(1) 单一视角非特定目标检测与跟踪
感知单一视角下的非特定目标是计算机视觉领域的一个关键问题。在该领域,我们的研究聚焦于非特定视频目标检测与跟踪方向。
① 针对非特定场景下的动态目标检测方面,我们提出了以下方法:1. 粘聚性超像素网格,显著提升了检测算法的运行时间和精度;2. 基于示例的信息选择模型,实现了显著性目标的检测;3. 基于显著性感知的外观模型,提升了目标跟踪的精度;4. 基于背景抑制的正则化方法,有效提升了算法的鲁棒性;5. 选择性空间正则化方法,实现了结合背景环境的鲁棒跟踪;6. 探究了运动模糊的影响,发现轻度模糊有助于实现鲁棒目标跟踪等。
② 在车道线检测方面,我们提出了以下创新:1. 提出基于视频的车道线检测,有效融合了时序信息;2. 收集和标注了视频级实例车道线检测(VIL-100)数据集,并基于时域连续性和几何一致性,提升了车道线检测的性能。
主要成果包括:
① Liang Li, Wei Feng, Liang Wan, Jiawan Zhang. Maximum Cohesive Grid of Superpixels for Fast Object Localization. In CVPR, 2013.
② Rui Huang, Wei Feng, Zezheng Wang, Yan Xing, Yaobin Zou. Exemplar-based image saliency and co-saliency detection. Neurocomputing, 2020.
③ Yujun Zhang, Lei Zhu, Wei Feng, Huazhu Fu, Mingqian Wang, Qingxia Li, Cheng Li, Song Wang. VIL-100: A New Dataset and A Baseline Model for Video Instance Lane Detection. In ICCV, 2021.
④ Mingqian Wang, Yujun Zhang, Wei Feng, Lei Zhu, Song Wang. Video Instance Lane Detection via Deep Temporal and Geometry Consistency Constraints. In ACM MM, 2022.
⑤ Qing Guo, Wei Feng, Ce Zhou, Chi-Man Pun, Bin Wu. Structure-Regularized Compressive Tracking with Online Data-Driven Sampling. IEEE TIP, 2017.
⑥ Wei Feng, Ruize Han, Qing Guo, Jianke Zhu, Song Wang. Dynamic Saliency-aware Regularization for Correlation Filters based Object Tracking. IEEE TIP, 2019.
⑦ Ruize Han, Qing Guo, Wei Feng. Content-Related Spatial Regularization for Object Tracking. In ICME, 2018. (Best Paper Award)
⑧ Qing Guo, Wei Feng, Ce Zhou, Rui Huang, Liang Wan, Song Wang. Learning Dynamic Siamese Network for Visual Object Tracking. In ICCV, 2017.
⑨ Zhihao Chen, Qing Guo, Liang Wan, Wei Feng. Background-Suppressed Correlation Filters for Visual Tracking. In ICME, 2018.
⑩ Ce Zhou, Qing Guo, Liang Wan, Wei Feng. Selective Object and Context Tracking. In ICASSP, 2017.
11 Qing Guo, Ruize Han, Wei Feng, Zhihao Chen, Liang Wan. Selective Spatial Regularization by Reinforcement Learned Decision Making for Object Tracking. IEEE TIP, 2017.
12 Qing Guo, Wei Feng, Ruijun Gao, Yang Liu, Song Wang. Exploring the Effects of Blur and Deblurring to Visual Object Tracking. IEEE TIP, 2021.
(2) 多视角协同的多目标关联感知
当前视频采集分析系统主要采用固定摄像头进行监控,然而,由于监控视角的固定、覆盖范围的有限以及远处分辨率不足等限制,视频分析仍然具有巨大的潜力待挖掘。为解决这一问题,我们团队面向多人场景公共安全维护的迫切需求,致力于构建主动式移动视觉系统,并创新性地提出了基于互补视角的多目标协同跟踪。我们实现了多源异步相机的全时空跨视角多目标关联,并采用了自监督框架,以实现精准的多视角视频时间同步。此外,我们通过基于多视几何基本原理的方法,成功实现了在大视角差异下的视频空间对齐。我们创建了一个虚拟-真实结合的多互补视角多目标关联跟踪基准,并提出了多视角多目标协同跟踪方法,可作为跨视角关联的通用范式。我们还提出了基于自监督学习的多视角多目标关联跟踪框架,以及目标重识别方法,实现了无重叠视域相机的多目标关联。
主要成果包括:
① Ruize Han, Wei Feng, Jiewen Zhao, Zicheng Niu, Yujun Zhang, Liang Wan, Song Wang. Complementary-View Multiple Human Tracking. In AAAI, 2020.
② Ruize Han, Wei Feng, Yujun Zhang, Jiewen Zhao, Song Wang. Multiple Human Association and Tracking from Egocentric and Complementary Top Views. IEEE TPAMI, 2021.
③ Liqiang Yin, Ruize Han, Wei Feng, Song Wang. Self-Supervised Human Pose based Multi-Camera Video Synchronization. In ACM MM, 2022.
④ Ruize Han, Yiyang Gan, Likai Wang, Nan Li, Wei Feng, Song Wang. Relating View Directions of Complementary-View Mobile Cameras via the Human Shadow. IJCV, 2023.
⑤ Ruize Han, Yiyang Gan, Jiacheng Li, Feifan Wang, Wei Feng, Song Wang. Connecting the Complementary-view Videos: Joint Camera Identification and Subject Association. In CVPR, 2022.
⑥ Ruize Han, Yun Wang, Haomin Yan, Wei Feng, Song Wang. Multi-view Multi-human Association with Deep Assignment Network. IEEE TIP, 2022.
⑦ Yiyang Gan, Ruize Han, Liqiang Yin, Wei Feng, Song Wang. Self-supervised Multi-view Multi-Human Association and Tracking. In ACM MM, 2021.
(3) 复杂场景多目标行为分析理解
对于复杂场景的多目标行为分析理解对于帮助各类系统理解目标的行为和状态至关重要。因此,我们展开了相关研究,旨在实现复杂场景中多目标的多尺度行为分析。我们团队提出了以下创新性方法:
1. 目标行为分析知识蒸馏框架: 通过基于人体姿态的单人行为识别,实现了对个体行为的精准识别。
2. 个体关系卷积模块: 引入了个体关系卷积模块,基于人体姿态,实现了群体行为的精准识别。
3. 多视角多人场景交互行为检测与识别: 提出了从多视角多人场景进行交互行为检测与识别,有效捕捉了不同目标之间的复杂关系。
4. 基于自监督学习的交互识别方法: 通过自监督学习方法,成功缓解了手工标注的难题,提高了交互行为的识别效果。
5. 互补多视角重点目标检测与定位: 引入互补多视角,实现了对重点目标的准确检测与定位。
6. 全景行为识别: 我们提出了全景行为识别,可获取不同颗粒度的行为识别结果,为多目标行为分析提供了更全面的视角。
主要成果包括:
① Cunling Bian, Wei Feng, Liang Wan, Song Wang. Structural knowledge distillation for efficient skeleton-based action recognition. IEEE TIP, 2021.
② Cunling Bian, Wei Feng, Song Wang. Self-Supervised Representation Learning for Skeleton-Based Group Activity Recognition. In ACM MM, 2022.
③ Jiewen Zhao, Ruize Han, Yiyang Gan, Liang Wan, Wei Feng, Song Wang. Human Identification and Interaction Detection in Cross-View Multi-Person Videos with Wearable Cameras. In ACM MM, 2020.
④ Jiacheng Li, Ruize Han, Haomin Yan, Zekun Qian, Wei Feng, Song Wang. Self-supervised Social Relation Representation for Human Group Detection, In ECCV, 2022.
⑤ Ruize Han, Jiewen Zhao, Wei Feng, Yiyang Gan, Liang Wan, Song Wang. Complementary-View Co-Interest Person Detection. In ACM MM, 2020.
⑥ Ruize Han, Haomin Yan, Jiacheng Li, Wei Feng, Songmiao Wang, Song Wang. Panoramic Human Activity Recognition. In ECCV, 2022.