Multi-target Tracking Of SAR Ship Based On KCF And Deep Learning

At present,the regulation of targets on the water has attracted more and more attention,and the research about multi-target tracking is the core problem.Compared with optical imaging,SAR imaging with the advantages of all-day,all-weather,and long operation distance can supply stable observations of ship targets.Therefore,this paper does the research about the multi-target tracking for SAR ship targets.Different from optical images,the difficulties of multi-target tracking with SAR images are mostly about the lack of training samples,the lack of image features,and the effect caused by the defocusing.The multi-target tracking is composed by the detection module and the association module,in which we do some research of the feasibilities of different methods.What’ more,in order to verify the performance of our multi-target tracking algorithm,we establish the SAR dataset of multitarget tracking with the format of MOT16.In the detection module,we respectively try the network method and the kernelized correlation filter method for ship target detection.In order to achieve high-quality detection,SAR image preprocessing is used to overcome the influence of the clutter and the land area,including the median filter,the OTSU operation,and the connected components analysis.The image preprocessing result is used for CFAR coarse detection.After that,we do the research about the method using the network features.The fine-tuned VGG16 network is used for false alarms elimination of CFAR detection.For the lack of training samples,the number of samples after the data augmentation is still not enough to train the VGG16 network,for which we apply the transfer learning on the open SAR dataset.In this process,the weighted cross-entropy loss function is used to improve the training performance with the imbalance training samples.The experiment is used to make comparisons of the crossentropy loss variation tendency in different situations and the precision variation tendency in different situations.Besides,the experiment also verify the effectiveness of the detection method using the network features.For the lack of training samples,we also do the research about the detection module without the network.In this part,the kernelized correlation filter is improved for false alarm elimination.As for the improved kernelized correlation filter,the template tensor is constructed using the small prior samples for generating the template filter.The maximum correlation values of CFAR detection results and the filter are used to distinguish the false alarm and the ship target by comparing with the pre-set threshold.There is a large difference in the distributions of the false alarms and the real targets in the coarse CFAR detection results,which proves that the threshold determination method is reasonable.In the experiments,the accuracy of the detection results is greatly increased after the false alarm elimination operation on the premise of ensuring the number of real targets,which indicates that our improved kernelized correlation filter is effective.In the association module that consists of Kalman filter and SIFT feature matching,by analyzing the characteristics of the ship target movement,different strategies are adaptively adopted to associate the targets based on the three intersection patterns between the prediction box and the detection box,which can reduce the impact of defocusing and target splitting in SAR images.As for the tracking interruption caused by the defocusing,the tracker’s time limit with Gaussian distribution is used to improve the robustness of the multitarget tracking.The experiment results demonstrate that the innovations in our association module have obvious effect on the tracking performance.