Joint deep neural network modelling and statistical analysis on characterizing driving behaviors

Abstract

Google defines the concept of autonomous driving as one of the applications of big data. Specifically, with the input sensor data, the autonomous vehicles can be provided with the semantic-level driving characteristics for an accurate and safe driving control. However, both the enumeration of handcrafted driving features with expert knowledge and the feature classification with machine learning for characterizing driving behaviors is lack of practicability under a complex scale. Therefore, this study focuses on detecting the sematic-level driving behaviors from large-scale GPS sensor data. Specifically, we classified different driving maneuvers from a huge amount of dataset through a layer-by-layer statistical analysis method. The identified maneuver information with the corresponding driver ID is useful for the supervised learning of high-level feature abstraction with neural network. With the aim of analyzing the sensory data with deep learning in a consumable form, we propose a joint histogram feature map to regularize the shallow features in this paper. Besides, extensive simulation is conducted to evaluate different machine learning and deep learning methodologies for optimal driving behavior characterization. Overall, our results indicate that Deep Neural Network (DNN) is suitable for the driving maneuver classification task with more than 94% accuracy, while Long Short-term Memory (LSTM) neural network performs well with a 92% accuracy in identifying a specific driver. However, LSTM shows degraded accuracy when the scale of the identification task becomes larger. In this case, a hierarchical deep learning model is proposed, and simulation results show that the combination of DNN and LSTM in this hierarchical model can well maintain the prediction accuracy even when the scale of the recognition task is four times larger.