Capstone/LightControl

1. Purpose of project

• Object Detection using Deep Learning Model (YOLOv5)
• Training YOLOv5 with custom dataset
• Object location coordinate transformation ( image > Work space )
• Arduino control using Object location

2. Deep learning model training

• we creating cutom dataset using roboflow

• Training YOLOv5-s model using colab
• YOLOv5 Github Link : https://github.com/ultralytics/yolov5

%cd /content

!git clone https://github.com/ultralytics/yolov5.git
!python train.py --img 416 --batch 16 --epochs 100 --data /content/dataset/data.yaml --cfg ./models/yolov5s.yaml --weights yolov5s.pt 

• real-time object detection

3. Project Process

[overview]

---

[Coordinate transformation]


To control robot arm, we need to coordinate (theta1, theta2)

So we converted object location coordinate(pixel) within the image to (theta1, theta2)






[Web cam to Real place]

	Web cam[px]	Real Value[cm]	Scale factor[px/cm]
1	(245, 270)	(26.5, 20)	(0.1082, 0.0741)
2	(195, 345)	(21.5, 25)	(0.1102, 0.0724)
3	(297, 198)	(31.5, 15)	(0.1060, 0.0758)
4	(135, 198)	(16.5, 15)	(0.1222, 0.0758)

So we get average scale factor (0.1116, 0.0745)



[ Real place to Polar]



[Polar to Arduino]

	Radius(r)	Theta2(degree)
1	52.5	92
2	48.5	93
3	15.5	94
4	41.5	95
5	39.5	96
6	35.8	97
7	30.5	100
8	29.5	101
9	26.5	103
10	23.3	106
11	20.0	109
12	17.7	112
13	14.5	117
14	11.0	125
15	9.3	130
16	7.4	135
17	4.7	145
18	2.3	155
19	0.0	175

import numpy as np

# pixel to cm
real_x = (pixel_x) * 0.11165
real_y = (pixel_y) * 0.0745

# cartesian to polar
R = (real_x**2 + real_y**2)**0.5 - 6.5
theta1 = 115 - np.degrees(np.arctan(real_y/real_x))

# r to theta2 (Regression)
if R >= 26.5:
   # 4 order polynomial regression
   theta2 = (-0.0000001)*(R**4) - (0.0004)*(R**3) + (0.0611)*(R**2) -3.1901*(R) + 152.57
else:
   # 3 order polynomial regression
   theta2 = (-0.0012)*(R**3) + (0.1378)*(R**2) -5.3419*(R) + 169.53

4. Arduino control

5. Final Result