(An) empirical driving maneuvers analysis by using selective training data-driven NLPCA-deep learning scheme

Abstract

With the development of the connected car technology, it becomes possible to connect the vehicle with various devices. Therefore, it is possible to provide application services using information collected from various devices. As the market for the connected car grows, demand for such application services is increasing, and especially demand for safety relevant services such as providing the driver assistance is expected to account for the largest portion. Accordingly, various methods of analyzing the driving maneuver using the information that can be collected from the vehicle and assessing the safety such as the degree of the abnormality and the collision risk from the analysis have been studied to provide the safety assistance service to the driver. However, in order to commercialize this, it is necessary to improve the accuracy of the driving maneuver analysis. Moreover, to provide the safety assistance service to the driver practically, not only the safety evaluation but also the corrective instructions for safe driving should be provided to the driver. In this thesis, we are to implement the abnormality analysis service for driving maneuver which can detect and analyze the driving maneuver to evaluate the driving abnormality and provide quantitative factors that the driver should fix for normal driving as a safe driving assistance service. To improve the accuracy of driving maneuver analysis, we firstly proposed a driving maneuver detecting model using the convolutional neural network. In addition, we propose a method that can provide the target values for normal driving by selectively training the nonlinear principal component analysis (NLPCA) - deep learning model using the data evaluated as normal driving according to the detected driving maneuver. Finally, the accuracy of driving maneuver detection using the proposed method was experimentally evaluated, and the validity and accuracy of the proposed model that can provide the target values for normal driving were experimentally evaluated.