Research On Target Tracking Technology Based On Siamese Neural Network

Target tracking technology based on siamese neural networks is currently one of the hot research directions in the field of deep learning.The siamese neural network is a siamese network architecture with a deep structure.Applying it to target tracking technology can improve the accuracy and real-time performance of target tracking.This article will study the target tracking technology based on siamese neural network from three aspects: loss function,feature extraction network and attention mechanism network.Firstly,the siamese region proposal network employs the two-class cross-entropy loss function to classify the background before and after.The loss function converges slowly in the training process,so that the trained model is not accurate enough to track the target under the condition of similar target interference.In this paper,the loss function in the siamese region proposal network is redesigned,and the variant focus loss function combining the focus loss function and the balanced cross entropy loss function is used as the classification loss function in this paper.The loss function solved the problem of sample imbalance in network training,improved the classification accuracy,and optimized the training model.Experimental results show that the redesigned loss function converges faster,its convergence value is closer to the true value,and the target tracking in the test scene is more accurate.Secondly,the siamese region proposal network adopts a relatively shallow convolutional neural network as the feature extraction network,and the extracted convolutional feature semantic information is not rich enough,and the discriminative power and positioning accuracy of the trained network model are not high enough.In this paper,an improved deep residual network is applied as the feature extraction network of the siamese region proposal network,so that the siamese tracker has the ability to obtain deeper convolutional features from the network,and improved the discriminative ability of the tracker.And because the siamese network has translation invariance,this paper adopts the spatial sensing sampling strategy to sample to alleviate the impact of the siamese network due to the translation invariance.The experimental results show that the improved deep residual network reduced the amount of calculation and ensured the effectiveness of feature extraction,which proved the rationality of the algorithm in this paper.Finally,this paper proposed a target tracking algorithm based on the siamese attention mechanism network to solve the problems of small targets,short-term occlusion,motion blur,and scale changes in UAV targets.In this paper,two channel attention mechanism networks are designed in the siamese neural network for feature selection,and the weights are re-assigned to different channels,which can make more effective use of features.Then,two region proposal networks with the same spatial resolution are designed to perform hierarchical feature fusion,so as to realize multi-layer feature collaborative estimation of the target location.Experiments show that the network structure of the siamese attention mechanism designed in this paper has excellent performance.Finally,the algorithm of this paper is applied to the UAV designated target tracking,which has good tracking effect and real-time performance.