Abstract
We previously performed pedestrian collision detection using a deep learning method based on dashcam video data. However, the detection accuracy was poor owing to insufficient training data. Herein, we attempted to improve the accuracy of the detection for Advanced Automatic Crash Notification System (AACN) using data augmentation, which increases the amount of data by adding artificially generated training data. As a result of comparing the effects of multiple image processing methods on the detection rate, the detection rate increased to 86.85% by adding training data with reduced contrast. This rate was 34.37 points higher than the conventional rate.
