WEKO3
アイテム
自動運転のための3次元物体検出フレームワークに関する研究
https://doi.org/10.18997/0002000292
https://doi.org/10.18997/000200029295bcd228-2f37-4f32-ad93-b0e6ccfe4347
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
kou_k_581.pdf (29.6 MB)
|
|
| アイテムタイプ | 学位論文 = Thesis or Dissertation(1) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 公開日 | 2023-11-29 | |||||||||
| 資源タイプ | ||||||||||
| 資源タイプ識別子 | http://purl.org/coar/resource_type/c_db06 | |||||||||
| 資源タイプ | doctoral thesis | |||||||||
| タイトル | ||||||||||
| タイトル | 自動運転のための3次元物体検出フレームワークに関する研究 | |||||||||
| 言語 | ja | |||||||||
| タイトル | ||||||||||
| タイトル | The Research on 3D Object Detection Framework for Autonomous Driving | |||||||||
| 言語 | en | |||||||||
| 言語 | ||||||||||
| 言語 | jpn | |||||||||
| 著者 |
李, 鎮
× 李, 鎮
|
|||||||||
| 抄録 | ||||||||||
| 内容記述タイプ | Abstract | |||||||||
| 内容記述 | In the last ten years, the commercial use of electric vehicles has gradually promoted the evolution of traditional vehicles development. The most crucial evolution direction is to make vehicles be intelligent, possess a human-like perception ability to the around environment. As reported by the U.S. Department of Transportation, more than 90% of car crashes are attributed to drivers’ errors. Therefore, the demand for safe driving is another main reason to facilitate the evolution of intelligent driving. 3D object detection is a critical intelligent driving technology that provides precise 3D scale, location and pose information of surrounding objects in real time. 3D detection of around objects and landform could make vehicles to realize the applications of navigation and obstacle avoidance in the transportation situations. With a high-performance 3D detector, intelligent driving could have the possibility to solving the most significant problems which are occurred by the traditional vehicles, such as driving safety, fatigue driving and transportation efficiency. Intelligent driving also can set people free from the heavy driving burden. Therefore, an accurate and efficient 3D object detection system embedded in intelligent vehicles is an indispensable technique for future human life. Autonomous driving has received enormous attention from academic and industrial communities. However, achieving full driving autonomy is not a trivial task, because of the complex and dynamic driving environment. Perception ability is a tough challenge for autonomous driving, while 3D object detection serves as a breakthrough for providing precise and dependable 3D geometric information. Inspired by practical driving experiences of human experts, pure visual scheme takes sufficient responsibility for safe and stable autonomous driving. In this paper, we proposed an anchor-free and keypoint-based 3D object detector with monocular vision, namely Keypoint3D. We creatively leveraged 2D projected points from 3D objectsʻ geometric centers as keypoints for object modeling. Additionally, for precise keypoints positioning, we utilized a novel self-adapting ellipse Gaussian filter (saEGF) on heatmaps, considering different objects’ shapes. We tried different variations of DLA-34 backbone and proposed a semi-aggregation DLA-34 (SADLA-34) network, which pruned the redundant aggregation branch but achieved better performance. Keypoint3D regressing the yaw angle in a Euclidean space, resulted in a closed mathematical space avoiding singularities. Numerous experiments on the KITTI dataset for moderate level have proven that Keypoint3D achieved the best speed-accuracy trade-off with an average precision of 39.1% at 18.9 FPS on 3D cars detection. With the continuous evolution of traditional automobile industry, autonomous driving has become an indispensable part of intelligent vehicles in the contemporary era. Three-dimensional (3D) object detection algorithms could provide precise 3D scales, locations and poses of surrounding objects. Therefore, an efficient and accurate 3D object detector is crucial for autonomous driving. Nevertheless, high precision 3D detectors generally consume overmuch inference time, unable to fulfill the tasks of critical real-time requirement. Existing 3D detectors demand for much more computations than 2D tasks for 3D bounding boxes regression. Hence, this paper proposed a single-stage, LiDAR-based, and efficiency-accuracy balanced 3D object detector with six Degree-of-Freedom, called 6DoF-3D. Our network structure consists of a backbone of CSPDarknet53 [?], a multi-scale features fusion neck, and a detection head using the proposed 2.5D Euler-Region-Proposal network (ERPN). For the lightweight network architecture, 6DoF-3D designed an entirely novel regression strategy that pruned redundant Degree-of-Freedom of 3D detection to reduce the complexity of 3D tasks. We designed a geometric transformation module to make up the lost Degree-of-Freedom to ensure detection accuracy without extra computation. Extensive experiments on KITTI data set prove that 6DoF-3D could achieve the state-of-the-art accuracy of 89.62% for 3D car detection at the moderate level. The mean average precision could achieve 66.72% at the real-time performance of 33.21 frames per second (FPS) for cars, pedestrians and cyclists. We leveraged two different types of data information and presented two 3D object detectors. The proposed two 3D object detectors achieved the comparable 3D object detection accuracy with other outstanding methods. Nevertheless, our proposed methods attached greater importance on real time inference speed for high efficiency demanding autonomous driving task. Both of our proposed methods have an obvious inference efficiency improvement and can basically realize real scenes application. For our future work, we plan to make the data fusion 3D object detection work, and further improve the detection accuracy based on our excellent inference efficiency. |
|||||||||
| 言語 | en | |||||||||
| 目次 | ||||||||||
| 内容記述タイプ | TableOfContents | |||||||||
| 内容記述 | 1 Research Background|2 The introduction of 3D object detection algorithms|3 3D Object Detection Using Monocular Vision|4 3D object detection using LiDAR device|5 Conclusion|6 Future work | |||||||||
| 言語 | en | |||||||||
| 備考 | ||||||||||
| 内容記述タイプ | Other | |||||||||
| 内容記述 | 九州工業大学博士学位論文 学位記番号：工博甲第581号 学位授与年月日: 令和5年9月25日 | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | Autonomous Driving | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | 3D Object Detection | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | Degree of Freedom | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | Gaussian Filter | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | Backbone Network | |||||||||
| 学位授与番号 | ||||||||||
| 学位授与番号 | 甲第581号 | |||||||||
| 学位名 | ||||||||||
| 学位名 | 博士（工学） | |||||||||
| 学位授与年月日 | ||||||||||
| 学位授与年月日 | 2023-09-25 | |||||||||
| 学位授与機関 | ||||||||||
| 学位授与機関識別子Scheme | kakenhi | |||||||||
| 学位授与機関識別子 | 17104 | |||||||||
| 学位授与機関名 | 九州工業大学 | |||||||||
| 学位授与年度 | ||||||||||
| 内容記述タイプ | Other | |||||||||
| 内容記述 | 令和5年度 | |||||||||
| 出版タイプ | ||||||||||
| 出版タイプ | VoR | |||||||||
| 出版タイプResource | http://purl.org/coar/version/c_970fb48d4fbd8a85 | |||||||||
| アクセス権 | ||||||||||
| アクセス権 | open access | |||||||||
| アクセス権URI | http://purl.org/coar/access_right/c_abf2 | |||||||||
| ID登録 | ||||||||||
| ID登録 | 10.18997/0002000292 | |||||||||
| ID登録タイプ | JaLC | |||||||||
