2023 Volume 79 Issue 20 Article ID: 23-20047
In recent years, AI-based image analysis methods have been introduced in automobile traffic volume measurement, establishing a system for automatic observation. However, the measurement accuracy for bicycles and pedestrians remains low, prompting high expectations for the development of multi-purpose traffic measurement AI capable of acquiring not only traffic volume but also movement speed, movement trajectory, and other information. This study aims to develop a multi-purpose traffic measurement AI that can estimate traffic volume and vehicle speed for bicycles and pedestrians. Specifically, we trained a convolutional neural network (CNN) using new training labels and leveraged information obtained from object tracking to facilitate traffic volume and speed estimation. Additionally, we proposed a self-learning method for each location to address differences in measurement accuracy among locations, and validated the efficacy of learning synthetic images using the background images of the observation locations.
