Be the first to like this
Smart phones are equipped with many sensors which provide detailed and continuous information of the device's location and movement. The use of such signals for vehicle movement inference presents many challenges due to signal noise, unknown phone orientation, varying device sensor quality and so on. Signal processing and feature engineering are generally difficult and require deep domain knowledge and manual pattern recognition. We discuss how deep learning can be leveraged in this context for automatic signal processing and feature engineering. We present several applications of deep learning in vehicle telematics as well as the deep learning architecture designed for learning sensor embeddings for vehicle movement events. One challenge we face is that model training requires huge volumes of sensor data, which must be processed efficiently. We present a solution using Spark for model development and batch deployment.