A deep learning approach for intelligent cockpits: learning drivers routines

Abstract

Nowadays an increasing number of vehicles are being equipped with powerful cockpit systems capable of collecting drivers’ footprints over time. The collection of this valuable data opens effective opportunities for routine prediction. With the growing ability of vehicles to collect spatial and temporal information solving the routine prediction problem becomes crucial and feasible. It is then extremely important to advance and take advantage of the capabilities of these cockpit systems. A vehicle that is capable of predicting the next destination of the driver and when the driver intends to leave to that destination can prepare the journey in advance. Previous studies tackling the next location prediction problem have made use of Traditional Markov models, Neural Networks, Dynamic models, among others. In this work, a framework based on the hierarchical density-based clustering algorithm followed by a Long Short-Term Memory (LSTM) recurrent neural network is proposed for spatial-temporal prediction of drivers’ routines. Based on real-life driving scenarios of three different users, the proposed approach achieved a test set accuracy of 96.20%, 90.23%, and 86.40% when predicting the next destination and a R2 Score of 93.69, 79.21, and 28.81 when predicting the departure time, respectively. The results indicate that the proposed architecture can be implemented on the vehicle cockpit for the assistance of the management of future trips.