Smartphone-based Risky Traffic Situation Detection and Classification

Although the number of traffic accidents occurring in Japan is decreasing, there still happen approximately 400,000 traffic accidents annually. Behind such accidents, there are frequent minor incidents (near-miss incidents) that may lead to such serious accidents. Analyzing such minor incidents is effective to reduce accidents, but the challenge is to design and deploy a method to collect and analyze such incident information. Drive recorders may be useful for such a purpose, but they cannot collect information from those vehicles without recorders. In this study, we propose the design and development of a platform that aggregates behavioral data from pedestrians and vehicle drivers using their smartphones, and automatically estimates risky traffic situations from the aggregated data. We present our preliminary result of detecting and classifying those events in a controlled environment and have achieved F-value 0.89 for four categories classification.

I. Introduction

It is said that behind traffic accidents, which are officially reported and recorded, a number of unrecorded minor incidents, which are significant signs for future serious accidents in the same or similar situations, have occurred. Actually, detection, collection and analysis of such minor incidents are not straightforward. For example, in Japan, there exists near-miss database [1] that collects videos from drive recorders installed in business vehicles, e.g., taxis. Those drive recorders are able to detect unusual stops (e.g., severe deceleration) of vehicles by built-in acceleration sensors and the video clips of several seconds before and after the events are stored in the local storage. Each video is then given to the manual classification by an administrator and is stored in the database if it is recognized as a case. However, it is reported that about 70% of such videos are false-positive data such as deceleration due to bumps, which are not actually the near-miss cases. Therefore, a considerable amount of human resources is required in data selection. Besides, drive recorders do not always capture all the scenes, as they record only the front views. Let us consider the fact that near-miss often occur due to pedestrians' unsafe behaviors (e.g., the sudden appearance of pedestrians from drivers blind spots). However, from the recorded scenes, the pedestrians' trajectories are not known and the deep understanding of the cause behind the near-miss is not possible. There are more complicated cases where multiple entities (vehicles, bikes and pedestrians) relate with each other to cause near-miss but the driver recorders may capture only a part of scenes.