Probability-Based Bounding Box With Confident Object Localization for Autonomous Navigation

Abstract The Object detection, especially for autonomous navigation, requires high accuracy of the localization accuracy on the objects detected. The bounding box is devised to provide the localization of the object detection. However, the confidence score of the bounding boxes does not provide the localization reliability although the bounding boxes are selected by the confidence score through a non-maximum suppression (NMS) task. In this paper, we propose a novel bounding box regression framework to build a probability-based bounding box. The proposed bounding box leverages the spatial and confidence information of the candidate boxes removed by the NMS task. The information enables to capture the object entirely through expanding the localization extent of the tradition bounding boxes reliably, where the extent is asymmetrically expansion on each side to avoids unnecessary localization. In addition, the regression task using the information of the candidate boxes allow us to the extent of the bounding box for desire confidence level. The proposed method is evaluated on sample images from Berkery DeepDrive (BDD) dataset, where the samples include objects only partially detected by the traditional bounding box. Our experiment presents that the probability-based bounding box would provide more reliable localization compared to the tradition bounding box.