Estimating Vehicle Speed on Highway Roads from Smartphone Sensors Using Deep Learning Models

Speed estimation is an open research area because of its importance in many applications and the necessity of replacing GPS due to smartphones battery drainage. Relying on integrating accelerometer values is challenging and generally requires continual utilization of external references to correct speed, because of error accumulation. Therefore, highway speed estimation, in particular, is a difficult problem due to the maintenance of high speeds by vehicles for a long time and the scarcity of reference points like uneven road surfaces, turns, and stops. In this paper, we investigate exploiting micro road surface unevenness that results in vibrations on the accelerometer readings without integrating the acceleration signal. In particular, we employ deep 1D convolutional neural networks to learn and extract robust features that learn the relation between such complex vibrations and speed. Also, the use of bidirectional LSTMs is investigated to benefit from both forward and backward dependencies in the sensed data, and allow a form of integration. Specifically, two highway speed estimation models are proposed. The first uses a deep convolutional neural network with 5 layers and the second uses a deep bidirectional LSTM neural network. The inputs to both networks are the readings from the accelerometer and gyroscope sensors of a smartphone. The methods achieved mean absolute error results of 5.53 km/hr and 3.71 km/hr, respectively; whereas a related LSTM based method, resulted in a high error rate of 68.05 km/hr. Finally, an implementation of the proposed CNN model on an android and iOS smartphones is described and analyzed.