Semantic Segmentation on Smartphone Motion Sensor Data for Road Surface Monitoring

Abstract

Improving road safety is one of the critical issues for road maintenance and management. Motion sensors embedded in smartphones to sense vibrations can be used to detect rough road surfaces when carried in moving vehicles. Finding segments in the signal which reflect the condition of the road surface, however, is a challenging task. This study proposes a modified U-Net architecture with integrated bidirectional Long Short-Term Memory layers to perform semantic segmentation on smartphone motion sensor data for road surface classification. Experiments show that using z-axis accelerometer and z-axis gyroscope features, the proposed method outperforms multiple existing semantic segmentation algorithms.