Behavior Prediction Using 3D Box Estimation in Road Environment

Abstract

Autonomous vehicle technology will make possibility of significant benefits to social welfare such as reducing traffic casualties, assisting the mobility of the elderly, and reducing the burden of driving. Among them, collision prediction and avoidance system are especially important topics in real road scenes. In order to realize a collision avoidance system, it is necessary to accurately grasp the surrounding environment of the self-location and predict the behavior of the target. Camera information or Light Detection and Ranging (LiDAR) information are used for behavior prediction. However, LiDAR is impractical due to its high cost. For camera information, the accuracy is lower than that of LiDAR, however there is a possibility that the accuracy can be compensated by introducing machine learning that has been developing in recent years. In this study, we investigate the usefulness of 3D box estimation in behavior prediction using camera information. In 3D box estimation, the dimensions and orientation of the target in the 2D box are regressed using a CNN model. We use MultiBin loss when we regress orientation. And we estimate final 3Dbox parameters based on regression values. Finally, we predict the trajectory using the center of the box and four vertices as inputs, and we verify its usefulness.