Recognizing Skateboard and Kickboard Commuting Behaviors Using Activity Trackers: #feasibility Study Using Machine Learning Approaches
Background: Active commuting, such as skateboarding and kickboarding, is gaining popularity as an alternative to traditional modes of transportation like walking and cycling. However, current activity trackers and smartphones, which rely on accelerometer data, are primarily designed to recognize symmetrical locomotive activities (e.g., walking, running) and may struggle to accurately identify the unique push-push-glide motion patterns of skateboarding and kickboarding. Objective: This study utilized machine learning techniques to evaluate the #feasibility of classifying skateboard and kickboard commuting behaviors using data from wearable sensors and smartphones. A secondary objective was to identify the most important sensor-derived features for accurate activity recognition. Methods: Ten participants (4 women, 6 men; aged 12-55 years) performed nine activities, including skateboarding, kickboarding, walking, running, bicycling, ascending and descending stairs, sitting, and standing. Data were collected using wearable sensors (accelerometer, gyroscope, barometer) placed on the wrist, hip, and in the pocket to replicate the sensing characteristics of commercial activity trackers and smartphones. The signal processing approach included the extraction of a total of 211 features from 10- and 20-second sliding windows. Random forest classifiers were trained to perform multi-class and binary classifications, including distinguishing skateboarding and kickboarding from other activities. Results: Wrist-worn sensor configurations achieved the highest balanced accuracies for multi-class classification (84-88%). Skateboarding and kickboarding were recognized with sensitivities of 93-99% and 97-99%, respectively. Hip and pocket sensor configurations showed lower performance, particularly in distinguishing skateboarding (49-58% sensitivity) from kickboarding (78% sensitivity). Binary classification models grouping skateboarding and kickboarding into a "push-push-glide" superclass achieved high accuracies (91-95%). Key features for classification included low- and high-frequency accelerometer signals, as well as roll-pitch-yaw angles. Conclusions: This study demonstrates the #feasibility of recognizing skateboard and kickboard commuting behaviors using wearable sensors, particularly wrist-worn devices. While hip and pocket sensors showed limitations in differentiating these activities, the broader "push-push-glide" classification achieved acceptable accuracy, suggesting its potential for integration into activity tracker software. Future research should explore sensor fusion approaches to further enhance recognition performance and address the question of energy expenditure estimation.
