State Prediction With Evaluation For Object Tracking In Intelligent Visual Computing

Object tracking is a key and core technology in intelligent vision,and it is also an important and challenging frontier research topic.Visual tracking is widely applied to complex engineering projects such as intelligent transportation systems,security monitoring,multimedia,robotics,etc.The state calculation approaches developed in this research aims to achieve accurate and stable automatic tracking of target objects in video sequences,and the main contents can be summarized as follows:(1)A visual target tracking method based on partition loss calculation is proposed.This method uses incremental discrete cosine transform and structured information loss representation.It can adaptively divide the sample image sequence into blocks combined with the partition coefficients,to strengthen the positive effects of more discriminative local patches within the locating foreground region of the target and weaken the negative impacts of disturbing context possibly included in the surrounding background area.The online state prediction process is to carry out the solution of the maximum a posterior through likelihood evaluation within the probabilistic inference framework of particle filtering.(2)A visual target tracking method based on dual-directional scaling calculation is proposed.This method uses discriminative correlation filtering model and multichannel image sample representation.It can adaptively perform efficient scale search along the horizontal and vertical directions according to the uniform or non-uniform scaling changes of the target appearance.The optimal sizes in different directions can be obtained through the evaluation of the scale state detection response and the conversion of the dual-directional scale to size module.The rectangular target bounding boxes with dynamic aspect ratios can be predicted online in real-time.(3)A visual target tracking method based on online fusion calculation is proposed.This method uses the evidential reasoning analysis strategy and decision-level system cue collaboration.It can adaptively exploit discriminative characteristics and complementary properties of multiple tracking predictors to overcome the limitation of a single deep learning scheme in dealing with complex environments and difficult situations.Through the online evaluation and estimation of target center and region,the potentially valuable decisions are fused,the decisions that may fail are suppressed,and the prediction results of the target state are refined in the tracking process.The proposed tracking algorithms can effectively enhance the state prediction ability of target tracking in real scenes,reduce the drift deviation of position during visual tracking,and flexibly cope with the free change of the width and height of the target region.Experimental results of qualitative and quantitative comparisons on object tracking benchmark data show the accuracy and robustness of the proposed visual target tracking approaches.