Multi-Level Feature Extraction and Classification for Lane Changing Behavior Prediction and POD-Based Evaluation

Abstract

Lane changing behavior (LCB) prediction is a crucial functionality of advanced driver-assistance systems and autonomous vehicles. Predicting whether or not the driver of a considered ego vehicle is likely to change lanes in the near future plays an important role in improving road safety and traffic efficiency. Understanding the underlying intentions behind the driver’s behavior is an important factor for the effectiveness of assistance and monitoring systems. Machine learning (ML) algorithms have been broadly used to predict this behavior by analyzing datasets of traffic and driving data related to the considered ego vehicle. However, this technology has not yet been widely adopted in commercial products. Further improvements in these algorithms are necessary to enhance their robustness and reliability. In some domains, receiver operating characteristic and precision-recall curves are commonly used to evaluate ML algorithms, not considering the effects of process parameters in the evaluation, while it might be necessary to access the performance of these algorithms with respect to such parameters. This paper proposes the use of deep autoencoders to extract multi-level features from datasets, which can then be used to train an ensemble of classifiers. This allows for taking advantage of high feature-extraction capabilities of deep learning models and improving the final result using ensemble learning techniques. The concept of probability of detection is used in combination with the networks employed here to evaluate which classifiers can detect the correct LCB better in a statistical sense. Applications on data acquired from a driving simulator show that the proposed method can be adopted to improve the reliability of the classifiers, and ensemble ANNs perform best in predicting the upcoming human behavior in this dynamical context earlier than 3 s before the event itself.