Object Segmentation Method For Complex Airport Environment

In the past two decades,with the rapid development of the civil aviation industry,traditional civil aviation monitoring methods have become increasingly unable to meet the requirements of modern airport scene surveillance.As a relatively mature technology,computer vision technology has begun to be applied to the intelligent construction of airport scenes.In the process of intelligent airport construction,the object segmentation technology as a computer vision technology is particularly important.However,due to the complex and changeable environment of the airport scene,the traditional object segmentation algorithm has appeared to be inadequate,and the effect is not good in practical applications.In recent years,due to its powerful feature expression capabilities for image data,convolutional neural networks have been widely used in related visual tasks and have achieved certain results.Therefore,based on deep learning related theories,it is very practical to establish a feasible and robust moving object segmentation method for the complex environment of the airport.The work of thesis has the following contents.Due to the vastness of the airport scene and the complex and changeable environment,objects of different scales will appear in the same surveillance screen,and the convolutional neural network will gradually lose detailed information in the feature extraction process.Therefore,in the field of image segmentation,multi-scale structures are used to obtain features of different scales in order to restore fine detection results.At present,in the task of object segmentation,features are not selected when fusing features of different scales.Therefore,thesis proposes to add the attention mechanism to the multi-scale structure to allow the model to learn important multi-scale features.In addition,since airport surveillance is a long video sequence,in some complex scenes,methods based on deep learning will segment the same type of stationary aircraft into moving targets.In order for the model to perceive motion characteristics,it is necessary to maintain a stable background model.thesis proposes a novel background estimation generation module to estimate a stable background.And use the contrast feature assimilation module to describe the semantically accurate foreground shape representation.Through a large number of experiments on the airport data set,compared with other classic methods,the robustness of this method is verified.In actual airport applications,it is necessary to stably maintain the fine segmentation of the target.In the commonly used feature extraction network,the internal correlation between features is not considered.For this reason,thesis proposes a learnable edge enhancement module.This module learns the inherent correlation between multi-frame encoders with multi-scale features,and performs feature decoding through skip connections.Aiming at the problem of unbalanced foreground and background samples in the airport environment,Focal Loss is used to guide the training of the network,while increasing attention to difficult samples in the data.In addition,thesis uses statistical methods to construct an airport background to guide the model to encode target features,and uses propagation-based ideas to guide moving target segmentation.The experimental results on the airport data set show the rationality of the model design in thesis.