Tracking Based On Detection With Deep Learning And Kernelized Correlation Filters

Nowadays,computer vision is used in various fields of people's life,which aims to replace human visual tasks with automated means.In this,the target tracking task has developed rapidly.Single object tracking is a basic task in computer vision.It usually includes the following initial settings:an initial target tracking location and a tracking box size;a video stream to track.The single target tracking algorithm will find the initial target in each frame of the video stream and draw the tracking frame.Because of its complexity,the target tracking task requires the algorithm to be robust and real-time.After a long period of development,the performance of the mainstream target tracking algorithm from the generation model to the discriminant model is also increasing.Based on the research and development process of tracking algorithm is summarized and analysis of a simple one,and pick one of the best algorithm improvements include:the introduction of a detection method for tracking correction,in order to solve the defect scale based algorithm,the tracking algorithm with scale change;and put forward a method based on tracking model confidence to achieve automatic model updating strategy,this strategy can solve the basic algorithm in background confusion and occlusion scene prone to loss of tracking shortcomings to a certain extent.The performance of the improved tracking algorithm has been greatly improved compared to the performance of the original algorithm.