Traffic accident statistics have suggested that failures in safety confirmation and/or hazard recognition resulting in a delayed response are predominant human errors in major crash types in Japan. A naturalistic driving data acquisition system was developed to investigate the human factors contributing to these accidents. A preliminary analysis was performed to evaluate the efficiency of the collected naturalistic data. An analysis of vehicle-to-pedestrian conflict data demonstrated that there are different types of recognition failure depending on the road section being travelled. This result suggests that naturalistic driving data can provide valuable information to investigate factors contributing to the risk of human error.