Feature Selection Based on Variance Distribution of Power Spectral Density for Driving Behavior Recognition

Abstract

Abnormal driving detection and recognition is a crucial area of research towards achieving safety in intelligent transportation systems (ITS). In this study, we propose a feature extraction approach and use the extracted features to train a deep learning model that is used for abnormal driving behavior recognition. The proposed approach derives the features based on variances calculated from each frequency bin containing the power spectrum data that is generated using the short time fourier transform. A subset of features is further selected based on variance similarity of the power spectral data. Similarity is realized by finding intersecting variance data of different variance samples obtained from defined data segments of a given driving behavior class. The driving behaviors considered are weaving, sudden braking and normal driving. Experiments were performed using an artificial neural network to test the efficiency of the proposed feature extraction approach. Results show that an accuracy of 91.0% can be achieved with accelerometer data. The accuracy is further improved to 96.1% by combining accelerometer with gyroscope data.