Research And Implementation Of Adaptive Target Tracking Algorithm Based On Correlation Method

Target tracking technology is an important branch of computer vision,which is widely used in intelligent monitoring,automatic driving,military defense and other fields,which has important research significance and practical value.Among them,the traditional and deep tracking algorithms based on correlation filtering have become a hot research direction in recent years,and many researchers have conducted in-depth research on it and achieved a lot of research results.But at the same time,it still faces many challenges.In the actual complex application scene,it is interfered by the internal factors of the target itself and the external factors of the environment.Fixed model updating is often difficult to cope with the challenges brought by tracking time-varying factors(target deformation,target occlusion,complex background,etc.).This paper focuses on the research of the traditional correlation filtering algorithm and the siamese network series tracking algorithm in terms of model adaptive problem,and studies the embedded platform deployment problem of the siamese network algorithm for the actual application requirements.For some traditional correlation filtering algorithms that use a fixed learning rate to update the tracking model,updating the model indiscriminately in target occlusion,target deformation,and fast motion scenarios leads to model pollution problems.This paper proposes a Gaussian error peak ratio based on the Staple algorithm.Combined with the edge half-peak ratio method to estimate the Gaussian kernel scale,it generates a standard Gaussian peak to calculate the energy distribution of the response map with the standard Gaussian distribution,which dynamically evaluates the tracking quality,adaptively adjusts the learning rate,and reduces the model pollution problem.Experimental results on the OTB100 data set show that the proposed algorithm improves the overall average tracking accuracy by 3.5% compared to the original algorithm.In response to the problem that the Siamese series of deep learning tracking algorithms cannot adaptively update the model,this paper proposes an adaptive tracking algorithm based on time-domain learning on the basis of the SiamRPN algorithm,and uses the Convolutional Gated Recurrent Unit(ConvGRU)to learn how to predict tracking model from the target sequence,And we fuse time-domain information of tracking targets in historical frames by data-driven method,which transforms the static one-shot detection in the inference stage into a dynamic few-shot detection problem.As a result,it learns timedomain correlation dynamic prediction from historical tracking frame data to adapt to the stable tracking model of the next frame,which can solve the problem of adaptive update of the model in Siamese series tracker.Experimental results on the OTB100 data set show that the proposed algorithm model has higher accuracy and success rate compared to the original algorithm,and the average tracking accuracy is increased by 1.2%.In order to meet the needs of miniaturization and low power consumption in tracking application,this paper deploys the siamesse network tracking algorithm on Hi Silicon Hi3559AV100 embedded platform.Combined with the hardware characteristics,we designed the siamese network tracking algorithm deployment scheme,and deployed the quantized neural network model through the neural network inference engine(NNIE).With the memory reuse strategy,we optimize the model branch and search branch feature extraction process,and combine NNIE and ARM cores to implement the forward inference process of siamese network.The experimental results show that the performance of the deployed algorithm can meet the task requirements of tracking the moving target for the ground-space background.