[3N3-IS-2e] Machine learning (5/5)

[3N3-IS-2e-03] Motion Classification for Automatic Driving

〇Takuma Hayakawa¹ (1. Nagoya Institute of Technology)

Keywords:Motion Classification, Automatic Driving

In recent years, automated automobile driving technology has been attracting attention in Japan. Although the technology of automated driving is steadily advancing, there are various problems that hinder the realization of automated driving. Among them is the problem that automated vehicles brake when they recognize pedestrians, but they also brake when pedestrians are trying to yield the right of way, resulting in a situation where both the vehicle and the pedestrian remain stationary without proceeding.
In this study, we focus on the motion of pedestrians and classify the motion of pedestrians based on the input image alone, without using any equipment such as motion captures. Among the pedestrian motions, we classify them into four types: walking, stopping, giving way, and stopping the car. If it is possible to recognize the gesture of a pedestrian giving way, the problem of giving way between a self-driving car and a pedestrian can be solved.
We found that movements can also be classified by object recognition techniques. However, it is difficult to classify continuously changing motions (e.g., "walking") with high performance even though poses can be classified with high performance.