Selective Visual Object Tracking With Dynamic Spatial Regularization

Visual object tracking is playing an increasingly important role in image recognition.Given exact target location of video sequence in the first frame,tracking methods are required to learn the information and predict target locations in the following frames.Traditional tracking methods based on correlation filter have actually achieved relatively high tracking performance.However,they are generally sensitive to target foreground appearance model.When target foreground appearance model is not reliable,these methods usually lose to track the target.Therefore,we propose to adopt selective object and context tracking(SOCT)to enhance tracking performance.Based on correlation filter,our approach dynamically track foreground or context of the target by judging the reliability of target foreground appearance model.We propose two metrics to judge the reliability of target foreground appearance model: occlusion and multiple similar objects around the target.In the first frame,we apply the initialized correlation filter to the search region and get a response map,by which we train the Ridge Regression model.In the following frames,we use the model and the distribution of response map in current frame to judge whether target foreground appearance model is reliable.If reliable,we use spatial regularization to penalize filter coefficients reside in the background.If not,we then enable context tracking and train a new context filter which use spatial regularization to penalize filter coefficients in the center region.Context tracking is continuous until target foreground appearance is reliable again.In addition,we use Gauss-Seidel method to quickly solve the quadratic minimization problem extracted from tracking problem.Comparing with VOT dataset,OTB dataset and some natural video sequences,experiments show that SOCT achieves better performance than 11 state-of-the-art methods.Especially when it comes to the sequence with attribute like occlusion and multiple similar objects around the target,SOCT can always accurately locate the target,while others may easily lose to track the target.