Channel Multi-video Target Search Method Research And Application

Water transportation is the foundation of China’s comprehensive transportation and an important component of China’s economic,foreign trade,and social development.The inland waterway is the foundation of China’s waterway transportation,and efficient and safe management of the waterway is of great practical significance.At present,the management method of inland waterways is still in a relatively early stage.For some illegal sand mining and illegal docking ships,target detection still remains in the stage of using manual video surveillance and manual patrol by patrol ships,resulting in low management efficiency.At the same time,due to the erosion of wind and sand in the inland waterway and the covering of goods,it has long been difficult to distinguish the ship numbers used to distinguish ships.Some dangerous goods ships and illegal vessels that require key monitoring cannot achieve identity identification across multiple videos,and cannot locate the sailing time and route of a certain ship on the waterway,posing significant safety hazards.Although there are many methods in the field of ship detection and identification,there are still the following shortcomings:(1)Although there are object detection algorithms for deep-sea ships and video surveillance systems involved in rivers and lakes,there have been no cases where the two are integrated and applied to inland waterways.Moreover,due to the environment of inland waterways,ordinary object detection methods cannot meet expectations well and are not conducive to real-time detection of illegal ships.(2)Although ship identification has been explored in both inland and offshore environments,it mostly relies on methods such as calculating ship trajectories and extracting ship identification numbers,which perform poorly in complex inland river scenarios.This study adopts deep learning technology to apply the methods of object detection and small sample learning to the management of inland waterways.The main focus is on ship target detection in fuzzy and multi-scale scenarios,as well as identification of different ships and even ship search in the entire channel segment.A ship target detection algorithm based on improved YOLOv5 s has been developed to address the above issues.The specific improvement is the introduction of the Sim AM attention module,which enhances the algorithm’s feature selection ability and makes the features more global;Using Bi FPN instead of the original PAFPN allows the network to maintain more of the original image features and reduce the possibility of missed detection in blurred scenes;Finally,Focal Loss loss function is used to focus the model on samples that are difficult to classify and improve the detection accuracy.In the end,the average accuracy of ship target detection in different scenarios reached 86.08%,achieving the expected effect.We have completed a multi video object search algorithm based on twin neural networks,which enables the identification of ship identities from different perspectives.Based on this,some improvements are also made: firstly,VGG16 is selected as the feature extraction network,and the triple neural network is also referenced as the performance evaluation standard;Secondly,the lightweight module Ghost Module is introduced into the backbone network to optimize the model parameters and the pyramid based feature extraction module PAFEM is introduced to accurately locate ship targets of different scales;Then,a spatial variation network(STN)is introduced to enhance the spatial invariance of the network;Finally,the GAM attention mechanism is introduced to improve the feature extraction ability of the network.In the end,the accuracy of ship identity identification reached 91.4%,and the identification effect was relatively obvious.The innovation of this paper lies in:(1)A ship target detection algorithm based on improved YOLOv5 has been developed for real scenes in inland waterways.(2)A multi perspective ship identification algorithm based on twin neural networks was developed to address the difficulty of identifying ships in multiple videos within inland waterways,and a ship search algorithm in multiple videos was implemented.The ship data involved in the research work of this article is limited to the real scene of the inland waterway,and only ship images from three different perspectives are collected in a fixed scene.Further consideration will be given to the ship data on the entire waterway in the future;How to improve the twin neural network and further improve the detection accuracy is the focus of later research.