IRJET- Autonomous Underwater Vehicle: Electronics and Software Implementation of the Proton AUV

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 06 | June 2019 www.irjet.net p-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 3433
Autonomous Underwater Vehicle: Electronics and Software
Implementation of the Proton AUV
Vivek Mange1, Priyam Shah2, Vishal Kothari3
1Student, Dept. of Electronics and Telecommunication Engineering, K J Somaiya College of Engineering,
Mumbai, India.
2,3Student, Dept. of Computer Engineering, K J Somaiya College of Engineering, Mumbai, India.
----------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - The paper deals with the software and the
electronics unit for an autonomous underwater vehicle. The
implementation in the electronics unit is the connection and
communication between SBC, pixhawk controller and other
sensory hardware and actuators. The major implementation
of the software unit is the algorithm for object detection
based on Convolutional Neural Network (CNN) and its
models. The Hyperparameters were tuned according to
Odroid Xu4 for various models. The maneuvering algorithm
uses the MAVLink protocol of the ArduSub project for
movement and its simulation.
Key Words: Autonomous Underwater Vehicle, Odroid
Xu4, Neural Network, Tensorflow Object Detection,
Pixhawk, Hyperparameter, ArduSub, MAVLink.
1. INTRODUCTION
Recent technological advancement and interest of people
towards the Single Board Computers have uplifted and
proved to be of great usefulness in the robotics and drone
industry. With the cutting edge technology for SBCs and
the simulation of the AUVs has led to various emerging
communities towards its development such as ArduPilot,
being one of the popular ones which comprise of
subdirectories like ArduSub, ArduCopter, etc.
Our System consists of Odroid Xu4 as a SBC, which acts as
a brain for controlling and communication. The Pixhawk is
connected to Odroid via Serial communication and act as a
flight controller for maneuvering and simulation. The
Tensorflow Object detection API, along with SSD
Mobilenet v2 model helped to achieve object detection
underwater. The hyperparameter such as learning rate,
batch size, number of epochs, etc was tuned according to
practical implementation on Odroid. The motion of the
system underwater was executed with the help of python
package Pymavlink. The simulation was carried out in the
SITL and QGC software before real-time testing. This
software provided the data of sensors like gyrometer,
accelerometer, etc.
Underwater vehicles and technologies like AI have caught
people's eye towards this field. Research and development
in this field have led to various emerging applications such
as cleaning water robots, oil excavation, etc.
2. SINGLE BOARD COMPUTER (SBC)
Single Board Computers (SBCs) are high-speed micro-
controllers. The SBC is responsible for manipulating
different aspects of the system and process the
information. Odroid Xu4 and Raspberry Pi 3B are two of
the processors on which we worked. Selection of a better
processor is based on the processing speed, GPU, external
ports, etc for successfully completing the vital tasks within
a given time period.
2.1 Raspberry Pi 3B
Raspberry Pi 3B is one of the latest models in the
raspberry series. RPi 3B is a 1.2Ghz quad-core processor
with Broadcom BCM2837 SOC 64bit CPU for fast
processing. RPi 3B has a 1GB LPDDR2 RAM as a storage
unit for faster data rate with low power consumption. It
has an ethernet connection with 100 Mbps transmission
but does not allow to power over Ethernet (PoE). The
power consumption of RPi 3B is around 260mA in idle
condition and up to 730mA, when processed at 400% CPU
load. RPi has inbuilt wireless LAN support specified under
IEEE 802.11b/g/n band and a Bluetooth module. It has
few external ports such as 1x Full-size HDMI port, 4x USB
ports with USB 3.0 and an external port for connecting an
RPi camera. It has a Micro SD card slot for loading the
operating system and storing the data.[4]
2.2 Odroid Xu4
Odroid Xu4 is the latest generation SBC. It possesses a
faster processing speed with dual quad-core processor.
The two processors are Samsung Exynos5422 CortexTM-
A15 2Ghz and CortexTM-A7 CPUs with additional GPU
support of Mali-T628 MP6. It comprises of 2GB LPDDR3
RAM for storage and high data rate. It has a Gigabit
Ethernet to possess a high-speed data transmission.
Odroid Xu4 requires high power input up to 5V/4A while
initial running, which leads to a high amount of heat
dissipation, and to overcome this issue we have an
‘External FAN’ or a ‘Heat Sink’ install on the board. It lacks
in the in-built assembly of any WiFi or Bluetooth system. It
has 1x HDMI display port, 2x USB 3.0 and 1x USB 2.0. It
also possesses an Emmc 5.0 module for flash storage.[5]

2.3 Analysis of Electronics
Initially, We used RPi 3B as our main processor to
perform various tasks such as Maneuvering, Image
processing, and Acoustic communication. The RPi was
integrated with Pixhawk to command the thrusters for any
particular motion resulting in the maneuvering of the
system.
Fig-1 : Architecture of Proton
Image Processing that comprises of the object detection
system faces few issues while working on RPi as the delay
in the output were, fewer frames per second ratio(FPS)
and increase in the system temperature. To overcome this
issue we upgraded our system to Odroid Xu4 which helped
in debugging the issues of the FPS up to 30 times faster
than RPi. Odroid surpasses RPi in processing speed as well
by providing approximately 7 times faster processing
power. Odroid Xu4 is integrated with Pixhawk using
MAVLink protocol and the complete architecture is shown
in (Fig-1). An external WiFi module and an O-CAM is
connected to perform the necessary tasks.
3. OBJECT DETECTION
Object detection is technology related to computer vision
and image processing that deals with detecting instances
of semantic objects of a certain class in a digital image.
Object detection is more powerful than classification, it
can detect multiple objects in the same image. It also tags
the objects and shows their location within the image.[2]
There are various techniques to detect an object in an
image, one of them is using neural networks. We have
used CNN to detect objects in the image.
Fig-2: Underwater Object Detection
3.1 Convolutional Neural Network (CNN)
The convnet is a special kind of neural networks which
contains at least one convolutional layer. A general
convnet structure will take an image, then pass it through
a series of convolutional layers, followed by pooling and a
fully connected layer to output the classification labels.
A neural network consists of several different layers
such as the input layer, at least one hidden layer, and an
output layer. Neural networks are best used in object
detection for recognizing patterns such as edges, shapes,
colors, and textures. The hidden layers are convolutional
layers in this type of neural network which acts like a filter
that first receives input, transforms it using a specific
pattern/feature, and sends it to the next layer. With more
convolutional layers, each time a new input is sent to the
next convolutional layer, they are changed in different
ways (fig-3). For example, identifying a blue car, from a
given dataset consisting images of cars and bikes, in the
first convolutional layer, the filter may identify shape in a
region to find a car, and the next one may be able to
conclude the color of the car, and the last convolutional
layer may classify the car as blue car (Fig-3). Basically, as
more and more layers the input goes through, the more
sophisticated patterns the future ones can detect.
Fig-3 : CNN

There are two types of deep neural networks. Base
network and detection network. MobileNet, VGG-Net,
LeNet are few examples of base networks. Base network
provides high-level features for classification or detection.
If you use a fully connected layer at the end of these
networks, you get an image classifier. But you can remove
the fully connected layer and replace it with detection
networks, like SSD, Faster R-CNN, and so on.
3.1.1 Faster R-CNN v2
R-CNN stands for Region Convolutional Neural
Network. In the object detection, an R-CNN is used to find
regions since determining the location of multiple objects
is essential to this type of model. An image is split into a lot
of different boxes to check if any of them have signs of the
object. The algorithm uses region proposal networks
(RPN) which ranks the specific regions that are most likely
to have the object and this is done using classifier and a
regressor (fig-4). A Classifier determines the probability of
a proposal having the target object. Regression regresses
the coordinates of the proposals.[1]
Performance and precision for an object detection task
using Faster R-CNN v2 model, scores over 28 mAP (mean
Average Precision) at 58 ms speed on COCO dataset.
Fig-4: Faster R-CNN
3.1.2 Single Shot MultiBox Detector (SSD)
MobileNet v2
Single Shot: this means that the tasks of object
localization and classification are done in a single forward
pass of the network.
MultiBox: this is the name of a technique for bounding
box regression developed by Szegedy et al.
Detector: The network is an object detector that also
classifies those detected objects, (fig-5).
Performance and precision for an object detection task
using SSD MobileNet v2 model, scores over 22 mAP (mean
Average Precision) at 31 ms speed on COCO dataset.
Fig-5: SSD Mobilenet
The Processing phase is faster in SSD MobileNet v2 model
over the Faster R-CNN v2 model in Odroid Xu4.
4. HYPERPARAMETER
Hyperparameters are the variables which determine the
network structure and how the network is trained.
Hyperparameters are set before training and before
optimizing the weights and bias for the layers. The
training parameters we used were found by practical
implementation and tuning of the model. [1]
4.1 Learning Rate
Learning rate controls the updation of the weight in the
optimization algorithm. Learning rate can be fixed
learning rate, gradually decreasing learning rate,
momentum-based methods or adaptive learning rates,
depending on our choice of optimizers such as Adam,
Adagrad, SGD, Ada Della or RMSProp.
Our Training parameters : a) Faster R-CNN v2 model, we
used manual_step_learning_rate with initial_learning_rate
as 0.0002. b) For SSD Mobilenet v2 model we used
exponential_decay_learning_rate with initial_learning_rate
as 0.004.
4.2 Number of epochs
The number of epochs is the number of times the
entire training set pass through the neural network. The
variation in epochs can be seen by increasing the number
of epochs until we see a small gap between the test error
and the training error.
Our Training parameters: The number of epochs = 1 for
both models.
4.3 Batch Size
Convnet usually prefers small batch size in the learning
process. A batch size of 1 to 3 is a good choice to test with
Odroid. The convnet is sensitive to batch size.
Our Training parameters: a) Faster R-CNN v2 model, we
used batch_size: 1 with momentum optimizer. b) SSD

Mobilenet v2 model, we used batch_size: 1 with
rms_prop_optimizer.
4.4 Activation Function
Activation function introduces non-linearity to the
model, which is required for complex functional mappings
between the inputs and response variable. Usually,
rectifier works well with convnet. Other alternatives are
sigmoid, tanh and other activation functions depending on
the task.
Our Training parameters: a) Faster R-CNN v2 model, we
used softmax/logistic regression. b) SSD Mobilenet v2
model, we used RELU_6.
4.5 Weight Initialization
To prevent dead neurons we have to initialize some
weight which can be a small random number, but not too
small as to avoid zero gradients.
Our Training parameters : a) Faster R-CNN v2 model, we
used weight:0.0. b) SSD Mobilenet v2 model, we used
weight:0.00004
4.6 Dropout for regularization
To avoid overfitting in deep neural networks, we can
use some regularization technique and most preferable is
the dropout technique. The method simply drops out units
in a neural network according to the desired probability or
threshold.
Our Training parameters : a) Faster R-CNN v2, we had
dropout_keep_probability: 0.1. b) SSD Mobilenet v2, we
had dropout_keep_probability: 0.8.
5. MANEUVERING
Maneuvering is the movement of the system. There are 5
degrees of freedom available in the System, that provides
movements in the respective direction as roll, yaw, heave,
dive, sway and surge (Fig-6). We maneuver our system
using a python script with the MAVLink protocol
functions. The script is uploaded on the Odroid Xu4 and
with the integration of Pixhawk, the commands to the
particular thrusters are provided with the PWM values.
The Pixhawk is connected to ESCs which intakes the input
in the form of PWM pulses and the output of ESCs are
connected to thrusters to provide the required thrust.
Fig-6: Maneuvering Movement
5.1 MAVLink
MAVLink (MICRO AIR VEHICLE COMMUNICATION
PROTOCOL) is a binary telemetry protocol. It is specifically
designed to establish communication with drones and
between its onboard components. MAVLink protocol was
designed to create a very lightweight messaging protocol,
which used limited system resources and bandwidth.
Telemetry data streams delivers a specific message
between the Odroid and the pixhawk controller. MAVLink
is used in number of autopilots, ground stations and
integration APIs.[7]
PyMavlink is a python library for handling ArduSub
vehicle communication and maintaining log files using the
MAVLink protocol. A python script is used to read the data
from different sensors and send commands to an ArduSub
vehicle.[7]
5.1.1 ArduSub
ArduSub is a sub-project under ArduPilot for
underwater projects. It uses the MAVLink protocol
,multiple tools and packages for allowing one to control
the system via python scripts based on certain logical
interpretation of working environment. ArduSub allows
custom package generation as per the developer [6]. Also,
adjustment of parameter values can be done based on
different geographical variation and application use.
5.1.2 Simulator (SITL & QGC)
ArduSub project comes with a Software in the Loop
(SITL) simulator to simulate the ArduSub vehicle. SITL
allows us to test the code without any hardware
implementation and to analyse the desired output.
ArduSub along with QGroundControl also offers us GUI
support for users on different platforms (Fig-7) [6].

Fig-7: Simulation of System using QGC
6. PRACTICAL IMPLEMENTATION
The overall implementation was carried out using Odroid,
Pixhawk, electronic equipment and Vehicle.
6.1 System Connection and Communication
The Odroid Xu4 with Ubuntu Mate 16.04 was
connected through UART0 port to the Telemetry port of
Pixhawk (Fig-8). Wifi Module 0 was used to communicate
between Odroid and the surface computer. This
communication is possible in two ways i) Wired
connection ii) Wireless connection. The Wired connection
is established through the ethernet connection using
Fathom X on both the end as it is used to provide the GUI
support on QGC. The wireless connection is established
using Putty software. The Pixhawk was configured using
QGroundControl software along with Ardusub. The
communication between both the devices was established
using the python package Pymavlink. The oCam: 5MP is
connected to the USB 3.0 of Odroid, which is used for
image and video capturing. It provides the frame rate of
1280×720@30fps with 65 degrees Field of vision (FOV).
Fig-8: Connections
6.2 Training Phase Implementation
The Tensorflow object detection code along with
training models and labeled underwater images were
trained on a laptop having a graphics card of NVIDIA GTX
1050 Ti [8]. The Faster R-CNN v2 and SSD Mobilenet v2
models were trained on COCO dataset with thousands of
classes. Then the models were retrained with 2 new
classes namely gate and flare using a large dataset of
underwater images respectively. Retraining of previously
achieved checkpoints was continued to achieve better
results. Using trial and error the inference graphs were
generated for each model, whereby the hyperparameters
were tuned for various batch size and tested on Odroid.
6.3 Testing Phase Implementation
The inference graph along with required code to test
was copied on the SD card inserted in Odroid. We took
input as Video capture using OpenCV library and oCam
camera. Then passed it to our main script for object
detection. If the tensorflow detects gate/flare above a
particular threshold of 75% then it creates a box around
the image with the label of detected object and detection
score.[Fig-2] The 4 corner coordinates of this box are then
used to find the center of the detected image. The center of
the main Video frame is also taken and is subtracted from
the detected image center which provides us whether the
detected object is on the Left-hand side, Right-hand side,
Above, Below or Exact front. Once we have the direction,
we now communicate with Pixhawk to forward the
required PWM pulses to the ESC. We have separate
function codes for each task like Arm/DisArm [On/Off],
surge-forward/backward [y-direction movement ], sway-
right/left [x-direction movement], heave/dive [ z-
direction], and yaw/roll [rotation]. All the above function
codes are called accordingly by the main code to do the
necessary movements.
7. RESULTS AND ANALYSIS
For better results, we have trained the system with
underwater images of the gate and flare with a batch size
of 2 and a checkpoint of 25k for SSD mobilenet v2 model.
The Threshold was set to a value greater than 75%. The
script was tested on underwater videos in real-time.
The system was able to detect objects and make further
decision of moving in various directions. The time-lag of
processing was around 0.5s on Odroid. We had a dataset of
around 5,800 underwater images. Distributed in 80:20
ratio for train and test respectively. About 4640 images for
training and 1160 images for testing. The dataset
consisted of images labeled gate, flare or both. The
training images were taken using GoPro Hero 3 at 60FPS.
Various factors were taken into consideration like distance
from the object, circular movement, critical sides, partial
images, etc. We have also tested the trained Faster R-CNN

model and the accuracy was greater than 80%, but the
processing time on the embedded devices at real-time was
very high around 2s. To carter with high processing speed
which is the main aim in real-time, we shifted to SSD
Mobilenet models which had comparatively less accuracy,
but faster processing speed which resulted in better
results.
To stabilize the system various parameters in Pixhawk
were manipulated with trial and error. For accomplishing
this benchmarks we manipulated various PID parameters
such as Roll axis rate with value 0.100, Yaw axis rate with
value 0.00 and the Pitch coupling factor with value 1.1 was
found to be precise. The different motor direction was
changed to generate the perfect torque to stabilize the
system.
8. CONCLUSION
We conclude that Odroid Xu4 gave us an optimal solution
based on its computing power, memory usage,
transmission rate, etc over RPi 3B. The SSD Mobilenet v2
model overcame difficulties found using Faster R-CNN v2
model. The tuned hyperparameter values led to better
results and accuracy in detection. Maneuvering of the
system was successfully done using Pymavlink. For more
flexible operations in maneuvering, one can go for the
integration of Odroid Xu4 and Arduino Mega.
REFERENCES
[1] Jonathan Hui, “Object detection: speed and accuracy
comparison (Faster R-CNN, R-FCN, SSD, FPN,
RetinaNet, and YOLOv3)”, 2018.
[2] M. Dąbrowski, J. Gromada, T. Michalik, O. Polska, and
C. Badawczo, “A practical study of neural network-
based image classification model trained with transfer
learning method”,FedCSIS, 2016.
[3] L. Meng, T. Hirayama, S. Oyanagi,”Underwater-Drone
With Panoramic Camera for Automatic Fish
Recognition Based on Deep Learning”,IEEE, 2018.
[4] Raspberrypi (n.d), Raspberry Pi 3 Model B, viewed 4th
June 2019, <
https://www.raspberrypi.org/products/raspberry-pi-
3-model-b/>.
[5] HardKernel (n.d), Odroid-Xu4, Hardkernel CO. LTD.,
South Korea, viewed 4th June 2019,
<https://www.raspberrypi.org/products/raspberry-
pi-3-model-b/>.
[6] Rustom Jehangir (n.d), ArduSub, BlueRobotics,
viewed 6th June 2019, <http://www.ardusub.com/>.
[7] Lorenz Meier (n.d), MAVLink, viewed 7th June 2019,
<https://mavlink.io/en/>.
[8] Tensorflow (n.d), Tensorflow, viewed 5th June 2019,
<https://www.tensorflow.org/>.

IRJET- Autonomous Underwater Vehicle: Electronics and Software Implementation of the Proton AUV

More Related Content

What's hot (18)

Similar to IRJET- Autonomous Underwater Vehicle: Electronics and Software Implementation of the Proton AUV (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Autonomous Underwater Vehicle: Electronics and Software Implementation of the Proton AUV