| Summary |
In the past, the data while driving was seldom collected, and proximate factors were often considered as the main factor in determining the cause of the accidents. However, proximate factors are not necessarily the main cause of the accident, and the prevention of proximate factors
alone may not be effective in preventing the accident. The project aims to shift road safety management from a nearsighted accident view to a comprehensive perspective that includes not only accidents but also dangerous driving events and driving behaviors in the analysis. This report
presents the efforts conducted in the second year of a four-year project, based on the first phase findings, with a focus on in-vehicle image recognition technology. The project discusses the limitations of existing in-vehicle equipment and digital driving recorders, and considers the
demand for analysis of abnormal events both inside and outside of the vehicle. The project aims to establish abnormal event image recognition technology training and integrate data sets to identify correlations between the types of abnormal events and their spatial characteristics both
inside and outside of the vehicle. Ultimately, the project provides recommendations of driver management for industry operators and of accidentprone locations for regulatory authorities.
To develop abnormal event image recognition technologies, the project team interviewed five national highway passenger transport operators to understand their current in-car device installation, common abnormal behaviors, driver safety management, and expectations for advanced driver-assistance systems (ADAS). Through a literature review, the team identified three types of distracted behaviors: cognitive, visual, and manual, and defined abnormal behaviors as those related to the degree of abnormality of manual distraction behaviors. The team then categorized in-car driving behaviors into three main areas: head orientation and deviation degree, hand position and whether it is on the steering wheel, and action items.
The present project continues to develop an in-car image recognition and analysis framework, which utilizes OpenPose to recognize a dataset of 18 joint points of the driver for deep learning image recognition. The framework infers the driver's hand and steering wheel movements, head movements, hand gestures, and body movements. However, the recognition rate is affected by changes in environmental lighting and imaging effects. The analysis results showed that the recognition rate was lower during periods of low light or poor imaging quality.
Comparing the results of manual verification with image recognition showed that while the significance of various hand and gesture movements varied between the two modes, the majority of parameters had consistent effect directions. Therefore, the developed image recognition
technologies could effectively recognize driver behavior and help detect abnormal events. It is recommended that companies consider technology, specifications, functions, and costs to develop and implement in-car driver behavior image recognition for fleet management and traffic safety.
This project utilized the 2,558 events from the previous phase to analyze the correlation between driver behavior and driving dynamics, as well as abnormal events. The overall correlation between driver behavior, driving dynamics, and abnormal events was then analyzed. The study
found that all driver behaviors while driving, except for operating in-car equipment, was significantly positively correlated with driver distraction. In the event sequence analysis, under different driving conditions, specific actions or action combinations by the driver could lead to abnormal events. For example, when changing lanes in front of the lead vehicle, high-speed approach to the lead vehicle (FCWH) can easily lead to abnormal events. When changing lanes with the host vehicle, being too close to the lead vehicle (SDW0.4) and operating non-lever incar equipment could easily lead to abnormal events. In the case of lane deviation of the host vehicle, actions such as leaning forward, hands leaving the steering wheel, and holding objects that couldn’t control the vehicle properly can easily lead to abnormal events. The results of the spatiotemporal analysis of abnormal events indicated that distracted behavior had a significant positive impact on warning and abnormal events,
from macro to micro perspectives. This means that certain conditions in the spatiotemporal environment make distracted events more likely to occur, and distracted events significantly increase the occurrence of warning and abnormal events. |
