Research On Detection,recognition And Tracking Method Of Surface Target For Unmanned Surface Vehicles

In recent years,various countries attach more and more importance to marine economy,and the popularization and application of marine equipment play an important role in promoting the development of marine economy.As one of the intelligent marine equipments,more and more researchers have devoted themselves to the research of the technology of Unmanned Surface Vehicle(USV),which is of great significance to the environmental sensing technology of USV.Aiming at the environment sensing technology of USV,this paper focuses on the detection,recognition and tracking algorithm of surface target in complex water environment.For the detection and recognition part,according to the twostage and single-stage detection algorithm in deep learning,two different solutions are proposed.The specific research contents are as follows:(1)Two-stage surface target detection and recognition algorithm based on Faster RCNN.When the fast RCNN algorithm is directly applied to the images on the water surface in the complex water surface environment,the problems of imprecise location and recognition,frequent missing detection of small targets and slow detection speed will appear,failure to provide stable and rapid detection and identification of water surface targets.So this chapter makes five improvements.In the stage of surface feature extraction,this chapter reconstructs the structure of Res Net101 feature extraction network by using multi-layer feature combination,which can further improve the detection and recognition quality.In the generation stage of candidate regions,a priori box is redesigned and the filter strategy of candidate regions is optimized to improve the detection accuracy and time performance.In the detection stage,this chapter uses the detection box voting strategy to correct the location of the target detection box,which is used to obtain better detection accuracy.In the experimental stage,various data extension methods are used to improve the robustness and generalization ability of the model.The experimental results show that the improved algorithm has the best performance in the complex water surface environment,and can meet the requirements of fast and reliable detection of water surface targets in the complex water surface environment.(2)The research of single-stage surface target detection and recognition algorithm based on Yolov3-TPDense.When Yolov3 Algorithm is directly applied to water surface images in complex water surface environment,it can achieve real-time detection effect,but the detection precision is low.Therefore,a YOLOv3-TPDense algorithm is proposed.Firstly,in the stage of feature extraction,the improved TPDense block is used to replace the low-resolution downsampling layer in Dark Net53 network in order to enhance feature propagation,promote effective feature reuse and improve network performance,this stage makes it easier to get rich and productive features.Then multi-scale feature is used to detect different size targets,which involves redesigning multi-scale priori boxes with K-means clustering algorithm in order to improve location accuracy and time performance.In the network training stage,this chapter improves the loss function to obtain a better detection model.In order to improve the robustness and generalization ability of the model,the data expansion technique is used to expand the training samples.Experiments show that the model has better balance ability in detecting quality and speed,and can meet the requirement of real-time and accurate detection of water surface targets.(3)Multi-target tracking algorithm based on KCF-Multi In SORT.If the detection efficiency is not high and the miss detection situation is frequent,and the ID switch is frequent and the target is occluded,the stability and real-time tracking effect of the water target will be seriously affected.Therefore,a tracking framework KCF-Multi In SORT algorithm(simple online and real-time tracking based on multi-information Association)is proposed.First,water video sequence is used to get detection information quickly and stably by Yolov3-TPDense Algorithm,and the detection information is sent to Multi In SORT algorithm and KCF Algorithm.When the detection information is lost,KCF will use the detection information of the previous frame to track the current frame,with this tracking information,the Multi In SORT algorithm is used to match and update the tracker;In contrast,when the detections are all successful,the detection information needs to be recorded in the KCF tracker,and the tracker does not need to be started.The experimental results show that this kind of detection algorithm,which is not completely dependent on the detection algorithm and combines the idea of deep learning tracking method and traditional tracking algorithm,can effectively improve the difficulties and shortcomings mentioned above.In the multi-target tracking scene on the water surface,the tracking effect can be improved obviously,and the tracking function is more stable.