基于水下原位摄像机采集的海洋原位视频数据

基于舟山海慧海洋科技有限公司自主研发的水下原位摄像机,采集了海洋原位鱼类视频数据，经网络传输，存储于云服务器上的数据库中。公司具有完善的数据管理制度，可以确保数据的安全性和完整性，同时采用规范化的数据采集流程于数据处理方式，保证原始数据的完整性和处理数据的准确性。预估每日产生的数据量在48段视频数据，每个视频数据对应43200张图片数据，经处理后得到的鱼类位置及类别数据量在10000条以上。该方法可以无接触、准确高效的实时获取水下原位环境中鱼类的种类及位置信息，不仅有助于科学的管理海洋鱼类的饲养状况，还可以为海洋牧场生态建设和海洋资源开发提供重要保障。确保海洋资源的有效利用和可持续开发，为政府及相关机构进一步制定相关政策，及海洋鱼类管理方案提供依据。实时采集的水下原位视频数据，通过公司自主研发的多个基于深度学习的深度网络模型，对于鱼类的类别及位置信息进行统计。算法整体分为3步：第一步，基于用于水下视频增强的深度学习网络模型对水下视频数据进行预处理，减弱水下原位图像的背景噪声并提高水下鱼类图像的信噪比；第二步，基于用于水下鱼类检测分类的深度学习网络模型对预处理后的视频数据进行处理，对视频内鱼类的位置及类别进行检测；第三步，基于用于鱼类跟踪的深度学习网络模型对检测后的视频进行处理，实现对鱼类的跟踪，最终，得到鱼类的种类、位置等相关信息。

This dataset is built upon in-situ underwater fish video data collected by in-situ underwater cameras independently developed by Zhoushan Haihui Marine Technology Co., Ltd. The captured video data is transmitted over networks and stored in a database deployed on cloud servers. The company has a comprehensive data management system to ensure data security and integrity, and adopts standardized data collection workflows and processing methodologies to guarantee the integrity of raw data and the accuracy of processed data. It is estimated that the daily data output consists of 48 video segments, with each segment corresponding to 43,200 image frames. The processed dataset containing fish location and category information includes more than 10,000 entries.

This non-contact, accurate, efficient and real-time approach enables acquisition of fish species and location information within in-situ underwater environments. It not only supports the scientific management of marine fish farming operations, but also provides critical backing for the ecological construction of marine pastures and the development of marine resources. This work facilitates the efficient utilization and sustainable development of marine resources, and provides a reliable foundation for governments and relevant agencies to formulate further policies and marine fish management plans.

The real-time collected in-situ underwater video data is analyzed using multiple deep learning-based network models independently developed by the company to statistically derive fish category and location information. The overall algorithm comprises three stages:
1. Preprocessing Stage: A deep learning network model for underwater video enhancement is applied to preprocess the raw underwater video data, reducing background noise in in-situ underwater images and enhancing the signal-to-noise ratio (SNR) of underwater fish images.
2. Detection and Classification Stage: The preprocessed video data is fed into a deep learning network model tailored for underwater fish detection and classification, to identify the location and category of fish present in the video.
3. Tracking Stage: The detected video frames are processed using a deep learning network model for fish tracking, enabling continuous tracking of individual fish and ultimately yielding comprehensive information including fish species and location.

搜集汇总

背景与挑战

背景概述

该数据集包含海洋原位鱼类视频数据，通过水下原位摄像机采集，数据规模为939条，每秒更新。应用场景包括海洋鱼类饲养管理和海洋资源开发，数据处理采用深度学习模型进行视频增强、鱼类检测分类和跟踪。