Water Surface Target Recognition And Tracking Method Based On Data Fusion

Unmanned surface vehicle(USV)is playing an increasingly important role in protecting marine territories and utilizing marine resources.By establishing a complete surface environment situation,the perception system of USVs provides a strong guarantee for reasonable mission planning and successful completion of the mission.However,the target recognition algorithm which only relies on optical information has poor accuracy in harsh ocean environment such as deck wetness and backlighting.The classical target tracking algorithm has poor robustness when the movement state of unmanned surface vehicle is complex and changeable.The RGB image data and lidar point cloud data have good complementarity,and the data fusion algorithm can improve the effectiveness of target recognition algorithm and target tracking algorithm in harsh water surface environment.Therefore,considering the characteristic of water surface environment,this paper has studied the method of water surface target recognition and tracking algorithm based on multi-sensor data fusion.The main research contents of this paper are as follows:First,in order to solve the problem of noise and tail wave interference in the original point cloud data and the problem of effective matching between RGB image data and lidar point cloud data;a Euclidean clustering method using point cloud intensity weighting is proposed to filter and cluster the point cloud original data.It can effectively solve the problem of tail wave interference during the clustering process of surface targets.A point cloud data preprocessing method based on binary rasterized images is proposed to reduce computing power consumption.The camera imaging model is used to jointly calibrate the optical camera and lidar,and the pixel-level fusion of RGB image data and lidar point cloud data is realized.Second,in order to improve the recognition effect of target recognition algorithm in the harsh marine environment with frequent rising waves and backlight phenomena.The projected point cloud data is processed by a combination of Gaussian filtering and dilation erosion to obtain a data matrix.A multi-source data fusion neural network for synchronous feature extraction is constructed,and the stable depth and intensity features of 3D point cloud in various harsh situations are used to improve the accuracy of the target recognition algorithm.Third,the irregular motion of the pixel position of the water surface target in the RGB image data leads to the failure of the target tracking algorithm.The Deep-SORT algorithm is improved by using the latitude and longitude calculated by the Haversine method and the target size extracted from the target point cloud cluster.Using normalized variance matching and average color difference to measure texture similarity and color similarity,a calculation method of target matching cost matrix is proposed to improve the robustness of water surface target tracking method.Based on the above research content,a complete unmanned surface vehicle perception system is designed,and the field test is carried out using the 3T unmanned surface vehicle test platform.It verifies the effectiveness and robustness of the water surface target recognition and tracking method based on data fusion proposed in this paper.