簡易檢索 / 詳目顯示

回結果列表
研究生： 陳昶安
Chen, Chang-An
論文名稱： 結合穩健深度估計之截斷感知單眼三維車輛偵測
Truncation-Aware Monocular 3D Vehicle Detection with Robust Depth Estimation
指導教授： 楊家輝
Yang, Jar-Ferr
學位類別： 碩士
Master
系所名稱： 電機資訊學院 - 電腦與通信工程研究所
Institute of Computer & Communication Engineering
論文出版年： 2026
畢業學年度： 114
語文別： 英文
論文頁數： 94
中文關鍵詞： 電腦視覺 、單眼三維物體偵測 、車輛偵測 、截斷感知 、中心偏移 、深度估計 、穩健深度監督
外文關鍵詞： Computer vision, Monocular 3D object detection, Vehicle detection, Truncation-aware detection, Center offset, Depth estimation, Robust depth supervision
相關次數： 點閱：39  下載：2 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 近年來，單眼三維車輛偵測在自動駕駛與智慧交通系統中受到廣泛關注。此任務僅使用單張 RGB 影像推估車輛於三維空間中的位置、尺寸、方向與深度，具有硬體成本低與部署彈性高等優點。然而，單眼影像缺乏直接的深度量測資訊，使模型必須從影像外觀、物體尺度、透視關係與場景脈絡中推論空間距離，因此在深度估計與三維定位上仍具有高度不確定性。尤其在真實駕駛場景中，車輛可能因影像邊界裁切、前景遮擋或視角限制而產生截斷現象，使可見外觀與幾何資訊不完整，進一步影響投影三維中心定位、深度估計與三維邊界框回歸的穩定性。
    本文提出一種結合穩健深度估計之截斷感知單眼三維車輛偵測方法，用以改善截斷車輛於單眼三維偵測中的定位穩定性。本研究以 MonoDETR 作為基準模型，針對邊界截斷車輛中投影三維中心可能位於影像範圍之外的問題，設計截斷感知中心目標生成與中心偏移學習策略。具體而言，本文將影像外的投影三維中心裁切至有效影像範圍內作為中心監督參考點，並以原始投影中心與裁切後中心之間的位移作為中心偏移目標進行監督。此設計使模型能在保留有效影像內中心參考點的同時，學習邊界截斷所造成的幾何偏移。
    此外，本文引入穩健深度監督策略，結合正規化深度誤差與對數深度誤差，使模型同時考量深度相對誤差與尺度變化。透過距離、遮擋與截斷程度所形成的樣本權重設計，該策略可降低困難樣本與深度離群誤差對物件層級深度預測的干擾，進而提升三維定位的穩定性。
    實驗結果顯示，本文所提出之方法能改善截斷車輛與困難樣本下的偵測表現。在針對近距離嚴重截斷車輛所建立的評估子集合中，本文方法相較於基準模型明顯提升二維邊界框平均精度與平均方向相似度，顯示所提出之中心偏移學習能改善截斷車輛的影像平面定位與方向估計。在 KITTI 驗證集上，本文方法於 Moderate 與 Hard 難度下提升二維偵測表現，並於 Hard 難度下改善三維平均精度。整體而言，本文方法透過截斷感知中心偏移監督與穩健深度監督，提升單眼三維車輛偵測在截斷、遮擋與深度不確定場景中的定位穩定性。

    Monocular 3D vehicle detection has received increasing attention in autonomous driving and intelligent transportation systems. This task estimates the 3D position, dimensions, orientation, and depth of vehicles using only a single RGB image, offering advantages such as low hardware cost and flexible deployment. However, since monocular images do not provide direct depth measurements, the model must infer spatial distance from image appearance, object scale, perspective relationships, and scene context. Therefore, monocular 3D vehicle detection still suffers from high uncertainty in depth estimation and 3D localization. In real driving scenes, vehicles may also be truncated due to image boundary cropping, foreground occlusion, or viewpoint limitations. These conditions result in incomplete visual appearance and geometric information, further affecting the stability of projected 3D center localization, depth estimation, and 3D bounding box regression.
    This thesis proposes a truncation-aware monocular 3D vehicle detection method with robust depth estimation to improve the localization stability of truncated vehicles. The proposed method is built upon MonoDETR as the baseline model. To address the problem that the projected 3D center of a boundary-truncated vehicle may lie outside the image region, this thesis designs a truncation-aware center target generation and center offset learning strategy. Specifically, the out-of-image projected 3D center is clipped to the valid image region and used as the reference point for center-related supervision. The displacement between the original projected center and the clipped center is then defined as the center offset target for additional supervision. This design enables the model to learn the geometric displacement caused by boundary truncation while maintaining a valid in-image center reference.
    In addition, this thesis introduces a robust depth supervision strategy that combines normalized-depth error and log-depth error. This strategy allows the model to consider both relative depth deviation and scale variation during training. By further incorporating sample-wise weights based on distance, occlusion, and truncation levels, the proposed depth supervision reduces the influence of difficult samples and depth outliers on object-level depth prediction, thereby improving the stability of 3D localization.
    Experimental results show that the proposed method improves the detection performance of truncated vehicles and difficult samples. On the evaluation subset constructed for near-distance severely truncated vehicles, the proposed method significantly improves 2D bounding box average precision and average orientation similarity compared with the baseline model. This indicates that the proposed center offset learning strategy improves image-plane localization and orientation estimation for truncated vehicles. On the KITTI validation set, the proposed method improves 2D detection under the Moderate and Hard settings and 3D average precision under the Hard setting, demonstrating more stable localization under truncation, occlusion, and depth uncertainty.

    摘要 II Abstract IV 誌謝 VI Contents VII List of Tables X List of Figures XI Chapter 1 Introduction 1 1.1 Research Background 1 1.2 Motivations 4 1.3 Thesis Organization 6 Chapter 2 Related Work 9 2.1 Monocular 3D Object Detection 10 2.2 Faster R-CNN 11 2.3 Anchor-Free Object Detection: FCOS and CenterNet 13 2.4 Transformer-based Monocular 3D Object Detection 15 2.5 Depth-guided Monocular 3D Object Detection 17 2.6 Lightweight Depth Predictor 20 2.7 Truncation-aware Object Detection 20 2.8 Scale-aware Depth Learning and Robust Regression Losses 23 2.9 Baseline Loss Functions 26 Chapter 3 Proposed Truncation-Aware Monocular 3D Vehicle Detection Method 31 3.1 Overview of the Proposed Truncation-Aware Detection Framework 33 3.2 Bipartite Matching Using the Hungarian Algorithm 36 3.3 Truncation-Aware Center Target Generation and Offset Learning 37 3.3.1 Motivation for Boundary-Truncated Center Localization 38 3.3.2 3D Center Projection and Clipping 39 3.3.3 Center Offset Target and Mask Generation 41 3.3.4 Center Offset Prediction and Regression Loss 43 3.4 Robust Depth Supervision for Object-Level Depth Estimation 45 3.4.1 Motivation for Robust Depth Loss 46 3.4.2 Matched Object-Level Depth Target 48 3.4.3 Normalized-Depth and Log-Depth Error Formulation 49 3.4.4 Sample-wise Reliability Weighting and Robust Depth Loss 51 3.5 Overall Training Objective 54 3.6 Inference-Time Driving Warning Description Generator 55 Chapter 4 Experiment Results 58 4.1 Environment Setup and Datasets 59 4.2 Training Details and Hyperparameter Settings 60 4.3 Evaluation Metrics 61 4.4 Qualitative Analysis on Boundary-Truncated Vehicles 62 4.5 Benchmark Evaluation on Near-Distance Boundary-Truncated Vehicles 66 4.6 Ablation Study 67 4.7 Hyperparameter Analysis 71 4.8 Comparison with Representative Methods on KITTI Validation Set 73 Chapter 5 Conclusions 76 Chapter 6 Future Work 78 References 79

    [1] A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? The KITTI vision benchmark suite,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2012, pp. 3354–3361.
    [2] A. H. Lang, S. Vora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “PointPillars: Fast encoders for object detection from point clouds,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 12697–12705.
    [3] G. Brazil and X. Liu, “M3D-RPN: Monocular 3D region proposal network for object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 9287–9296.
    [4] Y. Lu, X. Ma, L. Yang, T. Zhang, Y. Liu, Q. Chu, J. Yan, and W. Ouyang, “Geometry uncertainty projection network for monocular 3D object detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 3111–3121.
    [5] Y. Zhang, J. Lu, and J. Zhou, “Objects are different: Flexible monocular 3D object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 3289–3298.
    [6] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in Proceedings of the European Conference on Computer Vision (ECCV), 2020, pp. 213–229.
    [7] X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable DETR: Deformable transformers for end-to-end object detection,” in International Conference on Learning Representations (ICLR), 2021.
    [8] R. Zhang, H. Qiu, T. Wang, Z. Guo, Y. Tang, X. Xu, Z. Cui, Y. Qiao, P. Gao, and H. Li, “MonoDETR: Depth-guided Transformer for Monocular 3D Object Detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023, pp. 9155–9166.
    [9] A. Mousavian, D. Anguelov, J. Flynn, and J. Kosecka, “3D Bounding Box Estimation Using Deep Learning and Geometry,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 7074–7082.
    [10] Z. Liu, Z. Wu, and R. Tóth, “SMOKE: Single-Stage Monocular 3D Object Detection via Keypoint Estimation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020.
    [11] Y. Wang, W. Chao, D. Garg, B. Hariharan, M. Campbell, and K. Q. Weinberger, “Pseudo-LiDAR From Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 8445–8453.
    [12] X. Ma, Y. Liu, T. Xia, Y. Wang, Z. Tang, and L. Chen, “Delving Into Localization Errors for Monocular 3D Object Detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 4721–4730.
    [13] Y. Wang, V. Guizilini, T. Zhang, Y. Wang, H. Zhao, and J. M. Solomon, “DETR3D: 3D Object Detection from Multi-view Images via 3D-to-2D Queries,” in Proceedings of the Conference on Robot Learning (CoRL), 2021.
    [14] Y. Liu, T. Wang, X. Zhang, and J. Sun, “PETR: Position Embedding Transformation for Multi-View 3D Object Detection,” in Proceedings of the European Conference on Computer Vision (ECCV), 2022, pp. 531–548.
    [15] M. Ding, Y. Huo, H. Yi, Z. Wang, J. Shi, Z. Lu, and P. Luo, “Learning Depth-Guided Convolutions for Monocular 3D Object Detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 1000–1001.
    [16] C. Reading, A. Harakeh, J. Chae, and S. L. Waslander, “Categorical Depth Distribution Network for Monocular 3D Object Detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 8555–8564.
    [17] K.-C. Huang, T.-H. Wu, H.-T. Su, and W. H. Hsu, “MonoDTR: Monocular 3D Object Detection with Depth-Aware Transformer,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 4012–4021.
    [18] Z. Qin, J. Wang, and Y. Lu, “MonoGRNet: A General Framework for Monocular 3D Object Detection,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 9, pp. 5170–5184, 2022.
    [19] Y. Tang, S. Dorn, and C. Savani, “Center3D: Center-based Monocular 3D Object Detection with Joint Depth Understanding,” arXiv preprint arXiv:2005.13423, 2020.
    [20] D. Eigen, C. Puhrsch, and R. Fergus, “Depth Map Prediction from a Single Image using a Multi-Scale Deep Network,” in Advances in Neural Information Processing Systems (NeurIPS), 2014.
    [21] H. Fu, M. Gong, C. Wang, K. Batmanghelich, and D. Tao, “Deep Ordinal Regression Network for Monocular Depth Estimation,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 2002–2011.
    [22] P. J. Huber, “Robust Estimation of a Location Parameter,” The Annals of Mathematical Statistics, vol. 35, no. 1, pp. 73–101, 1964.
    [23] R. Girshick, “Fast R-CNN,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015, pp. 1440–1448.
    [24] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal Loss for Dense Object Detection,” in Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2017, pp. 2980–2988.
    [25] H. Rezatofighi, N. Tsoi, J. Gwak, A. Sadeghian, I. Reid, and S. Savarese, “Generalized Intersection over Union: A Metric and a Loss for Bounding Box Regression,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 658–666.
    [26] A. Kumar, G. Brazil, E. Corona, A. Parchami, and X. Liu, “DEVIANT: Depth EquiVarIAnt NeTwork for Monocular 3D Object Detection,” in Proceedings of the European Conference on Computer Vision (ECCV), 2022, pp. 664–683.
    [27] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” Advances in Neural Information Processing Systems, vol. 28, pp. 91–99, 2015.
    [28] Z. Tian, C. Shen, H. Chen, and T. He, “FCOS: Fully Convolutional One-Stage Object Detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9627–9636, 2019.
    [29] X. Zhou, D. Wang, and P. Krähenbühl, “Objects as Points,” arXiv preprint arXiv:1904.07850, 2019.

    下載圖示
    校外：立即公開
    QR CODE