Beach Surveillance: A Contribution to Automation  

作　　者：Maria da Conceição Proença Ricardo Nogueira Mendes Maria da Conceição Proença;Ricardo Nogueira Mendes(Marine and Environmental Sciences Centre & ARNET, Aquatic Research Infrastructure Network Associated Laboratory, Lisbon, Portugal;Department of Physics, Faculty of Sciences, University of Lisbon, Lisbon, Portugal;Institute for Nature Conservation and Forests, Algs, Portugal)

机构地区：[1]Marine and Environmental Sciences Centre & ARNET, Aquatic Research Infrastructure Network Associated Laboratory, Lisbon, Portugal [2]Department of Physics, Faculty of Sciences, University of Lisbon, Lisbon, Portugal [3]Institute for Nature Conservation and Forests, Algs, Portugal

出　　处：《Journal of Geoscience and Environment Protection》2024年第12期155-163,共9页地球科学和环境保护期刊(英文)

摘　　要：The problem of human overload in many habitats is becoming increasingly urgent, as it is the driving force that destroys ecosystems beyond repair. This paper describes a possible workflow for beach surveillance, using a deep learning solution available online that runs on a standard laptop with RGB images acquired with a standard camera. The software is YOLO v7, a state-of-the-art real-time object detection model presently used for autonomous driving, surveillance, and robotics. The workflow and parametrization needed for building a model are described, along with examples of the results over 180 test images that ensures an overall precision of 0.98 and recall of 0.94 (F1 = 0.96). The model was parametrized to focus on a minimum number of false positives;from the 5672 possible detections identified by human curation, 5285 were correctly identified and located, 387 missed and there are 116 mistakes. A minimum of computational skills is needed to reproduce this implementation in any user data of the same kind.The problem of human overload in many habitats is becoming increasingly urgent, as it is the driving force that destroys ecosystems beyond repair. This paper describes a possible workflow for beach surveillance, using a deep learning solution available online that runs on a standard laptop with RGB images acquired with a standard camera. The software is YOLO v7, a state-of-the-art real-time object detection model presently used for autonomous driving, surveillance, and robotics. The workflow and parametrization needed for building a model are described, along with examples of the results over 180 test images that ensures an overall precision of 0.98 and recall of 0.94 (F1 = 0.96). The model was parametrized to focus on a minimum number of false positives;from the 5672 possible detections identified by human curation, 5285 were correctly identified and located, 387 missed and there are 116 mistakes. A minimum of computational skills is needed to reproduce this implementation in any user data of the same kind.

关 键 词：People Counting Beach Surcharge Human Detectors Deep Learning Methodologies 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]