Specific Target Tracking Based On Detection With Deep Learning And Kernelized Correlation Filters

In this society,the role of target tracking in the field of computer vision is growing,and the development of target tracking related algorithms is also very rapid.Target-specific tracking is an important non-branch in the target tracking field.It generally has the following requirements:the video stream used for target tracking,the target position and size of the initial frame.The task of a specific target tracking is to accurately track the position and size of the target in subsequent sequence frames.The target tracking algorithm has also undergone a long development.From the generation model to the discriminant model,the performance of the model has also been greatly improved.In this paper,a simple analysis and research on some target tracking algorithms in recent years is carried out,and the KCF algorithm is selected as the basic algorithm of this thesis.Then the algorithm improvement scheme is proposed for the shortcomings of the basic algorithm.The specific contents include:introducing the target loss early warning mechanism,the model update strategy based on the early warning mechanism and introducing an auxiliary detection algorithm.The target loss warning mechanism can calculate the PSR and PSRP values by using the response graph,and determine whether the current tracking state is stable by comparing the PSRP value with the set threshold.When the tracking status is determined to be unstable,a new model update strategy is initiated.When the target is lost or occluded,the underlying algorithm is incapable of retrieving,so a detection algorithm is introduced to assist the correction.In addition,the introduction of detection algorithm for target tracking compensation can also solve the problem that the basic algorithm lacks scale transformation,so that the algorithm satisfies the requirements of scale transformation.The experimental results show that under the three complex tracking scenarios of target loss,background confusion and occlusion,the overall effect of the improved algorithm is optimized compared to the initial algorithm.