
Hello, I am Jiaxin
I work with Computer Vision, Machine Learning, Autonomous Driving, UAV, and Robotics
I am currently a Research Scientist at Hyundai-Aptiv Joint Venture / nuTonomy.
I obtained a PhD from the National University of Singapore, Electrical and Computer Engineering.
And the Bachelor of Engineering from Tsinghua University, Beijing, China.
Publications
ICCV 2019
Jiaxin Li, Gim Hee Lee; USIP: Unsupervised Stable Interest Point Detection from 3D Point Clouds, ICCV 2019
CVPR 2018
Jiaxin Li, Ben M. Chen, Gim Hee Lee; So-net: Self-organizing network for point cloud analysis, CVPR 2018
ICRA 2019
Jiaxin Li, Yingcai Bi, Gim Hee Lee; Discrete Rotation Equivariance for Point Cloud Recognition, ICRA 2019
IROS 2017
Jiaxin Li, Huangying Zhan, Ben M. Chen, Gim Hee Lee; Deep learning for 2D scan matching and loop closure, IROS 2017
ICCA 2018
Jiaxin Li, Yingcai Bi, Kun Li, Kangli Wang, Feng Lin, Ben M. Chen; Accurate 3d localization for mav swarms by uwb and imu fusion, ICCA 2018
IMAV 2016
Jiaxin Li, Yingcai Bi, Menglu Lan, Hailong Qin, Mo Shan, Ben M. Chen; Real-time simultaneous localization and mapping for uav: a survey, IMAV 2016
IECON 2016
Jiaxin Li, Mo Shan, Menglu Lan, Yingcai Bi, Hailong Qin, Feng Lin, Ben M. Chen; Semi-dense motion segmentation for moving cameras by discrete energy minimization, IECON 2016
IECON 2015
Jiaxin Li, Feng Lin, Ben M. Chen; A statistical approach for trajectory analysis and motion segmentation for freely moving cameras, IECON 2015
JIRS 2019
Yang Xu, Shupeng Lai, Jiaxin Li, Delin Luo, Yancheng You; Concurrent optimal trajectory planning for indoor quadrotor formation switching, Journal of Intelligent & Robotic Systems 2019
Past Projects
Point Cloud-based Object Classification
Trained with PyTorch, deployed with TensorRT with INT8 calibration. Performance: ~2ms for ~400 objects on Nvidia GTX1080.

Nearest Neighbor Search Algorithm in Cuda
Efficient Octree based nearest neighbour searching algorithm with CPU and GPU. Performance: 30k 8-NN searches on 30k points takes 1.4ms on 1080Ti. Building the Octree takes 1.4ms on 3.6GHz CPU for 30k points.

Indoor Multi-Drone Light Show
Lead a team to perform a multi-UAV light show inside an exhibition hall, based on Ultra-Wide-Band (UWB) localization, and multi-agent splines based trajectory generation. Developed UKF and least-square optimization algorithms to integrated UWB with IMU readings, which achieved the accuracy of 10cm for indoor 3D localization.

Ground Removal for Autonomous Driving
Lidar based ground plane / drivable surface estimation. Performance: 10ms on 2.2GHz CPU 4 threads for 80k points. Global patent application filed. ICRA paper in progress.


