An Underwater Object Detection and Tracking using Bounding Box Prediction and Convolutional Neural Network

Nowadays, an Object detection and tracking stays energetic and inspiring task in present processer imagination with extensive applications in various industries such as investigation, independent robot directions, and vehicle navigation. The increase in availability of computing power and vast public datasets has made it possible to develop and train robust object identification and tracking systems, driving innovation and advancements in these fields of tracking objects. Object detection faces several challenges such as, where objects are partially or fully hidden by other objects, and variation in pose and scale, which affects an object's appearance. Cluttered backgrounds and changing lighting conditions are also make detection difficult. To overcome these issues, the method of Convolutional Neural Network for Underwater (UW) Object Detection and Tracking (CNN-UWODT) is proposed with DenseNet feature extraction and bounding box for tracking the objects. The process involves with collecting input images from UOT-32 dataset converting to frames extracting features with DenseNet; then predicting the bounding boxes with CNN, and classifying objects with Kernel Extreme Learning Machine (KELM) component to identify key metrics from the input data. The proposed CNN-UWODT presents better results than the traditional You Only Look Once, Version 3 (YOLO-V3)in terms of precision ($94 \%$), recall ($96 \%$) and f1 score ($95 \%$) respectively.

I. Introduction

Nowadays, Investigation of the UW surroundings has increasing on the rising deficit of natural incomes. In addition, several ocean related requests consume gradually increased with dependencies on UW images captured by various forms of techniques. It also includes a varied of grounds with UW landscape scanning [1]. Automated object recognition and computer vision techniques are gaining attention from marine biologists as a rapid and efficient method for estimating new innovations in marine ecosystems, offering a promising alternative to traditional methods, which are inefficient and labor-intensive [2]. UW imaging faces challenges such as poor illumination, color degradation, and haziness, making object detection difficult. Existing methods are limited to two detection types. Most of the objects are detected by the robust. It has a limitation of range and false detections because of marine depth and water levels that will achieve less accurate reliable detection complex [3]. Despite various automated object recognition and computer vision techniques falling into various limitations while addressing the objects in detection, they struggle with detecting small objects with blurry images additionally less effective in this critical field of marine research and conservation [4].