실시간 라이다 객체 탐지를 위한 인코더 모델 개선

Abstract

LiDAR has the advantage of providing accurate distance information between the sensor and objects, which has led to extensive research utilizing LiDAR in the field of object detection. Point cloud, which is a data format acquired by LiDAR sensor, is not structured data like a tensor but unstructured data in the form of a set of points, making it necessary to represent a point cloud into a tensor, such as voxels and pillars, through an encoder to leverage Convolutional Neural networks (CNNs). While much research in LiDAR object detection has focused on improving performance by modifying CNN backbones or head networks, little attention has been given to improving the encoder. In contrast to existing approaches, we propose an enhanced encoder that more efficiently embeds point cloud data into CNN networks and improve object detection performance. Our proposed encoder demonstrates notable improvements in Average Precision (AP) on the KITTI dataset, achieving this with fewer parameters and computational overhead, while also operating in real-time.