Two-level distribution alignment for unsupervised domain adaptation of LiDAR-based object detector

Abstract

In autonomous driving, object detection technology is an essential part of recognizing the driving environment, and a high-performance deep learning-based method attracts significant interest. Thus, a vast amount of learning data is essential, and the demand for quality data is also increasing. However, data suitable for various sensor configurations and the purpose of use cannot meet this demand. An adaptation method based on transfer learning that enables heterogeneous data was proposed to solve this problem. In the case of LiDAR, many changes occur in the distribution of the point cloud created according to the number of channels, mounting positions, and beam angles of the product. Due to this change, the training data suitable for the user's LiDAR configuration is more insufficient. Also, in point clouds, there is no standard large scale data set as an image domain, feature vector extraction method, and transfer learning is hardly attempted. Therefore, in this study, we aim to develop a transfer learning method for LiDAR object detection. The developed method is evaluated with two public datasets. Through this, it is expected that a deep learning model can be trained with sufficiently existing heterogeneous data and then adapted with a small amount of homogeneous data to develop an object detection algorithm.