Crash Severity Analysis of Child Bicyclists Using Arm-Net and Mambanet

Somvanshi, Shriyank; Chakraborty, Rohit; Das, Subasish; Dutta, Anandi K · 2025 · Unknown

DOI: 10.1109/cai64502.2025.00146

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical safety issue of child bicyclist crashes, focusing on predicting crash severity to inform infrastructure and policy interventions. Child bicyclists (aged 14 and younger) are among the most vulnerable road users, frequently suffering severe injuries or fatalities due to limited experience, poor hazard perception, and complex traffic environments. The research aims to identify key risk factors and evaluate the performance of advanced deep learning models in predicting crash outcomes, specifically comparing ARM-Net and MambaNet. The methodology utilized a dataset of 2,394 child bicyclist crashes in Texas from 2017 to 2022. Crash severity was categorized into three classes: Fatal/Severe (KA), Moderate/Minor (BC), and No Injury (O). To address significant class imbalance, the authors applied the SMOTEENN resampling technique, which combined synthetic oversampling with noise reduction, resulting in a balanced dataset of 3,567 samples. The data underwent preprocessing, including feature scaling and one-hot encoding. Feature importance was initially assessed using XGBoost and Random Forest, identifying top predictors such as weather, lighting, road alignment, traffic control, and helmet use. The core experimental design involved training ARM-Net and MambaNet models using a 60/20/20 split for training, validation, and testing. Hyperparameter optimization was conducted via random search over 100 configurations for each model, with training performed over 50 epochs using GPU acceleration. The results demonstrated that MambaNet outperformed ARM-Net in overall predictive accuracy and robustness. MambaNet achieved a validation accuracy of 92%, compared to 88% for ARM-Net. In predicting Fatal/Severe (KA) crashes, MambaNet attained 90% precision, 98% recall, and a 94% F1-score, whereas ARM-Net achieved 86% precision, 96% recall, and a 91% F1-score. For No Injury (O) crashes, MambaNet maintained high performance with 94% precision, recall, and F1-score, while ARM-Net scored 90%, 87%, and 89%, respectively. Both models struggled with Moderate/Minor (BC) crashes, likely due to overlapping characteristics with other categories; MambaNet achieved 63% recall and 74% F1-score for BC, while ARM-Net achieved 67% recall and 76% F1-score. Log loss analysis indicated that MambaNet exhibited better generalization and stability, with smoother loss reduction and less overfitting compared to ARM-Net. The significance of this study lies in its validation of advanced tabular deep learning methods for traffic safety analysis. The findings highlight that MambaNet is particularly effective for identifying extreme crash severities, making it a strong candidate for integration into Intelligent Transportation Systems (ITS) and connected vehicle frameworks for proactive crash mitigation. The identified risk factors, such as poor lighting and lack of traffic control, provide actionable insights for infrastructure improvements like better street lighting and protected bike lanes. The study concludes that while current models are effective, future research should incorporate continuous variables and real-time behavioral data to further enhance predictive accuracy and support equitable safety interventions.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-09; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• vru crash typology