Year: 2025
Type: dissertation
Abstract: Powered Two-Wheelers (P2Ws) are highly vulnerable to road accidents, yet developing reliable crash detection systems remains challenging due to the scarcity and imbalance of real-world crash data. Collecting such data is costly, logistically complex, and ethically constrained. This thesis explores synthetic data generation as a solution, evaluating two state-of-the-art generative models, TimeGAN and IMUDiffusion, for producing highfidelity IMU-based time-series crash data. Both models were implemented on proprietary datasets and assessed using a comprehensive framework combining statistical, structural, and temporal similarity metrics, alongside PCA, t-SNE, and histogram visualizations. Results show that IMUDiffusion outperformed TimeGAN in fidelity, temporal consistency, and class-conditional generation. Controlled generation further enabled targeted synthesis of crash, non-crash, and misuse scenarios. Synthetic data was integrated into Autoliv’s crash detection model, revealing nuanced trade-offs: TimeGAN reduced false positives but increased detection latency, while IMUDiffusion improved detection speed but slightly raised false positives. These findings highlight synthetic data’s potential to mitigate data scarcity but also underscore the limitations of current downstream models.
Author: Muzaffar, Muhammad Khizar Bin
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7111683933