PEDESTRIAN CROSSING INTENTION PREDICTION IN UNSIGNALIZED CROSSING ENVIRONMENT

Abstract

To ensure safe and secure coexistence between pedestrians and autonomous vehicles (AVs), AVs must be able to anticipate pedestrian behavior and respond to it. This research gathers video data from real traffic scenes to predict pedestrian crossing intentions at unsignalized crossings. Computer vision techniques such as YOLOv4, DeepSORT, and perspective transformation are employed for road user detection, tracking, and mapping image coordinates to world coordinates to prepare trajectory datasets. Using trajectory data, several features influencing pedestrian intention like walking speed, location in the road environment, count and direction of approaching traffic, speed and type of closest approaching vehicle upstream, etc., are extracted. The dataset for this study was obtained by analyzing 1,411 pedestrians, resulting in 223,136 samples. To predict pedestrian crossing intentions Bi-LSTM with an attention mechanism model was built and trained to anticipate the pedestrian crossing intention at unsignalized crossing. The Bi-LSTM with an attention mechanism model outperforms the LSTM model by successfully combining the characteristics and surrounding dynamics of pedestrians to produce accurate predictions, achieving an AUC of 95.3%, 91.1%, 89.2%, 87.5%, and 88% on the testing dataset at unsignalized crossing on the 0.6 sec, 1.2 s, 1.8 s, 2.4 s, and 3 s time horizons. These outcomes can be used to improve Connected and Autonomous Vehicle (CAV) technologies, infrastructure-to-vehicle (I2V) connectivity, and driver assistance systems to enhance pedestrian safety while navigating through pedestrian crosswalks.