Research On Target Tracking Algorithm Based On Trajectory Prediction And Correlation Filtering

The tracking of moving targets has always been an important branch of the field of computer vision.In an increasingly complex environment,it is still a great challenge to ensure the accuracy and speed of tracking at the same time.Correlation filtering(CF)target tracking algorithm has always been very popular in the tracking field,mainly because he has the advantage of processing speed and good tracking effect.This article is based on the design of the correlation filter tracker that can take into account both accuracy and speed in a complex environment,that is,to improve the accuracy of tracking while achieving real-time effects.The main work content is as follows:(1)The kernel correlation filter algorithm(KCF)has high computational efficiency,It uses the fast Fourier transform to make the algorithm further expand the high-dimensional features while maintaining real-time.but its shortcoming is that the tracking performance is poor in a complex environment.All cyclic shifts use the period assumption to introduce unnecessary boundary effects and lead to candidate sample responses.The degree of calculation is large and the response value result is low.Therefore,KCF has been optimized.Based on KCF,a trajectory prediction algorithm is designed to predict the trajectory of the target in advance,determine the direction for the next target tracking,reduce the capacity of the sample set,reduce the cycle period to avoid unnecessary boundary effects,and improve the accuracy of target tracking degree.(2)In the relevant filtering algorithm,the model update is updated after the end of each frame tracking.When the target is occluded or deformed,a model drift may occur,which may cause the subsequent frame to fail to track the target.Therefore,the update of the later model is based on multi-peak detection,and the APCE criterion in the correlation filter LMCF(Large Margin Object Tracking with Circulant Feature Maps)is introduced.According to the result of APCE criterion,the model should be updated to reduce unnecessary model updates,improve the target tracking speed and reduce the possibility of model drift.Finally,the experiment was conducted on the OTB100 data set.Experiments show that the tracking algorithm based on trajectory prediction and correlation filtering has better performance than the existing target tracking algorithms in the benchmark test sequence,and has better performance in short-term occlusion properties.It runs at over 80 frames per second,Compared with KCF,the accuracy is improved by 7%,and the success rate is increased by 13%.