Road Surface Recognition Based on DeepSense Neural Network using Time Series Accelerometer Data
Nowadays, Smart devices plays a big role in our lives, especially in our daily activities. Therefore, Smartphones can be considered as one of the most interesting sensor for depicting our activities and our surroundings. Furthermore, the computation power of smartphones has increased a lot recently as most of them have multiple sensors like accelerometers and gyroscopes. Besides, They are capable of processing more tasks than we ever imagined. Because of their advantages of convenience and low-cost, the portable computation platforms has been adopted in the development of autonomous vehicles. The most critical issue of the intelligent system assisted vehicles is that the safety problem. The recognition of the road surface is one of the components to ensure the safety drive. Most of the solutions use sensor fusion to recognize road surfaces such as combining cameras and LiDARs, which is costly for equipment and they usually need installations to re-equip existing cars, but these methods provide overall excellent results. This thesis proposes a method for recognizing the road surface based on using accelerometer data collected from smartphone. The process uses time series data collected from a smartphone’s accelerometer, followed by a massive time series feature extraction and selection. After that, the features are fed into trained DeepSense variant neural network framework to get the recognition of the road surfaces. The proposed method provides three classes recognition for smooth, bumpy and rough roads. Moreover, in this thesis we conducted a thorough evaluation and analysis of the proposed method by comparing it with conventional machine learning methods like SVM, random forest, fully connected neural network and convolutional neural network. The accuracy of the method in this thesis overmatch the compared examples. The road surface type will be classified into three categories which will indicate smoothness of the road surface.
The following license files are associated with this item: