Saliency-Based Tracking Algorithms

Target tracking is the basic issue of computer vision,which has a wide range of applications included video surveillance,human-computer interaction,intelligent transportation and robot navigation.In recent years,Researchers are doing in-depth study from the generation model and discriminant model,trying to track a general object in wild environments.Although great progress has been made in recent years,there is still much room for improvement.Due to the saliency of tracking object,that is the object has movement characteristics and visual characteristics such as texture,color different from environment,this paper has done a lot improvement.Specifically,tracking object has motion significance,which means that the tracking object has displacement changes during the tracking process.In this paper,we propose Moving Target Focused Tracker(MTFT)to improve efficiency in single target with large displacement scene by detecting moving objects to coarsely position the tracking object.Extensive experiments on TB50 show that the proposed MTFT improve a little in single target with large displacement scene,but the comprehensive performance decreases a little.Additionally,tracking object has visual significance,which means that the tracking target tends to distinguish in color distribution from the environment.The General Target Focused Tracker(GTFT)is proposed to improve efficiency in general scene by obtaining the foreground probability to enhance the foreground and suppress the background.The proposed GTFT has achieved 82.8% accuracy in TB50,3.5% higher than Staple.In addition,we propose the General Target Focused Long-term tracker(GTFLT)in order to solve tracking failure in GTFT.GTFLT use color confidence related to obscures and illumination,and tracking confidence related to deformation and rotation to judge whether the object is lost or not.At the same time,GTFLT updates the tracking object discriminant model online through the reliable object to suppress the possible error of detection.Compared with the state-of-art methods in the last two years,GTFLT has achieved the best results in TB50,got 89.6% accuracy.And GTFLT achieves the best results in 8 categories of sequences among all 11.