Driving Style Profiling Using Deep Autoencoders for Safety Applications in Urban and Highway Scenarios

Driver behavior profiling and analysis play a significant role in road safety and security applications. With enormous data being collected from sensors installed in vehicles, this work proposes an approach to create driving profiles for each user based on the data collected from driving a specific vehicle. The paper investigates the utilization of autoencoders known for learning patterns from data in creating benchmark models capable of detecting irregularities in driving patterns different from those on which the benchmark model was trained. Specifically, this research involves developing and analyzing deep learning autoencoders, such as Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), which can learn temporal dependencies and unique features of various users. By constructing benchmark models, specifically driving profiles for each user, we assess the similarity and irregularity between the user on whom the model was trained and another user who traversed the same path using the same vehicle. This proposed technique of developing driving profiles holds potential for application in safety contexts, regardless of the types of users and in different road scenarios. The developed models were experimented on two datasets. The first dataset comprises data collected from bike riders riding along the same path in an urban scenario. The second dataset comprises data collected from car drivers driving on a motorway. Experimental results reveal an effective application of our proposed approach in both urban and motorway settings, and our proposed approach could potentially be utilized in numerous vehicular applications.

Перевод пока недоступен