Research And Implementation Of Object Tracking Technology Based On Siamese Network

As one of the basic research contents in the field of computer vision,object tracking is widely used in the fields of intelligent video surveillance,autonomous driving,intelligent agriculture and intelligent transportation.In recent years,large-scale publicly labeled image datasets are emerging,and the computing power of computer hardware is also improving.These factors have promoted the development of object tracking technology.Although the current development trend of object tracking technology is positive,it is still a difficult problem to realize high-performance tracker.For example,there are too many types of application scenarios,and the laws of object movement often change.In addition,other factors in the scene always interfere with the execution of the object tracking task.Therefore,this paper studies the problem of object tracking in complex environment.Based on the Siamese network,we proposed an object tracking algorithm with attention mechanism and an anchor-free object tracking algorithm.Then,we designed a prototype system based on the above algorithm and implemented it.Finally,we selected authoritative datasets to validate our algorithm.The main contributions of this paper include the following two aspects:(1)In complex scenes,interference factors will affect the performance of the tracking algorithm.Because the interference factors easily lead to the decline of the discrimination ability of the algorithm.To solve this problem,we proposed an object tracking algorithm based on the Siamese network and fused with the attention mechanism.First,we release the translation invariance of the Siamese network,which enables the algorithm to extract features using deep neural networks.The features thus obtained have rich levels and different characteristics.The tracking algorithm can get more Information from it.Then,we introduce the attention mechanism and design the calculation method of the response map in two stages.We construct a cross-correlation module DCAM(Depth-wise Cross-correlation with Attention Mechanism,DCAM)in the first stage,and design an adaptive response map fusion strategy in the second stage.Such a design enables the response map to have a global perspective during the calculation process.By filtering the acquired information,the algorithm pays more attention to the relevant information of the target,so as to achieve a stable and accurate tracking effect.(2)When designing a tracking algorithm,we hope to reduce the complexity of the regression task.It is also critical to train a tracking model with better performance.Therefore,we proposed an anchor-free object tracking algorithm based on the Siamese network.First,on the basis of the previous algorithm,the algorithm calculates the object bounding box without anchor.This approach can not only simplify the design of regression tasks,but also ensure the accuracy of tracking results.Second,we adopt the dynamically scalable cross entropy loss as the loss function for the classification branch.Such a design alleviates the phenomenon of the class imbalance in the data set.It not only corrects the optimization direction of the model,but also improves the overall performance of the tracking algorithm.Finally,we design a corresponding learning rate adjustment strategy to stochastically average the weight of a number of models.The generalization ability of the tracking model has been improved.Based on the above research results,we designed a prototype system and implemented it.We set up the experimental environment and designed the comparative experiments in detail.Through experiments,we verified the effectiveness of the algorithm proposed in this paper.We selected the authoritative datasets VOT2018 and VOT2019 to test and evaluate the proposed algorithm.The experimental results show that compared with the existing algorithms,the proposed algorithm has obvious improvement in accuracy and robustness.Meanwhile,it also performs well in overall performance.