Integrated active and passive safety protection systems have made substantial contributions to reducing traffic accidents and mitigating human injuries. However, assessing such systems through vehicle collision tests is limited, as this approach cannot cover the wide range of accident scenarios. To address this gap, identifying and generating representative pre-crash scenarios from real-world accidents provides key boundary conditions for the setup of virtual test scenarios. In this study, we used the Future Mobile Traffic Accident Scenario Study (FASS) dataset to reconstruct 112 two-wheeler accidents. For each case, we extracted pre-crash dynamic information, static attributes, and environmental context. An autoencoder was employed to encode high-dimensional features of scenarios, and K-means clustering was applied to categorize the accidents into eight representative pre-crash scenarios. For each scenario, we examined the motion states of participants and further compared the behaviors and injuries between motorcycle and bicycle accidents. The results show that motorcycles have a higher average pre-crash speed than bicycles (22.4 vs 16.2 km/h), resulting in a greater proportion of Maximum Abbreviated Injury Scale (MAIS) 4+ cases (35.6% vs 18.9%). Furthermore, a Gaussian Mixture Model was applied to fit the scenario features. This model was then used to randomly generate the initial conditions of pre-crash scenarios, including the positions, speeds, and yaw angles of vehicles and two-wheelers. The proposed scenario generation method was first applied to pre-crash scenario construction for integrated safety systems tests, and it can create diverse, statistically grounded scenarios that reflect real-world accident distributions. These scenarios can broaden the test scope of integrated safety systems and support their implementation.