Optimization And Design Of Kernel Correlation Filtering Algorithm For Embedded System

With the development of 5G communications,the Internet of things and artificial intelligence,the computer vision is widely used.As one of the important directions in computer vision,target tracking has high academic value and economic value.In recent years,with the increment of target tracking scenario,the tracking system developed towards miniaturization,low power consumption and high robustness,while the current tracking algorithms only deployed on cloud or high-performance platforms.therefore,it needs to be further optimized to meet the needs of actual scenarios.By comparing the performance of various target tracking algorithms,the kernel correlation filtering algorithm is selected as the base of this paper.Combing with the requirements of target tracking algorithm,two optimization method are proposed to optimize the algorithm.Firstly,an adaptive fuzzy optimization based on classification threshold obtained by systematic clustering analysis was proposed,OTB-2015 data set was divided into high,medium and low clusters,and adaptive fuzzy optimization was applied to them respectively.Secondly the adaptive feature fusion and template updating method is adopted to improve the robustness of algorithm,meanwhile principal component analysis and kernel function optimization are used to reduce the computation,then the Average Peak-to-Correlation Energy and Peak Side Lobe Ratio are used to evaluate the confidence of the tracking results,which could realize adaptive update of tracking model and predicted position.Then the HOG feature acceleration unit is designed based on HLS tools.The unit can realize the computation of horizontal gradient and vertical gradient at the same time by the two layers of Line Buffer in one clock cycle.At the same time,the gradient amplitude comparison is realized by means of look-up table combinational logic.Compared with Software iterative approach,which greatly improves data processing speed.Finally,a hardware acceleration framework based on the PYNQ-Z2 platform was built through the collaborative development of software and hardware,the FPS of the optimized kernel correlation filtering algorithm under this acceleration framework is 17.72,and the tracking accuracy is 72.86%.Compared with the origin KCF algorithm on PS,the tracking real-time performance is improved by 4.4 times,and the tracking accuracy is improved by5.58%,which can meet the requirements of most target tracking scenarios.