In recent years, a variety of crowd-sourcing data have been addressed in road safety studies with the popularization of mobile intelligent terminals. The crowd-sourcing data usually contain the location, velocity, acceleration of the mobile terminals, which may be related to dangerous driving behaviors, and even traffic crash records reported by users. In contrast to traditional crash record data, these crowd-sourcing data have the potential to reflect more detailed insights of road safety performance. In this paper, we proposed a novel method to conduct road safety assessments based on crowd-sourcing data provided by AutoNavi software co. LTD. Basically, with the user-report crash records as a safety indicator, we adopted both of the traditional crash prediction model and the machine learning model to evaluate the safety performance by taking various risk factors into account. As repeated crash reports exist, a merging and screening process should be conducted carefully first. Alternatively, some dangerous driving behaviors such as speeding, rapid accelerations or sharp turn may also be considered as safety indicators. We thus examined the correlations of these indicators and traffic crash record, and in further, revealed the substitutability and applicability of dangerous driving indicators for particular road types. Findings in this paper can be used to disclose alternative indicators for traffic crashes and emergencies. We also believe that crowd-sourcing data are worth further exploration to bring its full potential in road safety assessments.

CICTP