Chenwei Lyu Email & Phone Number

Santa Clara, Ca, Us

Santa Clara, Ca, Us

San Jose, California, United States

At the Perception Team, I developed a deep learning model, Super-LSS, to transform camera-view features into Bird's Eye View (BEV) features, delivering high average precision, speed, and robustness in object detection tasks. I proposed a geometry-based approach to map multi-view features to 3D anchor coordinates and BEV space, leveraging LiDAR supervision to optimize depth distribution estimation, outperforming other state-of-the-art transformer-based methods. Additionally, I achieved efficient cross-modality fusion of camera-view and BEV features, significantly enhancing performance on small object detection.

Pittsburgh, Pennsylvania, United States

As a Research Assistant at Carnegie Mellon University's Robotics Institute, working under Sebastian Scherer, I am contributing to the project "Towards Universal State Estimation and Reconstruction in the Wild." I integrated perception networks with a SLAM system based on MAC-VO, enabling robust and dense stereo visual odometry. Additionally, I performed feature distillation from DINO-v2 to Efficient-ViT, achieving fast and reliable feature extraction for enhanced performance in SLAM tasks.

Pittsburgh, Pennsylvania, United States

I developed an innovative self-supervised network based on HRNet for depth estimation, which outperformed existing methods on the KiTTI-360 dataset. This work was published in IGARSS 2023 as first author. I implemented a dual-branch network architecture that integrates ResNet and Spherical-HRNet, enabling feature extraction and eliminating spherical distortion. Additionally, I proposed a channel attention decoder to enhance modality fusion. To achieve efficient self-supervision, I introduced photometric loss, SSIM loss, and optical flow loss into the loss function.

Pittsburgh, Pennsylvania, United States

I proposed a refined 2D-3D matching scheme to align local point clouds with a global dense map, significantly improving pose estimation accuracy for event frames and enhancing event camera localization on the DSEC dataset. I also developed a smart car system to collect data from LiDAR, event cameras, and IMU, providing critical inputs for evaluating the proposed algorithm.

Wuhan, Hubei, China

I implemented a search retrieval system using BERT, integrated into the website architecture with PyTorch, significantly improving the precision and contextual relevance of search results. I led the front-end development using Vue and optimized the back-end with Django, enhancing data flow and query efficiency. Additionally, I designed a scalable and robust database schema in MySQL, employing advanced indexing strategies and query optimization techniques to accelerate search response times.

Wuhan, Hubei, China

I led the design and development of two innovative robots: an intelligent garbage sorting vehicle and a logistics handling robot. I implemented an AI-powered sorting algorithm using PyTorch, ROS, and Jetson hardware, enabling efficient waste identification and relocation, which earned a Gold Medal (Top 1%) in the National OptoTech Innovation Contest. I enhanced object detection capabilities by refining the YOLOv5 algorithm, integrating techniques like flipping, cropping, and adaptive anchor boxes to achieve 95% accuracy in sorting tasks. Additionally, I developed a sensor fusion-based cross-calibration algorithm for real-time path-planning adaptation. I also Published 6 Chinese utility model patents.

Master'S Degree, Computer Vision, 4.06/4.0

Robotics Institute, School of Computer Science

Bachelor Of Engineering - Be, Electrical, Electronics And Communications Engineering, 3.95/4.0

Graduate from AI class, Electronic Information School National Scholarship [the highest scholarship awarded by Ministry of Education.