Graph Attention을 적용한 라이다 물체 인식 시스템LiDAR Object Detection Using Graph Attention Network
- Other Titles
- LiDAR Object Detection Using Graph Attention Network
- Authors
- 남택규; 손혁주; 허건수
- Issue Date
- Jun-2021
- Publisher
- 한국자동차공학회
- Keywords
- Object detection; LiDAR; Attention; Point Cloud; Graph Neural Network
- Citation
- 2021 한국자동차공학회 춘계학술대회, pp.504 - 508
- Indexed
- OTHER
- Journal Title
- 2021 한국자동차공학회 춘계학술대회
- Start Page
- 504
- End Page
- 508
- URI
- https://scholarworks.bwise.kr/hanyang/handle/2021.sw.hanyang/191364
- Abstract
- In order to perform fully autonomous driving, it is essential to recognize surrounding objects. With the recentdevelopment of deep learning, various studies are being conducted to classify objects around vehicles. In object recognition based on deep learning, there are a point cloud-based model1 and an image-based model2. The image-based model has a sensitive limitation to the light and weather of the camera sensor. On the other hand, since the LiDAR sensor is relatively less affected by light and weather, the reliability of the point cloud-based model is higher. However, the data in the point cloud has a non-uniform distribution. This causes information loss in the convolution process and uses a method of processing a point cloud as a graph to effectively process it. In this paper, after making a graph using a point cloud as a node, we apply a Graph Attention Network
(GAT) that uses an attention mechanism to learn by assigning weights to important nodes to encode the characteristics of each node. This information is used to classify the object and predict the bounding box through two Multi-Layer Perceptrons (MLPs). As a result, the loss of node characteristic information is prevented by giving weight to the information of the target node and theneighboring nodes that are highly relevant.
- Files in This Item
-
Go to Link
- Appears in
Collections - 서울 공과대학 > 서울 미래자동차공학과 > 1. Journal Articles
Items in ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.