Underwater Object Detection through HOG-SVM and Deep Learning

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Underwater Object Detection through HOG-SVM and Deep Learning"

by

UY Mikaela Angelina Chan

Abstract:

In order to contribute to robotics research, specifically of Automated 
Underwater Vehicles, this study aims to develop a real-time system that 
can detect underwater divers over a video feed. To achieve this, a diver 
detection model had to be trained, and two approaches were studied: 1) 
HOG-SVM and 2) YOLO single-shot detector. The first approach uses a 
sliding-window approach that uses Histogram of Oriented Gradients (HOG) to 
extract features from each "window" of an input image and classifies it as 
a "diver" or "non-diver" through a Support Vector Machine (SVM). Data 
augmentation techniques were also used to increase the collected training 
data that were used to train the model. The results, however, showed a 
high rate of false positives. The second approach is the state-of-the-art 
YOLO single-shot detector, which is a 26 layer deep neural network 
architecture that takes an input image in one pass and outputs bounding 
box information of the objects detected. Transfer learning was used to 
fine-tune the pre-trained YOLO model to detect underwater divers, which 
was done by re-training the weights of last 2 layers using the collected 
training images. This approach can successfully detect underwater divers, 
and it was implemented together with a tracking algorithm that resulted in 
a video detecting system with an average frame rate of 45 fps, which 
satisfies the real-time system requirement.

Date                 : 4 May 2017 (Thursday)

Time                 : 15:00 - 15:40

Venue                : 2127C (via lift 19)

Advisor              : Prof. Chi-Keung TANG

2nd Reader           : Prof. Albert CHUNG