WEKO3
アイテム
Visual Transformerに基づく自動運転のための3次元物体検出に関する研究
https://doi.org/10.18997/0002000289
https://doi.org/10.18997/0002000289dcd3d80a-623d-4652-8711-7cd462ffaa31
| 名前 / ファイル | ライセンス | アクション |
|---|---|---|
kou_k_578.pdf (7.3 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 | Research on Visual Transformer-based 3D Object Detection Algorithm for Autonomous Driving | |||||||||
| 言語 | en | |||||||||
| タイトル | ||||||||||
| タイトル | Visual Transformerに基づく自動運転のための3次元物体検出に関する研究 | |||||||||
| 言語 | ja | |||||||||
| 言語 | ||||||||||
| 言語 | eng | |||||||||
| 著者 |
楊, 朔
× 楊, 朔
|
|||||||||
| 抄録 | ||||||||||
| 内容記述タイプ | Abstract | |||||||||
| 内容記述 | 3D object detection is essential in the field of autonomous driving technology. In this field, it is important to develop algorithms that can complete object detection on the road (such as pedestrians, vehicles, bicycles, etc.) in real-time and accurately. This is essential for ensuring the safe driving, path planning, and collision evasion in autonomous vehicles. In essence, autonomous vehicles must have an accurate perception of their environment for safe and reliable operation. 3D object detection provides vital environmental perception data, including the classification, location, dimensions, and positioning of objects. By facilitating the creation of high-accuracy maps, precise 3D object detection assists in exact positioning of autonomous vehicles. Dynamic object information with the map aids autonomous vehicles in planning and decision-making under complex road conditions. Typically, autonomous vehicles employ multiple sensors, such as LiDAR, RGB cameras, etc., to gather various types of data. The fusion of sensor data enhances the accuracy and robustness of object detection. Furthermore, 3D object detection algorithms must offer low latency and high-speed image processing as autonomous driving systems operate in real-time environments. Therefore, developing a high-accuracy, efficient 3D object detection method is of great practical significance in autonomous driving. With regard to 3D object detection algorithms, traditional object detection methods encounter challenges in practical applications due to their limited data processing capabilities and low generalization. Deep learning-based object detection algorithms have gradually become an important direction in 3D vision research, as they have advantages in handling large data sets and possess strong model learning capabilities, effectively addressing data processing challenges in autonomous driving scenarios. However, the existing deep learning-based detection algorithms exhibit problems such as high data processing complexity, disorganized feature extraction network interfaces, and inefficient operation. As a result, the application of most deep learning-based detection algorithms remains difficult in autonomous driving scenarios. Additionally, LiDAR-captured point cloud data provides a more vivid representation of the position, texture, shape, and movement characteristics of 3D objects. Therefore, point cloud data has emerged as a crucial research subject for addressing 3D object detection issues in autonomous driving scenarios. However, the unorganized, sparse, irregular, and discrete characteristics of point cloud data make it difficult for 3D object detection algorithms to effectively model target features. Thus, developing a resilient feature extraction network using deep neural networks has become a key approach to addressing this issue. In this work, we propose a Multi-Scale Prediction Region Convolutional Neural Network for 3D object detection based on autonomous driving, called as MSP-RCNN, focusing on small scale object feature representation and heightened precision detection in poor conditions. This is achieved by implementing the 3D Transformer Down-sampling Module, Key-point Sampling Module and Multi-scale Detection Head (3D Region Proposal Network). Our method represents an innovative end-to-end structure based on self-attention theory. In particular, our method integrates multi-module feature fusion through global 3D Transformer Down-sampling features and local key point feature fusion to achieve self-attention 3D object representation. Besides, the multi-scale region proposal detection head boosts small-scale object detection precision by facilitating bounding box regression at different levels. Comparative validation experiments carried out on the KITTI dataset confirm that our proposed method enhances detection accuracy by 3.4%. Furthermore, the robustness of our method was validated using the NuScenes dataset, demonstrating an improved detection accuracy of 3.6%. It also effectively deals with climate changes and day-night transition detection problems. |
|||||||||
| 言語 | en | |||||||||
| 目次 | ||||||||||
| 内容記述タイプ | Other | |||||||||
| 内容記述 | 1 Introduction|2 Theory and Concepts|3 Feature Extraction Network Based on Visual-Transformer|4 Multi-scale Prediction Region Proposal Network|5 Experiments | |||||||||
| 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578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 | Computer view | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | 3D点群処理 | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | 3Dオブジェクト認識 | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | 自動運転 | |||||||||
| キーワード | ||||||||||
| 主題Scheme | Other | |||||||||
| 主題 | Transformer | |||||||||
| 学位授与番号 | ||||||||||
| 学位授与番号 | 甲第578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 | 10.18997/0002000289 | |||||||||
| ID登録タイプ | JaLC | |||||||||
