21
A Hybrid Multi-Object Tracking System Heaven's Light is Our Guide Rajshahi University of Engineering and Technology Department of Computer Science and Engineering Presented by Md. Minhazul Haque Roll # 103001 Dept. of CSE RUET Supervised by Md. Arafat Hossain Assistant Professor Dept. of CSE RUET October 13, 2015

Multi Object Tracking | Presentation 2 | ID 103001

Embed Size (px)

Citation preview

Page 1: Multi Object Tracking | Presentation 2 | ID 103001

A Hybrid Multi-Object Tracking System

Heaven's Light is Our Guide

Rajshahi University of Engineering and TechnologyDepartment of Computer Science and Engineering

Presented byMd. Minhazul HaqueRoll # 103001Dept. of CSERUET

Supervised byMd. Arafat Hossain

Assistant ProfessorDept. of CSE

RUETOctober 13, 2015

Page 2: Multi Object Tracking | Presentation 2 | ID 103001

Outlines❏Objectives❏Overview of Object Tracking❏Existing Works❏Proposed Methodology❏Implementation❏Experiment Result❏Performance Analysis❏Future Work❏References

2

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 3: Multi Object Tracking | Presentation 2 | ID 103001

Objectives❏Create a Multi Tracking Method❏Better background subtraction method❏Reduce noise in input videos❏Minimize the error rate

3

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 4: Multi Object Tracking | Presentation 2 | ID 103001

Object

4

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

A bird

A car

Image Courtesy: 4freephotos

❏Object is a group of pixels with similar property, also called a blob

❏Anything can be an Object, A ball, car or bird

Page 5: Multi Object Tracking | Presentation 2 | ID 103001

Object Tracking

5

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

❏Locate Objects over time❏Save Object List into memory❏Set unique ID to each Object❏Loop until media/input ends

Page 6: Multi Object Tracking | Presentation 2 | ID 103001

Existing Works❏Tracking and Motion Estimation[1]

❏ Developed by MATLAB❏MOT Challenge 2015[2]

❏ Developed by❏ University of Adelaide Research Centre❏ Swiss Federal Institute of Technology, Zürich❏ TU Darmstadt

❏OpenCV Blob Tracker[3]

❏ Deprecated

6

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 7: Multi Object Tracking | Presentation 2 | ID 103001

Proposed Methodology

❏ A Hybrid System that contains❏ Background Subtraction Method❏ Noise removal method❏ Blob tracker method

❏Popular Tracker Algorithms❏ Kalman Filter❏ CamShift Algorithm❏ Optical Flow (Not used right now)

7

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 8: Multi Object Tracking | Presentation 2 | ID 103001

MOT: Block Diagram

8

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Start Initialize source media

Apply BGS

Apply Contour DetectionGet Object List

Track Objects

Update Objects

Delete Objects

Add Objects

Streamof frames

Get a frame

Loop untilend of media/frame

Page 9: Multi Object Tracking | Presentation 2 | ID 103001

Implementation❏Language: C++❏Framework: OpenCV 2.4.10❏IDE: Qt Creator❏Platform: Ubuntu 14.04 LTS

Data Resources❏MOT Challenge Website[2]

❏HSCC1000 Highway Surveillance Videos[4]

❏RUET Area Videos (Noisy)

9

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 10: Multi Object Tracking | Presentation 2 | ID 103001

Initialize Source Media❏Loaded video files using OpenCV

VideoCapture class

Apply BGS Method❏Tested the following methods

❏ GMM (Gaussian Mixture Model)❏ MOG2 (Mixture of Gaussian version 2)❏ ViBe (Used by American Navy, Closed Source)

Implementation [2] Initialization

10

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 11: Multi Object Tracking | Presentation 2 | ID 103001

Implementation [3] BGS Comparison

11

MOG2Perfect

GMM/MOGNoisy

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 12: Multi Object Tracking | Presentation 2 | ID 103001

Implementation [4] Noise Removal

12

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Noise Removal Techniques

❏ Apply medianBlur filter❏ Convert image to Binary

from Grayscale to remove shadow❏Apply erode and dilate technique

Image Courtesy: mimage.me, slice.org

Page 13: Multi Object Tracking | Presentation 2 | ID 103001

Implementation [5] Extract Blob Info

13

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

❏Apply Canny Edge filter❏Apply Contour detection using findContours❏Fit detected points into a boundingRect❏Store the following blob information into

memory❏ Center❏ Size❏ Angle/Orientation❏ Histogram❏ ID

Image Courtesy: opencv-srf.blogspot.com

Page 14: Multi Object Tracking | Presentation 2 | ID 103001

Implementation [6] Process Blobs

14

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

❏Match blob features and update properties❏ Match Histogram (CamShift)❏ Position (Kalman Filter)❏ Orientation (Fit Ellipse)❏ Motion/Trajectory

(Use previous position history, not implemented)❏ Screen position (Viewport edges)

❏Assign incremented ID to new objects❏Remove objects that are out of the screen

Page 15: Multi Object Tracking | Presentation 2 | ID 103001

Implementation [7] Result

15

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Time 00:00:11.3 Time 00:00:12.8

Page 16: Multi Object Tracking | Presentation 2 | ID 103001

Experiment Result

Result of Proposed System

16

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Videos Duration MOT

Highway_Surveillance_1.mp4 01:00 138

Highway_Surveillance_2.mp4 01:00 179

RUET_Area_1.mp4 00:49 20

Talaimari_Area_1.mp4 02:30 69

Page 17: Multi Object Tracking | Presentation 2 | ID 103001

Performance Analysis

Comparison of Object Tracking Result

17

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Videos Duration Actual Objects

Kalman Filter CamShift MOT

Highway_Surveillance_1.mp4 01:00 131 83 120 138

Highway_Surveillance_2.mp4 01:00 155 96 106 179

RUET_Area_1.mp4 00:49 24 21 23 20

Talaimari_Area_1.mp4 02:30 59 63 45 69

Page 18: Multi Object Tracking | Presentation 2 | ID 103001

Performance Analysis: Bar Chart

18

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Figure: Bar Chart using the data of multiple tracking methods

Number of Objects Detected

Page 19: Multi Object Tracking | Presentation 2 | ID 103001

Future Work❏Implement Optical Flow algorithm❏Calculate trajectory❏Create a GUI❏Export data to Excel and CSV format

19

Page 20: Multi Object Tracking | Presentation 2 | ID 103001

References

1.http://www.mathworks.com/help/vision/tracking-and-motion-estimation.html

2.https://motchallenge.net3.https://code.google.com/p/cvblob/4.https://www.youtube.com/user/HSCC1000

20

A Hybrid Multi-Object Tracking SystemOctober 13, 2015

Page 21: Multi Object Tracking | Presentation 2 | ID 103001

Thank you!

Any question?

21

A Hybrid Multi-Object Tracking SystemOctober 13, 2015