CNN-based System to Identify Bicycle Riders and Pedestrians: Toward Minor Collision Prevention on Sidewalks

In recent years, minor bicycle collisions on sidewalks have become a critical issue. To prevent them, various researches on person and bicycle detection have been conducted. However, existing methods cannot distinguish between bicycle riders and bicycle pushers because they rely on either the bicycle's shape or people's shape for detection. Therefore, we propose a CNN-based system that uses video frame images to identify not only walkers and bicycle riders, but also bicycle pushers, who have a shape similar to that of bicycle riders. Our system learned the features of 15,000 images of bicycle riders, bicycle pushers, and walkers and can automatically detect and identify the bicycle riders, bicycle pushers, and walkers in each frame of the input video. In an evaluation experiment, we recorded a video of pedestrians and bicycle riders on public roads to assess the bicycle rider identification rate of the proposed system. The bicycle rider identification rate achieved using the proposed method was 80.7%, which illustrates the effectiveness of our CNN-based identification approach.