Leveraging Saliency Prediction to Enhance Behavioral Decision-Making for Parked Vehicle Avoidance on Two-Lane Two-Way Roads

Abstract

One of the most challenging collision avoidance scenarios is to avoid parked vehicles on Two-Lane Two-Way (TLTW) because of dynamic scenarios where divers must evade the parked vehicles while remaining alert of oncoming vehicles. Traditional behavioral decision-making systems for automated vehicles typically rely on processed image data. While these systems have made significant advancements, they often struggle to effectively prioritize and process the most relevant aspects of the scene, such as parked vehicles, oncoming vehicles, and pedestrians. This limitation can impact the overall safety and efficiency of collision avoidance. Saliency prediction offers a promising solution by identifying and highlighting the most important regions in a visual scene. By focusing on these salient features, saliency prediction can enhance the decision-making process by emphasizing what is most relevant in a given context. This study proposes a novel decision-making framework that utilizes saliency prediction to improve decision-making performance. Additionally, this study is aimed at developing a novel saliency prediction model for autonomous driving: Weighted Salient Object Detection (WSOD). WSOD gives varying importance levels to salient objects within a scene. This approach is expected to enhance decision-making performance by prioritizing salient objects.