Research On Automatic Annotation And Multi-target Tracking Algorithm For Ship Video Target Detection

Video target detection technology can accurately locate and classify each target of interests in video or image sequences.Different from image target detection,video target detection faces more challenges,such as camera shake,motion blur and other factors,which may lead to false detection or missed detection.With the development of deep learning technology,the performance of target detection has been significantly improved,but the first task is to build a large number of annotated training data sets.In order to solve the problems of low efficiency of manual annotation and false detection or missed detection in video target detection,the research on automatic annotation and multi-target tracking algorithm for ship video target detection has been done in this paper.The main research contents are listed as follows:(1)In order to improve the efficiency of video image annotation,a semi-automatic video target annotation framework by combining detection and tracking is proposed.Firstly,manually annotated samples are used to train the improved YOLOv3(You Only Look Once version3)detection model offline,which is used as an online annotation detector.Then during online annotation,the target position and label are determined manually in the first frame.In the subsequent frame,target position is determined automatically according to the IOU(Intersection-Over-Union)of the detection box and the tracking box,and the response of the improved KCF(Kernel Correlation Filters)tracker is used to judge the target disappearance so that the current target annotation is stopped automatically.Finally,a key-frame extraction algorithm based on the target saliency is used to select key-frames.The above framework and method is integrated into the ship video target semi-automatic annotation software.The performance of improved YOLOv3 detection model and improved KCF tracking algorithm are firstly verified through experiments,and the feasibility of the semi-automatic annotation method is then verified through the ship video annotation task.(2)In order to improve the performance of ship video target detection,and reduce the rate of false detection and missed detection,a ship video target detection method based on improved DeepSORT(Deep association Simple Online and Real-time Tracking)multi-target tracking algorithm is proposed.Firstly,the DeepSORT multi-target tracking algorithm is improved by using the mean value of cosine distance to construct cost matrix.Secondly,combined with the improved YOLOv3 detection algorithm,the rate of false detection is reduced by target trajectory optimization.Using MOT2016 data set to test the improved DeepSORT,the number of identity switches(IDs)is reduced by 7.5%.Compared with the improved YOLOv3 model,the proposed target detection algorithm based on multi-target tracking improves the m AP and accuracy by 0.3% and 0.8% respectively,and reduces the rate of false detection by 11.2% when using the self-built ship video target detection data set.Finally,this paper is summarized and prospected.