sean m. ficht. problem definition previous work methods & theory results
TRANSCRIPT
Sean M. Ficht
Problem Definition
Previous Work
Methods & Theory
Results
Track and follow specific person with a mobile robot
Cluttered environments
Brief occlusion
Long occlusion
Cooperative user
Helper robot
Carry items for a person• Example: Hospital situation
Problem Definition
Previous Work
Methods & Theory
Results
Person following with a mobile robot
• Appearance based
• Optical flow based
• Stereo vision based
Segmentation of image
Classification
Detection
Limitations
Sidenbladh, Kragic, and Christensen; ICRA; 1999Tarokh and Ferrari; Journal of Robotic Systems; 2003Schlegel, Illman, Jaberg, Schuster, and Worz; British Machine Vision Conference; 2005
Calculate optical flow
Use to segment image
Limitations
Chivilo, Mezzaro, Sgorbissa, and Zaccaria; IROS; 2004Piaggio, Fornaro, Piombo, Sanna, and Zaccaria; IEEE ISIC/CIRA/ISAS joint conference; 1998
Find features
Segment from background
Use to track
Limitations
Zhichao and Birchfield; IROS; 2007
Problem Definition
Previous Work
Methods & Theory
Results
Kinect
Provides a depth image
Provides a RGB color image
Packaged solution
Detection and Tracking• Generic detector
• Specific appearance model
• Integrating particle filter
Robot Control
HOG person detector (OpenCV)
• HOG descriptoro Cells -> Block -> Windowo 4 cells in a blocko 105 blocks in a windowo 64x128 window
• Training
Dalal and Triggs, CVPR, 2005
Gradient of the Image
Binning of pixels in cells
Grouping of cells into blocks
Normalization
Kernel convolution
• Magnitude = (gx2 + gy
2)
• Angle = arctan(gy/gx)
Directional change in intensity
Bins apply to each cell
Nine separate bins
Gradient magnitude added to bin
Cells grouped into blocks
4 cells per block
Blocks overlap one another
HOG person detector
• HOG descriptoro Cells -> Block -> Windowo 4 cells in a blocko 105 blocks in a windowo 64x128 window
• Training
Support Vector Machine (SVM) classifier
Binary classifier
Trained on images
Detection and Tracking• Generic detector
• Specific appearance model
• Integrating particle filter
Robot Control
Color Histogram
Segmentation by depth to create template
Represents distribution of colors
10 bins for each color channel• 1000 element color histogram
Pixel classification
2 bin example• Bin 1: 0-127• Bin 2: 128-255
Average depth
Threshold (0.3 meters)
Template used to make color histogram
Detection and Tracking• Generic detector
• Specific appearance model
• Integrating particle filter
Robot Control
System State
Motion model
Observation model
Expected state
Resample
Hybrid state space
X and Y in image coordinates• Scaled according to depth
Z in depth coordinates
Detection and Tracking• Generic detector
• Specific appearance model
• Integrating particle filter
Robot Control
Input: tracking information from tracking algorithm
Uses tracking information to make movement decisions
Executes movement and returns to tracking algorithm
Problem Definition
Previous Work
Methods & Theory
Results
No occlusion
• Other people present (different depth)• Other people present (similar depth)• Pose change
Brief occlusion
Long occlusion
Initial template
Initial template Non-occluded target Occluded target
Average between 73% and 74%
Problem• Follow a person in different scenarios
System• RGB-D sensor• Generic detector• Specific appearance model• Particle filter• Robot control architecture
Performance• Performed in three separate test scenarios• Rapid side to side target motion trade-off• Large target scale changes
Train a new HOG detector to handle scale issues
Using more particles
KLT features for trajectory histories
Adaptive appearance model