Research On Improved Deep Correlation Filters Via Conditional Random Field

Visual Object tracking is one of the most important tasks in numerous applications of computer vision.It plays an indispensable role in a wide range of applications involving national defense and military,video surveillance,visual reconstruction and medical analysis,and contains enormous commercial value and great potential.It is challenging as objects often undergo significant appearance changes caused by deformation,abrupt motion,background clutter and occlusion.Therefore,to build a robust object appearance model for visual tracking is very important.Discriminative correlation filters(DCF)with deep convolutional features have achieved favorable performance in recent tracking benchmarks.The object in each frame can be detected by corresponding response map,which means the desired response map should get a highest value at the location of the object.However,the limitation of feature extraction leads to the multi-peak phenomenon of response map,which results in the target position prone to drift.To address these issues,we propose an improved deep correlation filters via conditional random field(CRF).On one hand,considering the continuous characteristic of the response values,it can be naturally formulated as a continuous CRF learning problem.Moreover,the integral of the partition function can be calculated in a closed form so that we can exactly solve the log-likelihood maximization.On the other hand,we can smooth the response map based on the similarity of neighboring superpixels.Thus we design an end-to-end deep convolutional neural network for estimating response maps from input images by integrating the unary and pairwise potentials of continuous CRF into a tracking model.With the combination between the initial response map and similarity matrix which are obtained through the unary and pairwise potentials respectively,a smoother and more accurate response map can be achieved,which improves the tracking robustness.We evaluate the proposed approach against 9 state-of-the-art trackers on OTB-2013 and OTB-2015 benchmarks.Qualitative and quantitative evaluations demonstrate that the proposed tracking algorithm performs favorably against several state-of-the-art methods.