Facial Expression Analysis from Upward-view Images for Comfortable Bicycle Riding

In urban settings worldwide, the advent of new mobility services presents both opportunities and challenges for regulatory bodies and established operators. Improving the cycling experience requires a deeper insight into riders' emotional states throughout their journey. This paper presents a novel method for analyzing drivers' facial expressions captured from low-angle cameras, particularly useful in situations where aligning the camera directly with the driver is impractical, such as on two-wheeled vehicles. Traditional sentiment analysis techniques, reliant on front-facing camera footage and direct pixel value application, demonstrate reduced accuracy when applied to upward-view images. In contrast, our study focuses on extracting facial features and utilizing their positional information. Facial features are obtained using the Practical Facial Landmark Detector (PFLD), resulting in 67 coordinate values subsequently trained by a neural network. The labeling framework employs binary classification, distinguishing between comfortable or not comfortable states. To evaluate the effectiveness of our method, we conducted comparative assessments under various conditions, including different data partitioning for training and testing, as well as diverse camera installation angles. Results indicate that our approach achieves superior accuracy, with an F-score of 0.90 under consistent conditions. For varying individuals within the training dataset, the F-score reached 0.51, while training and testing across multiple individuals yielded a notable F-score of 0.94.