Deep Learning Approach for Detecting Lane Change Maneuvers Using SHRP2 Naturalistic Driving Data

Abstract

Changing lane is a complex driving maneuver and could have a significant effect on traffic safety. Therefore, developing accurate and timely lane change detection systems could assist drivers to perform and complete this complicated driving task safely. This study proposes reliable lane change detection models based on deep learning (DL) and including features from vehicle kinematics, machine vision, roadway geometries, and driver demographics using the trajectory-level SHRP2 Naturalistic Driving Study and Roadway Information Database. A cutting-edge technique named “DeepInsight” was applied to transform numeric lane change data into image data. The generated image data sets were trained, validated, and tested using a novel DL architecture called ResNet-18, considering six categories of features. To balance the lane change database, two data balancing methods—the synthetic minority oversampling technique (SMOTE) and random majority under sampling (RMUS)—were considered and tested with various sampling ratios. In addition, wrapper-based Boruta and eXtreme gradient boosting (XGBoost) algorithms were used to extract relevant features in each category. A recall of 82% and overall accuracy of 77.9% were found using a ratio of 1:1 for the model based on vehicle kinematic features suggesting that the developed model could be used in the absence of other data. However, the best detection performance was observed using the same ratio for a reduced model based on XGBoost, which produced a recall and overall accuracy of 98.8% and 95%, respectively, considering all the features. The proposed detection system could be effectively leveraged to monitor lane change behavior and provide appropriate control strategies in a connected vehicle (CV) environment.