Research And Implementation Of Target Tracking Algorithm Based On Correlation Filtering

As the "eyes" in the era of artificial intelligence,computer vision is widely used in many fields such as machine learning,image signal processing and control engineering.Target tracking technology is the main branch and research hotspot of computer vision,is also applied in the fields of unmanned driving,military attack,video surveillance and human-computer interaction.The target tracking algorithm based on correlation filtering is a common visual tracking method.It takes the target region and the region to be measured as signals with the help of the correlation index reflecting the similarity between targets,and calculates the correlation between them.The higher the correlation is,the stronger the response of the filter model will be,and the response peak region will be predicted as the tracking target region.Kernelized Correlation Filter(KCF)algorithm is a classical correlation filter tracking algorithm with high accuracy and speed.However,the algorithm also has some problems,and the work carried out on it is as follows:(1)In the tracking process,the appearance model of the algorithm is not enough to deal with the diversity of the appearance of the tracking target,and the algorithm is not robust enough.The Gaussian mixture model was used to obtain the appearance sample set of multiple tracking targets.Based on the sample data generated by the model,the filter model was trained and updated to cope with the changes in the appearance diversity of targets and improve the robustness of the algorithm.In addition,the algorithm uses manual features with poor ability to represent the target to extract the target information,which makes it difficult to distinguish the target from the background information effectively in the complex scene.The depth feature of input image is extracted by Res Net-50 network layer,and the manual feature used by KCF algorithm is replaced to strengthen the characterization ability of target.(2)When the tracking target information is lost,the algorithm will continue to train until the target is completely lost based on a filter model with large error,and there is no tracking redetection.The tracking error is calculated and used as the tracking confidence to judge whether the tracking of two adjacent frames of video images is successful.If the calculated tracking error is larger than the set threshold,the tracking algorithm is judged to be a failure.On this basis,the tracking failure redetection mechanism is carried out,the current target prediction position is abandoned,and the target tracking box of the previous frame of image is regarded as the central point.The search area of the current frame target is enlarged to relocate the target For the above optimization scheme based on the Object Tracking Benchmark(OTB2015)data set,simulation experiments were carried out on the MATLAB platform,which proved the feasibility of the optimization scheme.At the same time,the optimized KCF algorithm is compared with some filtering tracking algorithms to ensure the reliability of the optimized.(3)Since the tracking target frame is always a fixed size in the tracking process of the algorithm,the performance of the algorithm will be degraded when the tracking target scale changes.For example,part of the target information will be lost when the scale increases and exceeds the target frame,and too much redundant background information will be doped when the scale decreases to a low proportion of the target frame.By adding an additional scale filter,when the tracking target is determined by the position filter,the algorithm takes the estimated target position as the center,carries out scale transformation of different proportions on the target frame,and operates the transformed sample data related to the scale filter,and selects the response peak sample as the final scale output.So that the tracking target frame can adapt to the change of target scale.The optimization ideas of KCF algorithm are summarized and simulation experiments are carried out on MATLAB platform based on Object Tracking Benchmark(OTB2015)data set,which proves the feasibility of the optimization scheme.At the same time,the optimized KCF algorithm is compared with a variety of correlation filtering and deep learning tracking algorithms to ensure the reliability of the optimization scheme.