Research On Video Object Tracking Algorithm In Complex Scenes

Visual object tracking is an important research direction in the current computer vision field and plays an important role in artificial intelligence and big data applications.The task of object tracking is to estimate and locate the target state in subsequent frames of a video sequence based on the initial state of the object in the first frame.Over the past decades,a large number of theories and algorithms on object tracking have been proposed and the performance of algorithms have been continuously improved.However,when the object tracking problems in industrial application scenes are dealt,unpredictable interference factors always occur,the effectiveness of the algorithm are affected,and great challenges to the tracking are brought.How to further improve the tracking performance of the object tracking algorithm in complex scenes such as deformation,low illumination,background clutter,fast motion,occlusion,and low resolution,and achieving a balance between the real-time performance and robustness of the algorithm remains urgent problems to be solved.The study takes the realization of object tracking in complex scenes as the main research line,drives by the relevant knowledge of generative and discriminative models,and aims to improve the precision rate,success rate and robustness without affecting the real-time performance to design the object tracking algorithm.The main research in the study is summarized as follows:(1)To solve the problem that traditional Camshift algorithm can not achieve effective object tracking in complex scenes because of its simple design of appearance representation model and repositioning strategy,a generative tracking algorithm is proposed,which combines improved local texture features with auxiliary relocation,to optimize the object appearance representation model and improve tracking accuracy.A local texture feature extraction mode based on improved particle swarm optimization algorithm and enhanced correlation between neighboring pixels and central pixels is designed.A dynamic weighted multi-feature fusion scheme is designed according to the correlation between feature contribution and Bhattacharrya distance,and the final position of the object is estimated by combining the candidate regions of multi-feature convergence with Kalman position compensation model.When tracking drift or failure,the external similarity between the current frame and the historical target template and the internal similarity of the current frame candidate region are studied,and the historical tracking traces are effectively used to provide reference for object relocation.The experimental results show that the proposed algorithm improves the distance precision rate and overlap success rate of the existing models,and performs well in tracking accuracy.(2)To overcome the shortcomings of correlation filter in considering the correlation and diversity of input features when dealing with boundary effect problems,a correlation filter tracking algorithm based on dynamic spatial regularization and target saliency guidance is proposed to optimize the feature selection model of the filter and improve the overlap success rate.Starting with the construction of the filter target function with spatial regularization matrix,a more discriminative filter model with time series constraints is learned.Based on the analysis of the intrinsic correlation between different feature characterization capabilities and filter responses,the object tracking results are calculated by the filter model.When tracking drifts or fails,the previous frame and the first frame are used as the input guidance of target saliency detection to obtain better re-detection results.The experimental results show that the proposed algorithm improves the overlap success rate of the existing models,and has some advantages in tracking success rate and real-time performance.(3)To solve the problem that a single correlation filter tracking model is sensitive to background interference,low resolution and other complex scenes,a correlation filter tracking algorithm coupled with multi-feature and scale adaptation is proposed to optimize the multi-feature filter model and improve the precision rate.By constructing a multi-feature coupled filter objective function,the Lagrange function is used to optimize the objective function,and two separate discriminative filters are trained to estimate the target location according to the correlation between the contribution of different features and the corresponding maximum response value.The average peak correlation energy introduced to judge the oscillation degree of the filter response and the maximum filter response are combined to update the target model,and the candidate region suggestion scheme is combined to effectively reduce the tracking drift.The experimental results show that the proposed algorithm improves the distance precision rate of the existing models.(4)To overcome the shortcomings of traditional correlation filters in modeling and target appearance representation,a low-rank correlation filter tracking algorithm based on hierarchical convolutional features is proposed to optimize the modeling method of the filter and improve the robustness.Starting with the mathematical modeling of the filter,the objective function is designed by using the Lasso regression modeling to learn a sparse and low-rank filter.By analyzing the different layered features of convolution neural network to represent the target,the target is coarsegrained positioned with the advantage of rich semantic information in the high-level features,and then the target is precisely positioned with the advantage of rich location information in the low-level features to achieve dual complementary enhanced tracking.The experimental results show that the proposed algorithm enhances the interpretability and robustness of the filter.