Application of Deep Learning Methods for Pedestrian Collision Detection Using Dashcam Videos.

Area of Science:

• Computer Science
• Artificial Intelligence
• Traffic Safety

Background:

• Pedestrians constitute the majority of traffic fatalities in Japan.
• Advanced automatic collision notification systems require robust detection methods.

Purpose of the Study:

• To evaluate deep learning for detecting pedestrian collisions using dashcam videos.
• To assess the feasibility of deep learning in automatic collision notification systems.

Main Methods:

• A dataset of 1,212 images was created from 78 dashcam videos of pedestrian-automobile accidents.
• Deep learning models were trained on this dataset to identify pedestrian collision features.
• Model performance was evaluated using varied test sets under different conditions.

Main Results:

• High-precision detection achieved for daytime, clear trajectory accidents, including location identification.
• Nighttime conditions and unclear data led to false or missed detections.
• Reduced exposure value and resolution negatively impacted detection rates.

Conclusions:

• Deep learning shows promise for pedestrian collision detection using dashcam videos.
• Further research is needed to improve performance in adverse conditions like nighttime.
• This technology could enhance automatic collision notification systems for improved road safety.