Image Dataset for Neural Network Performance Estimation with Application to Maritime Ports

Abstract

Automated surveillance systems based on machine learning and computer vision constantly evolve to improve shipping and assist port authorities. The data obtained can be used for port and port property surveillance, traffic density analysis, maritime safety, pollution assessment, etc. However, due to the lack of datasets for video surveillance and ship classification in real maritime zones, there is a need for a reference dataset to compare the obtained results. This paper presents a new dataset for estimating detection and classification performance which provides versatile ship annotations and classifications for passenger ports with a large number of small- to medium-sized ships that were not monitored by the automatic identification system (AIS) and/or the vessel traffic system (VTS). The dataset is considered general for the Mediterranean region since many ports have a similar maritime traffic configuration as the Port of Split, Croatia. The dataset consists of 19,337 high-resolution images with 27,849 manually labeled ship instances classified into 12 categories. The vast majority of the images contain the port and starboard sides of the ships. In addition, the images were acquired in a real maritime zone at different times of the year, day, weather conditions, and sea state conditions.

Every Vessel Counts: Neural Network Based Maritime Traffic Counting System

“…For model training, publicly available pre-trained weights were obtained. Then, the model was trained on the Split Port Ship Classification Dataset (SPSCD) [ 2 ]. The dataset consisted of 19,337 images in 1920 × 1080 resolution with 27,849 labeled ships, categorized into 12 classes.…”

“…For reference, examples of each class can be found in Appendix A . The dataset contained images taken under different weather conditions (e.g., rainy, cloudy, sunny, foggy, …), different illumination effects from the sun and sea surface reflections (different times of the year and day), and different sea state conditions [ 2 ]. In addition, the vast majority of the images contained the port and starboard sides of the ships.…”

“…Not all vessels are required to have AIS on board, resulting in incomplete data [ 1 ], and it can be difficult for VTS to accurately identify and track vessels in crowded or complex port environments. This problem is particularly relevant in Mediterranean passenger ports such as the Port of Split, where maritime traffic is diverse and consists of a large number of vessels that are not covered by traditional vessel monitoring systems [ 2 ]. To overcome these limitations, advances in computer vision combined with video surveillance have emerged as a promising solution.…”

“…This maritime zone presents unique challenges for accurate traffic monitoring and vessel counting due to the specific configuration of maritime traffic. As described in [ 2 ], due to the port structure, traffic ranges from smaller boats of all types, such as sailboats and speedboats to passenger ferries and large cruise ships. Traditional traffic monitoring systems have proven inadequate for monitoring and counting this variety of vessels, resulting in an incomplete or inaccurate traffic analysis [ 1 ].…”

“…The main scientific contributions can be summarized as follows: Enhanced counting and classification of vessels into 12 categories as they traverse the maritime zone, utilizing an adjustable counting zone that differentiates between vessels entering and leaving the port; Introduction of stability assessment as a complement to tracking algorithm to eliminate of false positives on unwanted objects, such as seagulls flying in front of camera, reflections from sea surface, etc. ; Training and validation of YOLOv4 on the newly compiled SPSCD [ 2 ] dataset. …”