Admin
Members
Home
About Us
News
Member
Course
ProjectList
Publication
Demo
Links
Private
Sponsors
Honor
Research Direction
Smart Building
Image
Analysis/Restoration
AR-VR
Smart Robot
Smart Traffic
Previous Research
Research Direction
Download researches detail :
TSMC Smart Office
Goal
This system can automatically generate the parameter settings of AHU to lower the power consumption and satisfy the expected setting at the same time. Besides, this system could improve itself using reinforcement-learning.
Challenge
The extreme and rare dataset problem
Museum Guidance System
Goal
1. Achieve personalized guidance.
2. Improve the efficiency and the quality of the museum services.
Challenge
1. Indoor localization under large scenes
2. Need to consider user requirement and environmental conditions
Transfer Learning
Goal
Use a deep learning framework to learn data relationship among different domains.
Challenge
1. Lack of training data for new domain
2. Transfer label and knowledge across domains
3. The negative transfer
4. Partial transfer learning
Medical Image Analysis
Goal
Build up a deep learning model for accurate and robust segmentation ofabdominal organs on CT scan.
Challenge
1. Weak boundaries of organs, Clustering background, High appearance similarity between organ and tissue and Appearance variation caused by external factors
2. Large variation of organ size and shape through the longitudinal axis
Image/Video Compression & Restoration
Goal
Learn Deep Image Structure Prior for Ultra-Low Bit Rate Image Compression.
Challenge
1. The image quality is usually degraded at very low bitrate
2. How to find out that which necessary information should be transmitted and which prior information that deep learning model provides
Skeleton-based Human-Computer interaction
Goal
Create a motion tracking system with multiple Kinects to track the user in 360 degree (marker-free system).
Challenge
1. self-occlusion problem
2. left-right problem
Pilot training system (head pose estimation)
Goal
Estimate head position and rotation using code/marker based localization.
Challenge
1. High refresh rate estimation
2. Multi-cameras data fusion
3. Wide range operation of human head
Cleaning robot
Goal
Simultaneous Localization and Mapping (SLAM).
-Automatically build a 3D map of the mobile robot’s surroundings
-Simultaneously localize the robot
Challenge
1. Quickly extract the landmarks and build 3D map
2. Quickly estimate robot location
UAV based density estimation
Goal
Given the crowd image, we build the Deep Learning model to estimate the crowd density map and count number of people.
Challenge
1. The significant scale variation in highly congested crowd
2. The estimated density map has low resolution
Parking Lot system
Goal
Use a camera (or multi-camera) to detect the entire parking space with deep learning method.
Challenge
1. Outdoor lighting variation, inter-object occlusion and perspective distortion
2. Non-unified vehicle size and uncontrollable parking displacement
Automatic Management of Roadside Parking Spaces based on Deep Learning, Geomagnetic Sensor Networks, and LoRa Communication
Goal
Propose a well-designed deep learning networks for recognizing the sequential patterns of magnetic signals.
Challenge
1. Environment noise
2. The diversity of magnetic signals due to sensor locations
3. The interruption from environment magnetic fields
4. The variety of magnetic signals due to vehicle types
5. The interruption by moving vehicles
6. The non-unified coordination of magnetic sensors
7. The annoying magnetic responses caused by the status changing of neighboring spaces
Autonomous vehicle - lane detection
Goal
Create a robust vision-based lane detection and tracking in different scenarios.
Challenge
1. Noises from various lane marking:
Texture marking, Zebra crossing, Crossroad signs, Intersection, Curve lane
2. Shadows cast from vehicle, tree and building
Autonomous vehicle - Depth completion
Goal
Borrowing useful information from RGB image to complete the sparse depth image.
Challenge
1. Distorted and blurry edges emerge in the depth maps
2. Excessively rich texture details on color images cause undesired depth estimation results
3. Spatial-scale offset
Previous Research
Image Processing
Novel illumination-normalization method for Face Recognition
Image white balance based on a non-diagonal model
X-Ray image contrast enhancement based on tissue attenuation
HDR Compression based on Image Matting Laplacian
Computer Vision
A Cascaded Hierarchical Framework for Moving Object Detection and Tracking
Probabilistic Modeling of Dynamic Traffic Flow between Non-overlapping FOVs
Moving Targets Labeling and Correspondence over Multi-Camera Surveillance System with Ghost Suppression
Multi-Target Detection and Tracking Over Multi-Camera Surveillance System
Bayesian Hierarchical Detection Framework for Parking Space Detection
Software Development Toolkit for Video Surveillance
A Surface-based Vacant Space Detection for an Intelligent Parking Lot
Computational Photography
Wireless Communication
Multi-Dimension Kernel Density Estimation for RSS Based Indoor Localization
Dynamic Sampling Rate Adjustment for Compressive Spectrum Sensing over Cognitive Radio Network
Automatic Landmark-based RSS Compensation for Device Diversity in an Indoor Positioning System
Landmark-Based Device Calibration and Region-Based Modeling for RSS-Based Localization
Control and Sensor
Image Registration among UAV Image Sequence and Google Satellite Image Under Quality Mismatch
Track and Reconstruct 3D Instrument in Minimally Invasive Surgery by using Integrated System.
A Hybrid Method for Visitor Localization and Tracking in a Museum Environment
