Does LSTM outperform 4DDTW-KNN in lane change identification based on eye gaze data?

Abstract

Lane change identification based on long short-term memory (LSTM) neural networks has received increasing attention. The most commonly known disadvantage of the approach is the “black box” nature. This study, to improve the interpretability, proposed a novel method for identifying lane changes. The philosophy was to measure the similarity of the glance behavior during the maneuver. The glance behavior in the same maneuver is more similar, while that in different maneuvers is different. First, a driving simulator was used to collect driving behavior data. The eye gaze time series were captured by the eye tracker. The median of the eye gaze sequence of 3 s before the initial moment of lane change was obtained. With the median as the center, the driver's visual field plane was divided into several grids with the side length of 600 pixels. A sliding space-time cuboid algorithm was proposed to extract the scanning path. Second, four-dimensional dynamic time warping (4DDTW) distance was used to measure the similarity of the scanning paths. Third, the K nearest neighbor (KNN) algorithm was used to classify the driving maneuver into the lane-keeping (LK), the right lane change (RLC), and the left lane change (LLC) based on the 4DDTW distance of the scan paths. LSTM was also used to classify the driving maneuver into LK, RLC, and LLC. The performances of 4DDTW-KNN and LSTM were compared. The accuracies of 4DDTW-KNN and LSTM were 86.50% and 86.33%, respectively. LSTM is not better than 4DDTW-KNN in lane change identification based on eye gaze data with a comprehensive consideration of the time efficiency, the accuracy and the interpretability.