A Video Target Detection Technique For A Common Border Smuggling Boat Type

Smugglers in border waters often use a high-powered vessel that is illegally modified,which brings great difficulties and dangers to customs anti-smuggling work.In order to detect the smuggling boat type in time,the public security organs set up a large number of monitoring probes on the shore,and the staff watched the video surveillance and issued an alarm.However,the number of surveillance sites is large,and the video data is huge.When watching video surveillance,the staff not only has a large workload,but also tends to be distracted,resulting in missed inspections.In order to meet the actual needs of common smuggling boat type detection in border waters,reduce the workload of border guards to identify the type of smuggling boat.Based on deep learning algorithms,this paper studies a variety of convolutional neural network models and Transformer models based on attention mechanism,and applies them to common smuggling boat-type video target detection tasks in border waters.The main research contents and methods of this paper are as follows:Firstly,to address the problem that there is no dataset dedicated to the detection task of border smuggling boat targets,this paper collects image data containing common smuggling boat targets from surveillance videos and the internet according to the detection scenario of border waters,uses Labelimg to label smuggling boat targets,and adopts two online data enhancement strategies,Mosaic and Mix Up,to build a dataset dedicated to the detection task of common border smuggling boat targets.Secondly,in order to solve the problem that the latest target detection algorithms have not been applied in the field of ship images,this thesis makes a comparative analysis of three target detection algorithms: Faster R-CNN,YOLO V4 and YOLO X.The experimental results show that YOLO X achieves the highest detection accuracy even though there are some misses in the public maritime dataset and the common smuggling boat type dataset,and it is the best basic model for the common smuggling boat type target detection tasks.Thirdly,aiming at the problem that YOLO X algorithm fails to detect small-scale common smuggling boat targets,this thesis uses the method of replacing feature extraction network to improve.The attention mechanism-based Swin Transformer model is embedded in YOLO X as feature extraction network.The attention mechanism and hierarchical pyramid structure of the Swin Transformer model are used to improve the feature extraction ability of the network.The experimental results show that the improved model reduces the leakage of small-scale smuggling boats and has higher detection accuracy.Finally,for YOLO X＿Swin model has a large number of parameters,which requires high computing performance of shore-based monitoring equipment.In this thesis,the model parameters are reduced by using hyper-parameter adjustment in the feature extraction network part.In the part of feature fusion network,the model parameters are reduced by using deep separable convolution instead of ordinary convolution,Resnet structure is increased in deep separable convolution,the number of network channels is increased,and image features of common smuggling boat types are fully fused.The experimental results show that the lightweight improved model reduces the model parameters and improves the detection speed while the detection accuracy is basically unchanged.