Kernel Correlation Filtering Target Tracking Algorithm Research And Embedded Implementation

Target tracking technology is an important part of the field of machine vision and has been widely used in military and civil fields.Visual tracking technology needs to face challenging scenes such as the occlusion of the target,illumination change,rapid movement,etc.At the same time,most of the existing computing platforms are systems with high power consumption and large volume.Therefore,it is necessary to design a robust multi-strategy target tracking algorithm,and to use the embedded development system with low power consumption and small volume.This paper focuses on the research and design of a multi-strategy target tracking system that can solve the problems in the actual scene,improve the kernel correlation filtering target tracking algorithm,and complete the algorithm transplantation on the embedded system to achieve the target tracking in the actual scene.The main work of this paper is as follows:1.Based on the kernel correlation filtering tracking algorithm,this paper designs a target tracking algorithm combining multi-feature fusion and multi-peak detection and update mechanism.Each feature dimension is used to allocate the weight of the feature,and the feature of each weight is linear fusion instead of a single feature.In addition,multi-peak confidence updating method is adopted to reduce the probability of tracking failure caused by wrong frame.Finally,multiple kernel filters are evaluated and the best one is selected from each frame of the video sequence as the tracking result.2.In this paper,the kernel correlation filtering tracking algorithm is transplanted to Zynq-7020 embedded system.In terms of hardware design,FPGA of the PL processes the video sequence collection,features processing and HDMI output display.ARM of the PS is responsible for the target tracking algorithm part to realize the prediction of the target position,and the communication between PS and PL part is realized through the AXI4?STREAM protocol to further improve the tracking ability.3.This paper tests the improved kernel correlation filtering algorithm on the OTB-2013 public dataset,and the success rate and accuracy are respectively 5.8% and 6.6% higher than that of the Kernerlized Correlation Filter tracking algorithm,and the embedded system can complete the real-time tracking of the target in the actual scene,the tracking speed is about 30 FPS.