側面衝突事故時の車両の回転挙動を考慮したΔV算出方法の検討

棚瀬　昇・勝間田　静江・安藤　貴裕・長岡　靖（トヨタ自動車）

D-Call Net(AACN)の死亡重症率の計算には，衝突事故の厳しさを示す指標として疑似ΔVが用いられている．疑似ΔVの算出は「衝撃が重心に作用」「車両の回転は伴わない」と前提を置いている．その前提の下では，車体のどの非変形部位でも概ね同じΔVが発生する．よって疑似ΔVを車両データと比較する場合はエアバッグECUのΔVを用いてきたが，前提とずれた事故の場合，一定の誤差が発生し，死亡重症率が正しく見積もれない可能性がある．

深層学習による車両損壊程度の自動検知に関する研究

西本　哲也・久保田　和広・中尾　賢人（日本大学）・本村　友一（日本医科大学）・Martion Elsegood・Sam Doecke・Giulio Ponte（アデレード大学）

日本で発生した交通事故に基づいて深層学習により車両損壊程度を判別する画像認識アルゴリズムを構築した．その検証ために南オーストラリアで収集した700台のEDRデータと車体損傷画像に適用して，実用性を検討した．その結果，破損程度を70％の的中率で瞬時に認識できることが分かった．

How does pre-crash environment affect injury risk? Injury prediction and analysis based on graph neural network

JUNHAO WEI･Yusuke Miyazaki (Institute of Science Tokyo)･Fusako Sato (JARI)

Injury severity in vehicle crashes is influenced by various factors within the pre-crash environment including environment, vehicle and driver attributes, together with other pre-crash attributes. To investigates the relationships among the factors to identify key determinants of injury outcomes, this study employs graph neural network to model complex interactions and dependencies from police-reported tabular database. The analysis reveals critical contributors to injury severity, uncovering relationships among variables in pre-crash environment. These findings provide actionable insights for enhancing traffic safety and developing effective injury prevention strategies.

Investigation of Severe Injury Probability Prediction Models by Body Parts Through Decision Tree-Based Machine Learning Approach

YIMENG MEI･Haruhiro Fukushima･Yusuke Miyazaki (Institute of Science Tokyo)･Fusako Sato (JARI)

Traffic accidents result in countless economic losses and casualties worldwide annually, making reducing casualties in road accidents always been a hot topic in research. In this context, quick and accurate prediction of occupant injury severity in traffic accidents helps emergency services respond more effectively. However, previous research effects have mainly concentrated on predicting overall injury severity rather than looking at injuries to specific body parts, which limits the precision of injury assessment and targeted emergency response. In this study, we aim to develop specific Random Forest-based models to predict injury severity for different body parts, including the head, face, neck, thorax, abdomen, spine, and limbs. These models facilitate the analysis of correlations between various accident risk factors and body part-specific injuries. Furthermore, they enable emergency services to implement more precise and targeted response strategies after accidents happened.