Li He (何力)
Associate Professor (with exceptional promotion)
Email: heli [at] gdut [dot] edu [dot] cn
Room 206, Dept. of Electromechanical Engineering
Guangdong University of Technology, Guangzhou, China, 510006
I am an Associate Professor in the Department of Electromechanical Engineering, Guangdong University of Technology. I received my B.Sc., M.Sc. and Ph.D. from Department of Automation at Northwestern Polytechnical University, Xi’an, China, in 2006, 2009 and 2014, respectively. I was a visiting Ph.D. Student in the Department of Computing Science, University of Alberta, Canada, from 2010 to 2011 and then served as a Postdoctoral Fellow from 2014 to 2017. I joined GDUT in 2017.
My research interests include machine learning, visual SLAM and computer vision. I have worked in a number of areas in machine learning and applied to various vision-based tasks in both academic and industrial fields.
- Two papers accepted by ICRA 2019. Congratulations to Zhuang Dai and Xinghong Huang (and the co-authors), both second-year graduate students.
Google Scholar & Github
Google Scholar: https://scholar.google.com/citations?user=U3Ek_lUAAAAJ&hl=en
Kernel Mapping, Nystrom Approximation and Spectral Clustering
Efficient algorithms focus on acceleratiing Spectral Clustering (SC) and kernel mapping. SC suffers from a cubic growth of time consumption that prevents it from large-scale tasks. Approximation algorithms are developed to handle this problem.
Large-scale Data Clustering
Kernel mapping implicitly uplifts data to the kernel space in which data are assumed to be linearly separable. Explicit Feature Mapping (EFM) is able to explicitly generate the uplifted data.
We developed an efficient algorithm to run SC on large-scale data by using EFM. Our method is parallel-computing-friendly and requires a constant memory consumption. In contrast, the standard SC has a square growth on the memory. Our method obtains the embeddings of 3.3M handwritten digits in 25 second.
Nystrom approximation is the most popular way in approximating eigen-decomposition of a kernel matrix and hence widely used in SC. We proved several theoretical results over Nystrom approximation, such as the approximation error, the optimal sampling and the selection of training size.
We propose a unified framework to present Nystrom. We then show that the error upper bound of this framework can be optimized by kernel k-means sampling. This result is suitable for most sampling methods, such as uniform, k-means, and also valid for common kernels, such as Gaussian and polynomial.
We are interested in developing mobile robots capable of navigating autonomously using vision or Lidar. This problem is recognized in the research community as SLAM (simultaneous localization and mapping).
M2DP: Point Cloud Global Descriptor
Light Detection And Ranging, or LiDAR, is widely used in mobile robot, AGV and auto-driving car. Compared with visual image, LiDAR system has its advantages in two folds: a) able to obtain accurate depth values and b) illumination invariant that works well in day- and night time.
We developed a novel point cloud global descriptor, entitled M2DP, to detect loop closures in SLAM. M2DP is 2-5 times faster than its counterparts and obtains the highest accuracy in many challenging data sets. M2DP is able to calculate the descriptor for 120K points in 0.35 second on a common laptop, indicating a real-time process ability essential for mobile robot.
M2DP is also robust to downsampling (1/100) and noise (50*resolution).
Vision-based Object Detection, Tracking and Recognition
We have worked on many vision-based industrial tasks in various purposes, mainly driven by oil sand mining and processing industry in Alberta, Canada and mobile manipulation in Guangdong, China.
We have developed a Missing Tooth Detection system for shovels used in Fort McMurray and worked on Bitumen Particle Tracking system (diameter of particles less than 10 um) for oil sand analysis in Edmonton.
- Li He, Nilanjan Ray, Yisheng Guan and Hong Zhang. Fast Large-Scale Spectral Clustering via Explicit Feature Mapping. IEEE Transactions on Cybernetics, Vol. 49, Issue 3, March 2019, pp. 1058-1071. (paper) (code)
- Zhuang Dai, Xinghong Huang, Weinan Chen, Li He and Hong Zhang. A Comparison of CNN-Based and Hand-Crafted Keypoint Descriptors. to appear in ICRA 2019.
- Xinghong Huang, Zhuang Dai, Weinan Chen, Li He and Hong Zhang. Improving Keypoint Matching Using a Landmark-Based Image Representation. to appear in ICRA 2019.
- Li He and Hong Zhang. Kernel K-means Sampling for Nystrom Approximation. IEEE Transactions on Image Processing, Volume 27, Issue 5, May 2018, pp. 2108-2120. (paper) (code)
- Li He, Haifei Zhu, Tao Zhang, Honghong Yang and Yisheng Guan. Projected Affinity Values for Nystrom Spectral Clustering. Entropy, 2018, 20(7): 519. (paper)
- Li He, Yi Li, Xiang Zhang, Chuangbin Chen, Lei Zhu and Chengcai Leng. Incremental Spectral Clustering via Fastfood Features and Its Application to Stream Image Segmentation. Symmetry, 2018, 10(7): 272. (paper)
- Wen JM, He L* and Zhu FM. Swarm Robotics Control and Communications: Imminent Challenges for Next Generation Smart Logistics. IEEE Communications Magazine, 2018, 56(7): 102-107. (paper)
- Li He, Nilanjan Ray and Hong Zhang. Error Bound of Nystrom-approximated NCut Eigenvectors and Its Application to Training Size Selection. Neurocomputing, 2017, Vol. 239, May: 130-142. (paper) (code)
- Xubin Lin, Weinan Chen, Li He, Yisheng Guan, Guanfeng Liu. Improving Robustness of Monocular VT&R System with Multiple Hypothesis. IEEE ROBIO 2017 (Finalist of T. J. Tarn Best Paper in Robotics). (paper)
- Li He and Hong Zhang. Iterative Ensemble Normalized Cuts. Pattern Recognition, 2016, Vol. 52, April: 274-286. (paper)
- Li He, Xiaolong Wang, and Hong Zhang. M2DP: A novel 3D point cloud descriptor and its application in loop closure detection. In Intelligent Robots and Systems (IROS), 2016 IEEE/RSJ International Conference on, IROS 2016, pp. 231-237. IEEE, 2016. (paper) (code)
- Li He, Shiru Qu. Iterative spectral clustering by inverse stereographic projection. Control and Decision, 2014, 29(3): 396-402 (in Chinese). (paper)
- Li He, Shiru Qu. Dimensionality Reduction of Vehicle Visual Features by PLS. China Journal of Highway and Transport, 2014(4): 98-105 (in Chinese). (paper)
- He Li, Qu Shiru, Zhang Daqi. Image Enhancement Based on Inter-Scale Correlations of Nonsubsampled Contourlet Coefficients. Journal of Northwestern Polytechnical University, 2010(1): 42-46 (in Chinese). (paper)
- Li He, Shiru Qu and Huaifeng Li. A Part-based Online Modeling Algorithm and Its Applications on Robust Visual Tracking. Chinese Control Conference, CCC 2013. (paper)
- Li He, Hui Wang and Hong Zhang. Object detection by parts using appearance, structural and shape features. 2011 IEEE International Conference on Mechatronics and Automation, ICMA 2011, Beijing, 2011: 489-494. (paper)