Research On Object Tracking Algorithm Based On Siamese Network

Object tracking is one of the most important research directions in the field of computer vision.With the development of artificial intelligence,deep learning methods have made breakthroughs in computer vision tasks such as target detection and semantic segmentation.Object tracking has also met new opportunities and challenges.In recent years,the emergence of multiple large-scale target tracking data sets has cleared the final obstacle to applying deep learning method in tracking tasks,and promoted the vigorous development of novel object tracking algorithms.Among these innovations,the siamese networks framework methods,represented by the fully-convolutional siamese networks(SiamFC)tracking algorithm,are widely accepted by scholars.In this thesis,we propose an object tracking algorithm based on the siamese network with mixed attention mechanism(SiamMA)under the classic SiamFC framework.According to the requirement of using a lightweight network,the backbone network with fewer parameters and higher efficiency is redesigned,and the algorithm performances,e.g.operating speed and tracking accuracy are overall improved.Aim to solve the imbalance between positive and negative samples in the training stage,we propose a stacking and cropping method to construct the self-confrontation training sample pairs,introduce more negative samples.For the purpose of improving the training efficiency of the network,we adjust the loss function according to the characteristics of the training sample pairs.Classical algorithms are easy to lead to tracking failure in complex scenes,which include objects deformation,occlusion,and fast motion.To solve this problem,a novel method by using the mixed attention mechanism(SiamMA)is proposed.According to the characteristics of different branch images of twin network,different attention modules are used to improve the network feature expressive effect,so that the algorithm can obtain stronger robustness in complex scenes.Five public test data sets,such as GOT-10 k and UAV123,are used to test the algorithm.The experimental results show that the performances of success rate and accuracy of our algorithm are better than the classic algorithms,such as SiamFC and KCF.At the same time,the algorithm's strong generalization can better adaptation to different tracking scenarios.Finally,we verify the effectiveness of the algorithm in a real scene from acquiring images by an unmanned aerial vehicle(UAV).The results show that the algorithm still has good tracking accuracy in the case of large angle of view of UAV's camera,and operating speed meets the real-time requirements,which verifies the generalization and practicability of the algorithm.