Recognition Of Moving Targets In Narrow Waters By Fusing Visible Light Images And LIDAR Data

With the research and application of intelligent technology in the field of water transport,many research institutions at home and abroad invest a lot of manpower and material resources in the development of intelligent ships.As an important part of intelligent ships,environmental perception is an important prerequisite for safe navigation of ships.In narrow waters,there are many uncertainties,static obstacles and dynamic ship targets,which bring difficulties to navigation environment perception.As an advanced high-precision environmental sensing equipment,LIDAR can perceive the three-dimensional information,intensity information and depth information of surrounding targets,and is not affected by the change of light.It performs well in the case of ship encounter and complex background.As an important means of deep learning in image detection,convolutional neural network has a good effect in realizing ship target recognition.In this paper,visible image data and point cloud data generated by LIDAR are used to realize target detection in narrow waters respectively,and fusion analysis is carried out.The main work is as follows:First,the LIDAR and camera was used to carry out the real ship test.Got the original point cloud and image data.Parse the raw point cloud data.In the process of data preprocessing,noise and redundancy,points were eliminated by statistical analysis filtering method.The strength values of point cloud were analyzed and the strength values of different objects were quantified.The cloud intensity value of the target point with a single material was treated as an average value.Secondly,the point cloud characteristics of ship wakes are analyzed.By using the random sampling consistency estimation algorithm based on plane model and taking the intensity value of point cloud as threshold parameter.An improved random sampling consistency estimation algorithm model suitable for this paper is obtained.A variety of target clustering methods were analyzed and compared.Euclidean clustering method was used to achieve geometric segmentation and clustering among objects.In the form of bounding box,the target clustering result was taken as the detection target.Target detection based on point cloud data was realized.Third,a large number of different types of ship image data were obtained by using related methods.Perform manual filtering and annotate filtered data using LabeImg plug-in.Obtain clear and effective network training data.The pre-weight value of the convolutional neural network was used to conduct 30,000 iterations training on the annotated data.An efficient and reliable target recognition and classification model was obtained.Finally,the time registration of LIDAR and camera was realized,and the fusion analysis of sensor detection target was carried out after registration.