Research On Vision-Based Marine Targets Detection And Tracking Technology On A USV

The vast sea area not only contains abundant resources,but also is one of the main channels of trade and cultural exchange.Vigorously developing marine equipment is a necessary means to enhance China’s marine development and control capabilities.Unmanned surface vehicle is a kind of unmanned vehicle for developing surface space.It has the advantages of small size,high speed,intelligence,better stealth,no casualties and so on.It can complete various scientific investigation and engineering tasks,and has a wide range of applications and important research value.The Qixi USV is the main research object in this paper.Taking the environmental information needed by USV to avoid obstacles during autonomous navigation as the starting point,this paper focuses on the visual detection and tracking technology of surface targets.It is devoted to detecting and tracking surface targets such as ships,islands and reefs from complex water environment.In order to obtain the state of the moving target,it is estimated continuously to prevent collision with the USV.The specific contents are as follows:Firstly,the advantages and disadvantages of common image segmentation algorithms in detecting surface targets are analyzed.According to the characteristics of water surface images disturbed by sunlight and reflection after threshold segmentation,the threshold segmentation algorithm is optimized.The algorithm is suitable for surface targets with a large grayscale difference from water surface environment.Secondly,considering the characteristics of deep learning target detection algorithm based on candidate regions and regression,YOLOv2 is used as the basic detection algorithm in this paper.In view of the shortcomings of this algorithm in detecting small targets on the water surface,two feature hierarchical cascading strategies are introduced:Top-Down Modulation and Feature Pyramid Networks.The improved network can combine low-level features with low semantics,high resolution,and high-level features with high semantics and low resolution to obtain more detailed information.The original network and two networks with hierarchical cascade strategy are trained by the self-made surface target data set.The performance of the test set shows that both strategies can effectively improve the ability of model detection for small targets on water surface,and YOLOv2 algorithm based on Feature Pyramid Networks has higher precision and speed.Then,by analyzing the applicability of various target tracking algorithms on the USV,the simple and efficient Mean-Shift target tracking algorithm is introduced in this paper.Aiming at the problem that the shape of the target varies in different directions due to the changing course angle of USV during navigation and the movement of surface target,this paper presents an anisotropic bandwidth adaptive Mean-Shift surface target tracking algorithm.The qualitative and quantitative analysis of the experimental results show that the improved algorithm effectively improves the accuracy,robustness and real-time performance of the tracking.Finally,the object of study of this paper,Qixi USV,and the hardware and software structure of the optical vision system developed in this paper are briefly introduced.The system provides effective technical support for path planning in Qixi’s autonomous obstacle avoidance experiment.In the experiment,the optical vision system accurately detects and tracks obstacles and calculates their relative positions.The position information of obstacles is transmitted to the planning and control system through network communication.Finally,Qixi successfully avoids obstacles.This proves the accuracy and real-time of the improved detection and tracking algorithm and the reliability of the optical vision system.