MODEL BASED VISION

Intersting Lable-Free Real-time Object Pose Estimation

Outdoor Vehicle Detection | Advisor: Prof. Maxim Likhachev

Research Staff, Search-Based Planning Lab    Robotics Institute, Carnegie Mellon University

• Proposed Vehicle-PERCH, a novel 3D vehicle detection framework that can detect vehicle 3D pose through an analysis-by- synthesis manner. The algorithm effectively integrates 2D and 3D information, thus provides REAL-TIME capability

• Applied unsupervised clustering method (Gaussian mixture models) to separate vehicles into twelve categories based on vehicle size information, then constructed a dozen vehicle 3D models (microcar, sedan, compact car, SUV, etc.)

• Experimented on the KITTI dataset. The results show that Vehicle-PERCH achieves ON PAR 3D detection & localization performance with the state-of-the-art learning-based methods, WITHOUT using 3D pose annotation data

• Submitted to ICRA 2021

Indoor object 6DOF Pose Estimation |Advisor: Prof. Maxim Likhachev

Research Staff, Search-Based Planning Lab    Robotics Institute, Carnegie Mellon University

• Studied Perception Via Search (PERCH) is a class of algorithms that is first rendering scenes with different object poses, then search for the best explanation of the observed scene in the space of possible rendered scenes, thus predict the object pose while accounting for occlusion

• Studied space rotation formalisms, took advantage of the object's geometric symmetry of the object to reduce redundant space rotation proposals, thereby achieving an algorithm speedup (over 50%)

• Tested on the open dataset (YCB), results show that our algorithm SURPASSES state-of-the-art 6-DOF pose estimation methods with remarkable margins without need for any ground truth pose annotations

• [ Paper] | [ Github] | [ Full Video Link]

SYSTEM ANALYSIS

Modeling and Analysis of Complex System | Advisor: Prof. Jitesh Panchal

Master Student, Design Engineering Lab at Purdue(DELP),    ME Department, Purdue University

• Solved the difficulty of service seekers when faced with a large number of service providers, also addressed the drawbacks of the First In First Out (FIFO) matching mechanism by developed a stable matching system based on utility theory to generate the preference lists of service providers and service seekers. The matching mechanism was accomplished based on different utility interests. Searched for the optimal matching frequency using the provided matching.

PROJECTS

Face Detection on Depth Image

• Independently Implemented a modified version of the multi-task cascade CNN model based on Caffe and Python from scripts, which solved the problem of facial detection on completed depth images; optimized the model to obtain 99.93% precision and over 97% recall.

• [GitHub]

Automonous Driving Robot

• A self-designed and manufactured, fully autonomous robot that can retrieve multiple randomly located tennis balls and put the tennis balls back to a random located container.

• I in charge of the computer vision part. Applied OpenCV (installed on Raspberry Pi) to locate and track the tennis balls based on color and outline shape..

• Overview of Robot