Real-Time 3D-LiDAR Object Detection in Autonomous Vehicle Systems Using Cluster-Based Candidates and Deep Learning

Recently, IT companies such as Google, NVIDIA, and NAVER have been also developing autonomous vehicle platform technologies. In particular, sensors for object detection in surrounding environments have been improved in recognition rates by applying multi-sensor systems using camera, LiDAR, and radar. With the increasing importance of recognition technology, 3D information-based recognition technologies have been actively advanced as a commercial product of 3D-LiDAR. In this paper, a candidate group of point-clouds from 3D-LiDAR is extracted using Euclidean clustering in order reduce the processing time delay in RPN (Region Proposal Network), which is one of the basic schemes for existing object detection. Then, it proposes types of input slicing, based on the extracted candidates. In addition, the accuracy and the processing time using four CNN networks (Basic CNN, ResNet, VGG16, and MobileNet) are compared over not only the private data(CVLab dataset) obtained in actual road environment but also the publicly open KITTI dataset.