3D Object Detection Method Using LiDAR Information in Multiple Frames

For a safe autonomous navigation, it is important to understand the configuration of the environment and quickly, accurately grasp the information regarding the location, direction, and size of each constituent object. Recent studies on autonomous navigation were performed to not only detect and classify objects, but also to segment and evaluate their properties. However, in these studies, pre-processing was required, which incurred a considerable amount of computational cost. Moreover, the 3D shape model was further analyzed. In other words, more computation cost and computing power are required. In this study, we propose a new method for detecting and estimating the pose of a 3D object using LiDAR information via charge-coupled-device (CCD) in real-time environment. We classified objects into classes (e.g., car, pedestrian, and cyclist), and the 3D pose of an object is quickly estimated without requiring a separate 3D-shape model. From the multiple frames obtained using the LiDAR and CCD, we design a method to robustly reconstruct the 3D environment in real time by aligning the object information of the previously obtained frames with the current frame through an optical-flow method. Our method helps in complementing the limitations of CCD-based classifiers and correcting the defects by increasing the density of the 3D-LiDAR point cloud. We compared the results obtained using our method with the state-of-the-art results of the KITTI data set; which were in good agreement in terms of speed and accuracy. This comparison shows that the 3D pose of a box can be generated with better speed and accuracy using the reconstructed 3D-point-cloud clusters proposed in our method.