Yifei Kang
Work & Education Experience
Changan Automobile 2020.11 - Present
- Modules: BEV Static, BEV Dynamic (Parking), Localization, Calibration
- Team Size: 68
Zongmu Technology 2019.7 - 2020.10
- Senior Localization Algorithm Engineer
JD X-Division Autonomous Driving Team 2018.7 - 2019.7
- Autonomous Driving Algorithm Engineer
Hualu Engineering Technology Co., Ltd. 2015.7 - 2018.6
- Automation & Instrumentation Engineer
- Software Product Development
Beijing Jiaotong University 2006.9 - 2015.7
- Automation Department
Bachelor's Degree
- Ph.D. in Transportation Information Engineering and Automation
- Key Courses: Multiple View Geometry in Computer Vision; FastSLAM
Project Experience
Parking End-to-EndChangan 2026
- Led parking end-to-end new pipeline design: full AI coding, new DAG architecture, 360x efficiency improvement over driving end-to-end pipeline
- Contributed to parking end-to-end model development: modified PlanR1 Geometry deep learning network; responsible for pruning traffic light modules for parking scenarios
- Agent Development:
- Completed: fully automated weekly report agent; automated commit summary agent
- In progress: fully automated engineering agents including code refactoring, testing, and graphical verification agents
Intelligent Driving Perception System Technical LeadChangan 2024-2025
- Core Achievements: Real-time semi-dense Mapping [Industry’s Only Mass Production], OCC Mapping Accuracy 2.78cm [Target <5cm Achieved, Industry-Leading], Real-time LiDAR Odometry 3.5‰ [Best-in-Class Mass Production]
- Led mass production vehicle calibration, after-sales calibration, and online calibration; [Mass Production: CHANGAN NEVO E07, Q07]
- Led Real-time crowdsourced mapping and localization, from vehicle sensor data real-time construction of HD-map-like maps; [Mass Production: CHANGAN NEVO E07]
- Led industry’s only mass-produced real-time semi-dense mapping (replacing OCC). Real-time semi-dense mapping around target parking spaces, effectively solving suspended obstacle detection challenges, precisely identifying umbrellas, ping-pong tables and other extreme cases; [Mass Production: CHANGAN NEVO Q07]
- Led large-scale high-precision OCC data pipeline design: mapping accuracy 2.78cm (achieving <5cm design target, industry-leading), benchmarked against Tesla 10cm; involved in sensor selection, bracket design, multi-LiDAR large-scale Bundle Adjustment (5 LiDARs, 600 frames, ~30 million point clouds, single LiDAR accuracy 2.5cm)
- Led design of best-in-class mass production real-time high-precision LiDAR odometry: mass production LiDAR accuracy 5cm, odometry achieving 3.5‰ accuracy (only 3.5cm error per 10m, FastLio 7cm and LOAM 12cm), benchmarked against Huawei SuperOdom; [Mass Production: CHANGAN NEVO Q07]
- Led BEV network ground height estimation module design for better PV collaboration; [Mass Production: CHANGAN NEVO Q07]
- Led BEV network static and dynamic model post-processing; [Mass Production: CHANGAN NEVO Q07]
Intelligent Driving Localization System Technical LeadChangan 2022-2023
- Real-time visual ICP with wheel odometry and imu localization algorithm based on HD maps, supporting mass production highway assisted driving; [Mass Production: Deepal SL03, S07]
- Multi Camera feature point-based SLAM underground parking garage mapping and localization algorithm, supporting mass production AVP; [Mass Production: Deepal SL03, S07]
Academic Research & Open Source Contributions
Multi-Camera Vision SLAM System Zongmu 2021 【IROS 2021 First Author Paper Engineering Implementation】
- Led multi-camera vision system mapping and localization algorithm development
- Led multi-camera vision system online calibration
- Published at IROS 2021, the top robotics conference, as first author
Open Source Core Contributions
- OpenCV Pull Request #14583: Fixed critical bug, improving computer vision core library stability
- g2o Pull Request #402: Optimized graph optimization library performance, contributing to SLAM core algorithm library
- msckf Pull Request #116: Improved visual SLAM algorithm implementation, enhancing localization accuracy
Other Experience
China Association of Automobile Manufacturers Committee Member Changan
Selected Publications
[1] Yifei Kang, Yu Song, WuWei Ge and Tong Ling, Robust Multi-Camera SLAM with Manhattan Constraint Toward Automated Valet Parking. International Conference on Intelligent Robots and Systems (IROS), 2021, 7592-7599.
[2] Yifei Kang, Yongduan Song, Yu Song, and Deli Yan, Stereo Visual Odometry Algorithm with Rotation-Translation Decoupling for Dynamic EnvironmentsSquare-Root Cubature Kalman Filter and Its Application to SLAM of a Mobile Robot. Robot,2014, 36(3), 758-768;
[3] Yifei Kang, Yongduan Song, Yu Song, and Deli Yan, Square-Root Cubature Kalman Filter and Its Application to SLAM of a Mobile Robot. Robot,2013, 35(2),186-193;
[4] Yifei Kang, Yongduan Song, Yu Song, and Deli Yan, SLAM without odometer. Journal of ZHEJIANG University, 2014, 48(3): 414-422.
[5] Y Song, Q L Li, Y F Kang and Y D Song, CFastSLAM: a new Jacobian free solution to SLAM problem. International Conference on Robotics and Automation (ICRA), 2012, 3063-3068.
[6] Y Song, Q L Li, Y F Kang and Deli Yan. Effective cubature FastSLAM: SLAM with Rao-Blackwellized particle filter and cubature rule for Gaussian weighted integral, Aug 2013.
[7] Y Song, Q L Li, Deli Yan and Y F Kang, Robust Visual Tracking with Classifier-like appearance Model and Entropy Particle Filter, Processing of the 10 th World Congress on Intelligent Control and Automation, 2012, 4853-4858.
[8] Y.D Song, W.C Cai, Y F Kang and Danyong Li, A System to Prevent the Controller from the Impact, invention patent.
[9] Y.D song, Y F Kang and Lei Wang, Real-Time Localization System without GPS for High Speed Train, invention patent.