Adaptive Scale And Angle For Sea Surface Object Tracking Based On Deep Learning

Unmanned Surface Platform can replace human beings to complete some dangerous tasks,such as maritime rescue and maritime surveillance.As one of the visual perception tasks for Unmanned Surface Platform,object tracking on the sea plays an important role in locating and monitoring targets continuously.It includes two tasks: object center location and object state estimation(scale and angle),which are coupled with each other.Ships that sail on water are common objects of seas,and their state changes greatly,which makes it very challenging to predict its state.This dissertation takes ships as research objects,and studies steady prediction of object state inspired by points representation of objects in object detection based on deep learning.Finally,this dissertation verifies the performance improvement of the proposed tracker through sequences of ships in real sea scenes.Generally,the main contributions of this dissertation are as follows:Firstly,a tracker with scale adaptation is proposed,which uses points to represent objects for objects on the sea.Aiming at solving the problem of poor accuracy of scale prediction of trackers which use anchors to represent objects when the scale of objects changes rapidly,this dissertation concludes that the root cause lies in the low overlap rate between anchors and object during tracking,which leads to insufficient information to represent objects.Based on the above analysis and inspired by points representation of objects in object detection,this dissertation constructs the points classification network and points regression network for tracking based on fully-convolutional siamese networks.We use Focal loss and Io U loss to realize classification and regression respectively,and formulate training strategy.Compared with the SiamDWRPN(based on anchor)in 12 image sequences with rapid scale variation,the tracker with scale adaptation proposed in this dissertation achieves 17.6% improvement on success rate.Secondly,a real-time end-to-end tracker with angle adaptation for objects on the sea is proposed.Aiming at solving the problem of the relative change of angle to the first frame in the current tracking task,based on the idea of angle adaptation in object detection,this dissertation takes angle prediction of objects as classification and adds the angle prediction network based on the tracker with scale adaptation,and formulates the training strategy,which finally realizes the end-toend prediction of angle in object tracking.Compared with the popular LDES in 4 image sequences with angle variation,the tracker with scale and angle adaptation proposed in this dissertation achieves 11.9% improvement on success rate.Moreover,in order to reduce the parameters and complexity of the model,the angle prediction network is further compressed.As result,the parameter is only about 1 / 88 of that before compression.Thirdly,a ship dataset for tracking is constructed.Since there is no ship dataset for tracking at present,we take image sequences of ships in real sea scenes to build a novel ship dataset.This dataset contains 30 sequences.The challenges of the dataset include rapid scale variation,angle variation,backlight,jittering,occlusion,out of view and so on.The accuracy and robustness of the tracker proposed in this dissertation are verified on this novel ship dataset.Compared with the SiamDWRPN(based on anchor)in 12 image sequences with rapid scale variation,the tracker with scale adaptation proposed in this dissertation achieves 17.6%improvement on success rate.Compared with the popular LDES in 4 image sequences with angle variation,the tracker with scale and angle adaptation proposed in this dissertation achieves 11.9%improvement on success rate and the speed can reach 80FPS(frame per second)tested on a single NVIDIA 1080 Ti GPU,which can meet the requirement of real-time tracking.