a, title = {TAG-BiLSTM: A Temporal Attention-guided BiLSTM for Real-time Rear-end Collision Prediction in Intelligent Transportation Systems}, author = {Penoo, Ebenezer and Thangarajah, Akilan}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {308--318}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/penoo26a/penoo26a.pdf}, url = {https://proceedings.mlr.press/v318/penoo26a.html}, abstract = {Rear-end collision prediction is a critical component of modern Intelligent Transportation Systems (ITS), enabling the early detection of hazardous driving conditions and supporting advanced driver assistance functions. Although recent deep learning (DL) approaches have shown strong predictive capability, many architectures, especially those relying on heavy attention or graph-based computations, suffer from high inference cost, limiting their suitability for real-time deployment on edge devices. To address this challenge, we propose a light-weight temporal attention-guided bidirectional long short-term memory (TAG-BiLSTM) network that efficiently models historical vehicle-kinematic sequences while capturing long-range temporal dependencies. The BiLSTM backbone encodes forward and backward motion context, while the attention gate selectively emphasizes the most safety-critical frames associated with rapid deceleration, headway compression, or abrupt motion changes. Experimental evaluation on benchmark open-source datasets, including the Next Generation Simulation (NGSIM)1 and highD2, demonstrates the robustness of the proposed model, achieving a mean performance exceeding 94%.} }
Endnote
%0 Conference Paper %T TAG-BiLSTM: A Temporal Attention-guided BiLSTM for Real-time Rear-end Collision Prediction in Intelligent Transportation Systems %A Ebenezer Penoo %A Akilan Thangarajah %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-penoo26a %I PMLR %P 308--318 %U https://proceedings.mlr.press/v318/penoo26a.html %V 318 %X Rear-end collision prediction is a critical component of modern Intelligent Transportation Systems (ITS), enabling the early detection of hazardous driving conditions and supporting advanced driver assistance functions. Although recent deep learning (DL) approaches have shown strong predictive capability, many architectures, especially those relying on heavy attention or graph-based computations, suffer from high inference cost, limiting their suitability for real-time deployment on edge devices. To address this challenge, we propose a light-weight temporal attention-guided bidirectional long short-term memory (TAG-BiLSTM) network that efficiently models historical vehicle-kinematic sequences while capturing long-range temporal dependencies. The BiLSTM backbone encodes forward and backward motion context, while the attention gate selectively emphasizes the most safety-critical frames associated with rapid deceleration, headway compression, or abrupt motion changes. Experimental evaluation on benchmark open-source datasets, including the Next Generation Simulation (NGSIM)1 and highD2, demonstrates the robustness of the proposed model, achieving a mean performance exceeding 94%.
APA
Penoo, E. & Thangarajah, A.. (2026). TAG-BiLSTM: A Temporal Attention-guided BiLSTM for Real-time Rear-end Collision Prediction in Intelligent Transportation Systems. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:308-318 Available from https://proceedings.mlr.press/v318/penoo26a.html.

Related Material