Research On Marine Small Target Tracking Based On Uav Video Analysis In Complex Environment

With the increase in international economic cooperation and maritime activities such as transportation and fishing,the number of maritime accidents has also increased.As a result,rapid and accurate accident location and emergency rescue are particularly important.Unmanned aerial vehicle(UAV)rescue systems have the potential to remotely and automatically locate,track,and perform search and rescue operations for maritime accidents.However,due to the wide perspective of drone cameras,small target sizes,and the complex and dynamic maritime environment,imaging quality is often poor,hindering subsequent target tracking.To address this issue,this paper proposes a novel approach based on Siamese convolutional neural networks,as well as the incorporation of second-order pooling networks and attention mechanisms,to enhance target tracking performance and enable real-time tracking.The specific research contents are outlined below:(1)To simulate the challenging scenario of sea fog,this paper employs two fog generation algorithms to process both the training and testing datasets.However,collecting a large amount of dataset in foggy conditions is notoriously difficult and updating data labels is costly.To circumvent this challenge,this study selects small targets from the outdoor video datasets,GOT-10 K and Tracking Net,as the experimental data.To prevent overfitting,the selected dataset is augmented using the two fog algorithms to create low,medium,and high concentration fog scenarios.(2)To mitigate the impact of complex and dynamic maritime environments on target tracking,this paper enhances the fog algorithm by incorporating the dark channel prior.Specifically,the proposed method employs guide filtering to enhance the speed of the fog removal process and gamma correction image algorithms to improve color reproduction.By comparing the effectiveness and processing time of three different fog algorithms,this study selects the algorithm that best aligns with the requirements of target tracking,and integrates it into the tracker data pre-processing to significantly enhance tracking performance.(3)To address the issue of suboptimal tracking performance for small targets,this study proposes a novel approach based on the Siam RPN algorithm.Specifically,the second-order pooling network is added to the end of the twin network feature extraction structure to refine and enhance the characteristics of small targets,and the extracted feature information is integrated to improve the tracker’s ability to identify small targets.Experimental results demonstrate the efficacy of the proposed algorithm.(4)To address the issue of tracking drift caused by drone interference and changes in viewing angle,this article proposes an anchor-free frame twin network(SANA)based on an attention mechanism.This network incorporates the band pool module and the context channel module at the end of the 3rd,4th,and 5th layers of the network from the benchmark tracker Siam CAR.In the template feature extraction branch,strip pooling is performed in two directions to weigh the space dimension of the target location and capture remote relationships of communities to prevent interference of the target location prediction from unrelated backgrounds,thus improving the ability to judge the target location.In the search area feature extraction branch,the context channel module integrates overall and local characteristics information to obtain remote context information,adjust the proportion of channels related to target characteristics to enhance channel strength,and accurately return the border frame of the target.The proposed SANA network effectively addresses the problem of tracking drift caused by drone interference and changes in viewing angle,as demonstrated by experimental results.