FUEL CAP DETECTION AND POSE ESTIMATION CSCI 512 COMPUTER VISION YI HSIU HUNG 05/04/2016

Transcription

1 FUEL CAP DETECTION AND POSE ESTIMATION CSCI 512 COMPUTER VISION YI HSIU HUNG 05/04/2016

2 INTRODUCTION SELF DRIVING GOOGLE CAR Self Driving Google Car: Self Drive over 1.5 million miles 2014 early construct of a prototype Delivered the first real build of prototype vehicle in December 2014 If we need the cars fully automatic, we need the cars to be charged or fueled automatically.

3 INTRODUCTION WIRELESS CHARGING TECHNOLOGY Wireless charging: 2015 Hevo Power s prototype 1.5 kw (Goal to reach 200 kw)

4 INTRODUCTION TESLA S ROBOTIC SNAKE CHARGER Tesla single Charger: 10 kw (29 miles range per hour) Tesla Dual Chargers: 20 kw (58 miles range per hour)

5 DESIGN A video simulating a car driving into a fuel station Fuel Cap Detection Circular Fuel Cap Extract features Eliminate noise Fuel Cap Detection Rectangular Fuel Cap Extract features Eliminate noise Pose Estimation Cartesian coordinates

6 FUEL CAP DETECTION CIRCULAR FUEL CAP First Trial Hough Transform imfindcircles function rgb2gray imadjust (Gamma = 4, To enhance the contrast of high intensity)

7 FUEL CAP DETECTION CIRCULAR FUEL CAP First Trial Hough Transform Hough Transform Detect Not detect (%) Circular Fuel Cap There True Positive Rate 73.12% Total Frames: 279 Not There False Positive Rate 1.79% (still have some false detections)

8 FUEL CAP DETECTION RECTANGULAR FUEL CAP First Trial Binary Image Processing bwlabel function (To label blobs) Canny edge detection regionprops function (To extract features)

9 FUEL CAP DETECTION RECTANGULAR FUEL CAP First Trial Binary Image Processing (Rectangular Fuel Cap) Binary Image Processing Detect Not detect (%) Rectangular Fuel Cap There True Positive Rate 81.01% Total Frames: 107 Not There False Positive Rate 0.00% First Trial Hough Transform (Circular Fuel Cap) Hough Transform Detect Not detect (%) Circular Fuel Cap There True Positive Rate 73.12% Total Frames: 279 Not There False Positive Rate 1.79%

10 EXTRACT FEATURES regionprops function Boundingbox Centroid Extrema MajorAxisLength MinorAxisLength Orientation imhist function Mean Standard Deviation Range value Uniformity (Circular Fuel Cap) (Rectangular Fuel Cap) (imhist)

11 NOISE ELIMINATION X Y Ratio Area Mean and Standard Deviation

12 NOISE ELIMINATION New Centroid in Old Bounding Box Separate Video into Two Parts

13 FUEL CAP DETECTION Second Trial Binary Image Processing + Extract Features (Rectangular Fuel Cap) Binary Image Processing Detect Not detect (%) Rectangular Fuel Cap There 79 0 True Positive Rate % Total Frames: 107 Not There False Positive Rate 0.00% Second Trial Binary Image Processing + Extract Features (Circular Fuel Cap) Hough Transform Detect Not detect (%) Circular Fuel Cap There True Positive Rate % Total Frames: 279 Not There False Positive Rate 0.00%

14 POSE ESTIMATION Problems of My Video No Dimension (No P_m) Changing Focal Length (No K)

15 POSE ESTIMATION Y (Rotate Y axis ) (Rotate Z axis ) Second Trial My Way Using Length proportion to find R = Rz*Ry*Rx Z X Y X az ay X

16 MY RESULTS Circular Fuel Cap Video Rectangular Fuel Cap Video

17 LIMITATIONS Lighting and Reflection Distance Change

18 LIMITATIONS Pose Illusion Eliminate noise setting X Y ratio Area Mean Standard Deviation?

19 CONCLUSION & FUTURE WORK Lighting is a factor which needs to be controlled. Fixing the focal length will make it easier to detected the fuel cap. Need a second camera or a depth sensor for better pose and depth estimation. Increase the fuel cap features, i.e. Open the fuel cap. Try take video with fixed focal length and known dimension. Try different shapes of fuel cap.

20 QUESTIONS

21 BACKUP Extrema Function MajorAxisLength MinorAxisLength Orientation

22 FUEL CAP DETECTION RECTANGULAR FUEL CAP First Trial Hough Transform imfindlines function (No Fuel Cap was detected) rgb2gray imadjust (Gamma = 4, To enhance contrast of high intensity) (reduce Threshold, increase FillGap, decrease MinLength)