学位论文简介
在环境气候复杂多变的现实交通场景下，对交通目标进行鲁棒检测对于自动驾驶车辆的安全行驶具有重大意义。虽然当前的检测模型已经取得了成功，但它们在覆盖交通目标广泛性、应对困难情况充分性、部署实际车辆可用性等多方面鲁棒性问题上仍存在缺陷和较大进步空间。基于以上分析，本文提出了一系列算法与模型来应对当前挑战。主要创新点如下：
（1）针对由于室外极端场景导致交通文本检测模型性能显著下降甚至完全失效的问题，提出一种极端环境下域自适应的交通场景文本检测模型，设计了多粒度文本提案网络与一种兼具领域间和领域内自适应的方法。
（2）针对由于激光雷达点云的“近密远疏”特性而导致其难以精准检测远处目标的问题，提出一种远距离对象增强的基于点的 3D 目标检测模型，通过设计的远距离对象增强的集合抽象和回归方式，分别解决了“实例点不平衡”与“候选框不平衡”两大难题。
（3）针对现有的自注意力机制存在的查询点感受野搜索不精准问题，提出一种基于监督式尺度感知 Transformer 的单目 3D 目标检测模型，解决了由于噪音特征干扰导致的小目标检测精度难以提升的难题。
（4）针对由于不同类别物体视觉外观差异显著导致的多类别统一检测难以实行的问题，提出一种模态辅助与多类别外观感知的单目 3D 目标检测模型，解决了现有模型需要针对性地为了不同类别物体调整超参数并重新训练的难题。
（5）将提出的模型部署在课题组自研的智能无人物流车上，并在现实园区中进行运行测试，证明了模型的高度实用性与强鲁棒性。
主要学术成果
Xuan He, Zian Wang, Jiacheng Lin, Ke Nai, Jin Yuan, and Zhiyong Li. “DO-SA&R: Distant Object Augmented Set Abstraction and Regression for Point-based 3D Object Detections,” in IEEE Transactions on Image Processing (2023), doi: 10.1109/TIP.2023.3326394.(第一作者，CCF A 类，中科院一区)
Xuan He, Fan Yang, Jiacheng Lin, Haolong Fu, Jin Yuan, Kailun Yang, and Zhiyong Li. "SSD-MonoDETR: Supervised Scale-aware Deformable Transformer for Monocular 3D Object Detection," in IEEE Transactions on Intelligent Vehicles (2023), doi: 10.1109/TIV.2023.3311949. (第一作者，中科院一区)
Xuan He, Jin Yuan, Mengyao Li, Runmin Wang, Haidong Wang, and Zhiyong Li. "A Text-Specific Domain Adaptive Network for Scene Text Detection in the Wild," in Applied Intelligence (2023): 1-13. (第一作者，中科院二区)
Xuan He, Zhiyong Li, Jiacheng Lin, Ke Nai, Jin Yuan, Yifan Li, and Runmin Wang. "Domain adaptive multigranularity proposal network for text detection under extreme traffic scenes," in Computer Vision and Image Understanding 233 (2023): 103709. (第一作者，CCF B 类，中科院三区)
Xuan He, Kailun Yang, Junwei Zheng, Jin Yuan, Luis M. Bergasa, Hui Zhang, and Zhiyong Li. “S3-MonoDETR: Supervised Shape&Scale-perceptive Deformable Transformer for Monocular 3D Object Detection,” in IEEE Transactions on Multimedia, Under Review. (中科院一区)
Jiacheng Lin, Zhiqiang Xiao, Xiaohui Wei, Puhong Duan, Xuan He, Renwei Dian, Zhiyong Li and Shutao Li.“Click-pixel Cognition Fusion Network with Balanced Cut for Interactive Image Segmentation,” in IEEE Transactions on Image Processing (2023), doi:10.1109/TIP.2023.3338003. (CCF A 类，中科院一区)
Li, Yifan, Xiaoyan Peng, Ziyan Wu, Fan Yang, Xuan He, and Zhiyong Li. "M3GAN: A masking strategy with a mutable filter for multidimensional anomaly detection," in Knowledge-Based Systems 271 (2023): 110585. (中科院一区)
Wang, Haidong, Xuan He, Zhiyong Li, Jin Yuan, and Shutao Li. "JDAN: Joint Detection and Association Network for Real-Time Online Multi-Object Tracking," in ACM Transactions on Multimedia Computing, Communications and Applications 19, no. 1s (2023): 1-17. (CCF B 类)
