Yu-Chi Lin
林友騏
COURSE PROJECTS
Waiter Robot
The waiter robot is a prototype robot to acting like a waiter in the restaurant. It can take orders and deliver dishes in a pre-defined restaurant environment. The robot is equipped with a Kinect to detect the table by feature matching. We utilize ITRI's speech recognition system to take orders from the customer. Small robot arms are mounted on the top of the robot to grasp bowls from the top platform of the robot. The robot is capable of patroling inside the restaurant and avoiding obstacles in the restaurant by sonar feedback.
3D Unorganized Point Cloud Based Segmentation for Unknown Objects
The project aims to segment unknown objects from object clusters on a table based on unorganized point cloud data. Compared to algorithms based on organized point cloud, which most existing 3D segmentation approaches belong to, the our approach has advantages in dealing with data fusion from multiple data frame or data sources, and make sensor choices more versatile.
The proposed approach first over-segments the point cloud into several surface patches by supervoxel algorithm and construct a graph of these patches. For every observed surface patch, there should exist part of object volume behind the surface that supports it. Since each 3D position is uniquely occupied by an object (or no object), the measurement of the overlapping volume behind the surface patch provides distinguishing feature for segmenting objects. Following this idea, a graph-based segmentation algorithm with the overlapping volume measurement is applied to every patch, and then to every object hypothesis to obtain the final result. The proposed approach is free from manual seeding, which is suitable for practical application. [Report, PDF]
Age Estimation based on Bio-inspired Features and SVM
The age estimation system is based on the approach proposed by Guo et al. in 2009. The input of the estimation system is a face image. We extracted the feature with Gabor filter to collect texture information of the face image in various scales and directions. We divided human age into 6 groups and classified each face image into these groups. We reduced the dimension of feature vector with PCA and trained an SVM to classify the age group of the face. The accuracy of the age estimation system reaches 92%.