A Multimodal Personalized Architecture for Irregular Bicycle Riding Form Detection

Due to the recent popularity of sports bicycles, many people have started tourism with cycling. Beginner cyclists often try to complete long distance cycling without sufficient training, causing overuse of leg muscles. Overuse of leg muscles caused by an irregular bicycle riding form leads to diseases such as iliotibial band syndrome. It is helpful to automatically identify beginner cyclists of irregular forms to avoid these diseases. However, none of the multimodal AI models recently attracting attention can achieve this. This paper proposes a system architecture for detecting individually optimized irregular bicycle riding form. Our proposed architecture uses bicycles installed with the Inertial Measurement Unit (IMU) and 2D Light Detection and Ranging (2D LiDAR) sensors. The sensor data obtained from IMU are analyzed by Recurrent Neural Network-based models, and the results are refined by multimodal classifier models combined with the 2D LiDAR data. The model selected by the proposed method obtained 0.900 in accuracy, 0.892 in precision, and 0.841 in recall for an f-score of 0.862, demonstrating the effectiveness of meta-model approaches.