Application of Deep Learning Methods for Pedestrian Collision Detection Using Dashcam Videos 2020-22-0008
The goal of this study is to clarify the usefulness of deep learning methods for pedestrian collision detection using dashcam videos for advanced automatic collision notification, focusing on pedestrians, as they make up the highest number of traffic fatalities in Japan. First, we created a dataset for deep learning from dashcam videos. A total of 78 dashcam videos of pedestrian-to-automobile accidents were collected from a video hosting website and from the Japan Automobile Research Institute (JARI). Individual frames were selected from the video data amounting to a total of 1,212 still images, which were added to our dataset with class and location information. This dataset was then divided to create training, validation, and test datasets. Next, deep learning was performed based on the training dataset to learn the features of pedestrian collision images, which are images that capture pedestrian behavior at the time of the collision. Pedestrian collision detection performance of the trained model was evaluated as the percentage of correct predictions of pedestrian collisions in image data according to varied test sets with different combinations of characteristics. Our results for the proposed method show high-precision collision detection for daytime, clear pedestrian wrap trajectory accident data, including accurate detection of pedestrian collision location information. However, nighttime, unclear accident data resulted in false detection or no detection. Reduction of exposure value and resolution was confirmed to reduce detection rate. The results of the present study suggest the possibility of pedestrian collision detection by deep learning using dashcam videos.