Research On Surface Multiple Moving Targets Detection And Tracking Method

China has a vast marine territory,and increasingly relies on marine space and its resources,so it is very important to safeguard national marine rights and interests and build a powerful marine country.As one of the marine equipment platforms,unmanned surface vehicle(USV)is gradually applied in the fields of surface unmanned patrol,seabed resource survey and polar scientific investigation due to its small volume,good maneuverability and unmanned casualties.In this paper,surface multiple moving targets detection and tracking method is studied.By detecting and tracking surface multi-moving targets,the USV can avoid them and complete specific tracking tasks.The specific research contents are as follows:First of all,the basic theory of multiple moving targets detection and tracking is studied,and the three aspects of moving target detection algorithm,tracking algorithm and data association method used in the tracking process are analyzed.The advantages and disadvantages of the three are summarized,which provides the theoretical basis for the further study of the paper.Secondly,this paper introduces convolutional neural network as the basic detection algorithm,aiming at the problem that the calculation of YOLO-V3 detection algorithm is large and the performance requirements of computing equipment are high,which leads to its slow running speed on embedded equipment.By introducing the expansion-projection(EP) module and the projection-expansion-projection(PEP)module,the original residual network in YOLO-V3 is replaced,we get light weight detection network(YOLO-V3-LT),while maintaining the detection accuracy of the network,improve the detection speed of the network so that it can be applied to embedded devices to carry on USV.Through training the light weight detection network on the surface ship target data set,the test results on the embedded device show that the light weight detection network keeps the accuracy of detecting the surface target,and the detection rate is effectively improved.Aiming at the problem of occlusion between ship targets with similar appearance,this paper introduces the Mask R-CNN algorithm to detect and segment multiple ship targets on the water surface,and obtains the contour mask eigenvector of each target,so as to enhance the stability of data association in the occlusion scene.Based on the YOLO-V3-LT algorithm and Mask R-CNN algorithm,this paper proposes a YOLO-V3-LT based water surface multiple targets tracking algorithm and a Mask R-CNN based water surface multiple targets tracking algorithm.The effectiveness of the proposed tracking algorithm is verified by the public data set.Then,this paper improves the traditional data association algorithm and proposes a data association algorithm based on information fusion:the motion model is used to predict the motion information of the target during occlusion,and the data association algorithm based on the motion information is constructed;the apparent feature information of the targets extracted by convolutional neural network and the mask information of the targets output by Mask R-CNN are used to construct the data association algorithm based on the feature information;the data association algorithm based on motion information is fused with the data association algorithm based on feature information by weight value,and the data association algorithm based on information fusion is obtained.The validity of the data association algorithm based on information fusion is verified by comparing with the traditional method.Finally,the software and hardware system for tracking surface multiple moving targets is built on the platform of USV-1,and the tracking system is verified by experiments.The experimental results show that the average accuracy of the water surface multiple targets tracking method based on YOLO-V3-LT is 88.74%in the test set,and the frame rate is maintained at about 40 fps in the real-time detection test.In the four scene tests on the water surface,the average accuracy of the tracking is 92.39%,and it can effectively reduce the influence of occlusion on the tracking;the average accuracy of the multi-target tracking method based on mask is 94.52%in the test set.In the four scene experiments on the water surface,the average accuracy of tracking is 98.16%,it can track the target with high accuracy.And in the scene of occlusion between many targets,using the mask information obtained by segmentation for data association can effectively reduce the occurrence of ID switch and tracking loss between targets.