Object Tracking Algorithm Based On Correlation Filtering And Metric Learning

Object tracking is to continuously predict the size and location of the target in a subsequent frame which is given the size and location of the target in the initial frame of a video sequence.Object tracking algorithm is widely used in video surveillance,intelligent transportation,intelligent human-machine interaction and other scenarios.We do not limit the types of targets and requires high accuracy.In video sequences,the scene of target motion is complex and often changes.At the same time,the target itself is prone to deformation,out-of-plane rotation and occlusion.So the task of target tracking is very challenging.So far,many research work has been proposed.Correlation filtering and deep learning algorithms have been widely used in previous research work and have achieved good results.The correlation filtering algorithm can update the model online,but the model only uses a single mat rix to represent the target,which leads to its poor discriminant ability.Deep learning algorithm can use a large number of offline samples to complete off-line training,but the overconsumption of online training time leads to slow tracking speed and uncertain target types,which makes feature extraction difficult to train.In this paper,we try to combine the two methods,and propose a target tracking algorithm based on correlation filtering and metric learning to solve the problem of deformation and out-of-plane rotation in target tracking.In this paper,correlation filtering algorithm is used to extract target candidate positions,and then metric learning network is used to calculate the similarity of samples in the target and candidate regions to find the specific size and location of the target,so as to achieve tracking.In the correlation filtering algorithm,background information is added to restrain the effect of target deformation and out-of-plane rotation.In the training of correlation filtering model,the samples similar to the target are first found and then trained as negative samples to the correlation filtering model.This can prevent the model from drifting when the target changes,thus optimizing the positioning effect.Finally,this paper establishes a metric learning network,and further judges the target position through the network output vector,so as to obtain the final tracking results.Among them,the metric learning network can better capture the contour information of the target.When the target is deformed or rotated out of plane,the contour information can be used to accurately determine the position of the target,so as to improve the tracking accuracy.Experimental analysis verifies the effectiveness of the proposed framework.The results of correlation filtering,metric learning network and other research points are validated,and grouping experiments are conducted according to different tracking scenarios.The overall accuracy and precision of OTB2015 data set were improved by 4.3% and 5.1% respectively.