Object recognition from on-the-road traffic data

Abstract

In order to detect the object and inspect the road conditions in real-time, the 2-dimensional (2D) and 3dimensional (3D) data obtained from the onboard sensors, LiDAR and digital cameras are analyzed for object recognition to assist driving. Due to the uncertainties of the dynamic objects, such as pedestrians, animals or vibrated vehicles, extraction of complete and clear objects from LiDARs datasets requires complex post-processing since LiDAR data can be used for scanning at long distances, i.e., 300m, which can alarm the driver timely to take necessary actions. The dynamic and static objects from the LiDARs point clouds can be detected with the teacher-student framework algorithm along with the KITTI dataset. Furthermore, a semi-supervised theory is utilized to improve detection performance.