Parallel Visual Tracking With Correlation Filters And Siamese Networks

Visual tracking is one of the fundamental problems in computer vision.In recent years,correlation filters-based trackers develop rapidly and attract much attention,but its limited recognition ability fail to deal with some complex scenarios.Meanwhile,convolutional neural network based trackers which benefits from a large number of training dataset,have been widely concerned due to its powerful ability of classification and recognition,but the complex model prevents them from being updated online,which cause the losing of domain specific information.How to effectively combine the advantages of these two methods is a research hotspot.In this paper,the verifier is used as a bridge to explore this hot spot in the way of multi-threaded parallel.The main tasks are as follows:(1)A three-branch dual-thread parallel tracking framework is proposed,and the correlation filtering algorithm of online updating model parameters is used as the main thread algorithm for fast tracking.At the same time,the second thread runs the verifier to verify the tracking results,and execute the correction algorithm to correct the tracking results after the reliability of the tracking results declines.The correction algorithm is divided into two sub-branches,which respectively deal with the scale inaccuracy of tracking box and the situation that target out of view.(2)Several factors affecting the effect of network features in the tracking algorithm are analyzed,and a more suitable feature network is obtained by adjusting the residual network by adding clipping operation,changing parameters such as convolution step size and so on.(3)A tracking model that can quickly update online is realized with full convolution network,and the optimization of the network is transformed into the optimization of gauss-newton sub-problem,which will implement by the back propagation function.In a broad sense,this model is equivalent to the correlation filter model.(4)The IoU prediction network is realized by using the siamese network structure to predict the IoU between the tracking box and the groundtrue during tracking,and the conjugate gradient method is used to iterate the boundary box to optimize the tracking results.Furthermore,this IoU prediction network serve as a weak corrector to correct small deviations such as scale deviation.Finally,a strong corrector is realized by siamese network with the regional pooling which continuously expands the search area.