Road safety analysis framework based on vehicle vibrations and sounds using deep learning techniques

Abstract

Road accidents in India occur due to potholes. These potholes are not repaired because the authorities will not be aware of it unless the public raises an issue. Lack of adequate techniques to identify potholes has caused huge trouble to the public. The primary goal of this study is to build a deep learning model that would analyze the patterns in the sound recording of the vehicles and label the Road Anomaly Events (RAEs). Deep learning techniques like Convolutional Neural Network (CNN) and BLSTM are used to classify the sound signature and then are labelled accordingly. The idea can be implemented in areas where there is regular movement of vehicles to identify the exact locations of the pothole and inform the concerned authorities so that the public can experience smoother roads. From the analysis, it is found that the model has an accuracy of 83% with ADAM Optimizer while RMSProp produces 54-60% accuracy.