Autonomous-Driving

Autonomous RC car project using Python3 on raspberry pi 3

Problems

• 서보 모터에서 발생하는 떨림 => PWM of Raspberry Pi is the structure that the software creates, so the DUTY value that determines the rotation angle of the servo shakes.
• 서보 모터에서 저전력이 발생했다는 경고가 자주 뜸
• 라즈베리파이의 처리속도 한계로 인한 낮은 FPS
• Driving 과 Object detection을 동시에 수행하기에 부족한 Raspberry pi의 자원
• 훈련을 위한 Track이 부족하다 / 제작하기엔 예산이 부족

Schedule

• Calander

April

May

Overview

• Use raspberry pi 3
• Use 2 motors 1 DC motor and 1 Servo motor
• Behavior cloning we are inspired by DAVE-2 System
• Object Detect we use YOLO algorithm v3
• Remote control control rc car using dualshock4 (here driver => ds4drv)

RC car Structure

Motor Detail

MDD10A

• M1A, M1B : Output motor 1
• M2A, M2B : Output motor 2
• B+, B- : PowerInput Max 25V, 10A
• PWR : Green LED, Power
• M1A, M1B: Test Button Motor1
• M2A, M2B: Test Button Motor2

Input Pins

• DIR1 : Direction input(motor1). low(0 - 0.5v), high(3 - 5.5v)
• PWM1 : PWM input for speed control (Motor 1). Max 20Hz
• DIR2 : Direction input(motor1). low(0 - 0.5v) , high(3 - 5.5v)
• PWM2 : PWM input for speed control (Motor 2). Max 20Hz
• GND : Ground

PWM & DIR control

-	Input	DIR	Output A	Output B
PWM	off	X	off	off
PWM	on	off	on	off
PWM	on	on	off	on

Autonomous Driving

End to End Learning for Self-Driving Cars

Block Diagram of DAVE-2

Neural Network Architecture

• source
  

The network consists of 9 layers, including a normalization layer, 5 convolutional layers and 3 fully connected layers. The input image is split into YUV planes and passed to the network. They use strided confolutions in the first three convolutional layers with a 2x2 stride and a 5x5 kernel and a non-strided convolution with a 3x3 kernel size in the last two convolutional layers.

Model Summary

Layer (type)	Output Shape	Param #	Units	Kernel size	Activation
conv2d_1(Conv2D)	(None, 31, 98, 24)	1824	24	5, 5	relu
conv2d_2(Conv2D)	(None, 14, 47, 36)	21636	36	5, 5	elu
conv2d_3(Conv2D)	(None, 5, 22, 48)	43248	48	5, 5	elu
conv2d_4(Conv2D)	(None, 3, 20, 64)	27712	64	3, 3	elu
conv2d_5(Conv2D)	(None, 1, 18, 64)	36928	64	3, 3	elu
flatten_1(Flatten)	(None, 1152)	0
dense_1(Dense)	(None, 100)	115300	100	-	elu
dense_2(Dense)	(None, 50)	5050	50	-	elu
dense_3(Dense)	(None, 10)	510	10	-	elu
dense_4(Dense)	(None, 1)	11	1	-

Total params: 252,219
Trainable params: 252,219
Non-trainable params: 0

• Optimizer: Adam
• Learning rate: 1e-4
• loss function: MSE
• Batch size: 100
• Steps per epoch: 200
• epochs: 10

Simulation

Detail for simulation
Player plays the simulation game manually and records car's information. Use 3 cameras(left, center, right)

Preprocessing

Original Image	Preprocessed Image

def img_preprocess(img):
  img = img[60:135, :, :] # height, width, channel # 카메라에 투영되는 차량의 앞부분을 제거한다
  img = cv2.cvtColor(img, cv2.COLOR_RGB2YUV) # recommended(nvidia model)
  img = cv2.GaussianBlur(img, (3, 3), 0)
  img = cv2.resize(img, (200, 66))
  img = img/255  
  
  return img

Data Augment

• imgaug is a library for image augmentation in machine learning experiments. It supports a wide range of augmentation techniques, allows to easily combine these and to execute them in random order or on multiple CPU cores, has a simple yet powerful stochastic interface and can not only augment images, but also keypoints/landmarks, bounding boxes, heatmaps and segmentation maps.

Results