Implementation and Evaluation of an Enhanced Intention Prediction Algorithm for Lane-Changing Scenarios on Highway Roads

Abstract

For an autonomous vehicle driving on a public road, the safety of the passengers and the efficiency of the trip taken are prioritized causing the main function of the autonomous vehicle to be interpreting and inferring the intention of surrounding vehicles, and warning the driver accordingly. Recent Advanced Driving Assistance Systems (ADAS) are capable of and usually limited to, support features like forward-collision warnings, alerting the driver of hazardous road conditions, detecting road markings, and warning the driver if they are changing lanes. However, modern ADAS are still unable to perform basic vehicle-behavior-prediction humans are capable of. In this paper, we introduce and compare the results of two different methodologies, Recurrent Neural Networks (RNN) and Long Short-Term Memory networks (LSTM), for lane-changing intention prediction of surrounding vehicles. For the LSTM model, the F1-score achieved was 0.944 for lane-keeping, 0.781 for left lane-changing, and 0.942 for right lane-changing. The RNN-based model reached an F1-score of 0.704 for lane-keeping, 0.533 for left lane-changing, and 0.714 for right lane-changing. The training process of these data-driven based methodologies can be implemented using sequences of changing centroids of vehicles along with the frames and labeling of the maneuvers introduced by the PREVENTION dataset.