🤖 AI Summary
Accurate vehicle speed estimation for mobile robots and autonomous ground vehicles using only low-cost triaxial accelerometers—such as those embedded in smartphones—remains challenging, especially without gyroscopes, wheel odometry, vehicle bus data, or external positioning signals.
Method: This paper proposes an end-to-end deep neural network that operates exclusively on sliding-windowed temporal accelerometer measurements. It incorporates long-range temporal modeling to enhance robustness in dynamic driving scenarios and directly learns the nonlinear mapping from acceleration sequences to instantaneous speed.
Contribution/Results: To our knowledge, this is the first work achieving sub-meter-per-second speed estimation (mean absolute error of 0.72 m/s) using smartphone-grade accelerometers alone, validated over 13 hours of real-world road testing across diverse conditions. The method delivers high-output frequency (10–100 Hz), significantly surpassing GPS (1 Hz), and runs entirely offline—requiring no vehicle interfaces or real-time external assistance.
📝 Abstract
We introduce the CarSpeedNet, a deep learning model designed to estimate car speed using three-axis accelerometer data from smartphones. Using 13 hours of data collected from a smartphone in cars across various roads, CarSpeedNet accurately models the relationship between smartphone acceleration and car speed. Ground truth speed data was collected at 1 [Hz] from GPS receivers. The model provides high-frequency speed estimation by incorporating historical data and achieves a precision of less than 0.72 [m/s] during extended driving tests, relying solely on smartphone accelerometer data without any connection to the car.