An Improved YOLO v4 Algorithm-based Object Detection Method for Maritime Vessels

Abstract

Ship object detection is the core part of the maritime intelligent ship safety assistance technology, which plays a crucial role in ship safety. The object detection algorithm based on the convolutional neural network has greatly improved the accuracy and speed of object detection, which YOLO algorithm stands out among the object detection algorithms with more excellent robustness, detection accuracy, and real-time performance. Based on the YOLO v4 algorithm, this study uses the k-means algorithm to improve clustering at the input side of image data and introduces relevant berth data in the self-organized dataset to achieve detection of ships and berths for the lack of detection of berths in the existing ship detection algorithm. The experimental results show that the mAP and F1-score of the improved YOLO v4 are increased by 2.79% and 0.80%, respectively. The improved YOLO v4 algorithm effectively improves the accuracy of ship object detection, and the in-port berth also achieves better detection results and improves the ship environment perception, which is important in assisting berthing and unberthing.

A streamlined approach for intelligent ship object detection using EL-YOLO algorithm

“…Due to its advantage in real-time processing, this series has found wider applications in practical object detection tasks, particularly in scenarios that demand real-time requirements, such as real-time object search for unmanned aerial vehicles 40 and object detection for unmanned vessels. For instance, He et al 41 employed an improved version of YOLOv4 to detect ships approaching harbor berths. Their approach enhanced the input data clustering using the k-means algorithm, thereby improving the detection accuracy.…”

Construction Method of Unmanned Surface Vehicles Power Control Platform Based on Environmental Force Feedback

“…Through specific sensing equipment and mechanical transmission devices, combined with computer technology, the force generated by the movement of unmanned ships is obtained, and the force is fed back to the operator through a specific actuator, so that the operator can feel the existence of environmental forces. It increases the operator's immersion in the process of controlling the ship, and improves the operation efficiency and accuracy [7]. The basic composition of the force feedback system and the control structure of the unmanned vessel is shown in Figure 1.…”

“…Zou et al studied finitetime output feedback attitude control of rigid spacecraft, which ensures that rigid spacecraft can track the attitude of time-varying reference attitude in finite time, thus improving the application efficiency of spacecraft [6] . In the field of motion assistant control of unmanned ships, He G et al used YOLO algorithm in ship object detection, which played a crucial role in ship navigation safety [7] . Lv C et al proposed a hybrid coordinated control strategy based on signal energy method for the speed and heading control problem of underactuated USV, and the results show that the algorithm has certain effectiveness and stability [8] .…”

A review of intelligent ship marine object detection based on RGB camera

“…Meanwhile, several improved YOLO series methods have been proposed to improve maritime detection accuracy and speed. Considering the widely distributed input objects that bring insufficient receptive field for detection, especially for small and unbalanced ones, the K-means clustering algorithm is applied to input part of YOLO to cluster initial data ( [90,92,93]). For the backbone part, Li et al [88] integrated Densenet into the Darknet-53 backbone network of YOLOv3 to replace 26 × 26 and 13 × 13 of the lower sampling layers, which made the feature transmission in the deep network structure direct to reduce loss, showing benefits for ship object detection and environmental awareness ability.…”

“…Meanwhile, several improved YOLO series methods have been proposed to improve maritime detection accuracy and speed. Considering the widely distributed input objects that bring insufficient receptive field for detection, especially for small and unbalanced ones, the K‐means clustering algorithm is applied to input part of YOLO to cluster initial data ([90, 92, 93]). For the backbone part, Li et al.…”