Research On Algorithms Of Marine Target Detection And Tracking Based On Deep Learning

As a large maritime country,my country has a vast sea area and abundant marine resources,and the development potential of the marine economy is huge.However,in recent years,there have been frequent maritime disputes in adjacent waters,so it has important and far-reaching practical significance to improve my country’s sea surface monitoring technology.Marine target detection and tracking technology plays an important role in the field of sea surface monitoring.This article will use intelligent monitoring and ship environment perception as the background,with marine targets as the research object,and study the deep learning-based marine multi-target detection and tracking method,combined with convolution The advantages of neural network in image feature extraction can improve the accuracy of marine target detection and realize real-time detection and tracking of marine targets.The specific work is as follows:First,a batch of maritime target data sets were collected through public data sets,sea trial pictures,and online searches.The data sets are divided into 10 categories: warships,passenger ships,speedboats,cargo ships and other maritime vessels,as well as buoys,There are about 12,000 non-ship marine targets such as reefs and islands.Use label Img software to label the data set,and make the data set according to the COCO format of the pictures and label data.Secondly,in view of the traditional maritime image target detection algorithm,the detection accuracy is not high,and it is easy to cause the problem of missed detection and false detection of the target.This paper introduces the SSD detection algorithm based on the anchor frame.Use multiple strategies to improve the SSD algorithm,improve the robustness of the standard SSD algorithm,serious imbalance of positive and negative samples,etc.,replace VGG16 with ResNet50 as the feature extraction backbone network,and add FPN(multi-scale feature pyramid)structure to the standard SSD network,And quoting the new Focal Loss loss function.The improved test results show that the detection effect of small targets in maritime images has been significantly improved,and the performance of the improved SSD algorithm is overall better than the standard SSD algorithm.Third,the multi-strategy improved SSD algorithm still faces problems such as the excessive number of anchor frames due to the anchor frame,which leads to the reduction of model inference speed and the accuracy of the algorithm depending on the manually set anchor frame scale.The non-anchor frame CenterNet algorithm with better detection performance is used to continue the target detection research of this subject.Adding mosaic-based enhancement of small targets and the introduction of CIOU loss function optimization and improvement,the experimental results show that the optimized CenterNet algorithm regression box is more accurate than before optimization,and the algorithm performance and small target detection accuracy have been greatly improved.Finally,build a tracker based on the optimized CenterNet detector: use the Kalman filter algorithm to predict frame by frame,use the Hungarian matching algorithm to realize the data association between the tracking information and the detection information,and the motion information and the appearance of the re-identification network extraction Information is evaluated,and the matching accuracy is further improved through the cascade matching strategy.Experimental results show that the tracking algorithm can better correct the size changes of ships in the tracking process,achieve effective tracking of small targets at sea,and to a certain extent alleviate the problem of poor tracking effect due to occlusion.