Research On Object Tracking Algorithm Based On Feature Adaptive Fusion

In the era of intelligent development of life,intelligent products are emerging in various fields.The development of intelligence is inseparable from the support of computer vision.Object tracking is the most basic branch of research in the field of computer vision.It has important research value in the fields of intelligent transportation,human-computer interaction,and intelligent medical diagnosis.Object tracking is currently divided into target tracking algorithms based on correlation filtering and target tracking algorithms based on deep learning.Although there are many good tracking algorithms in both research directions,object tracking still has many difficulties and challenges in practical applications.In the tracking process,many external factors such as illumination variation,motion blur,background clutter,and occlusion will interfere with the positioning of the target.At the same time,the target itself will also interfere with the tracking,such as object scale variation,posture change,and deformation.Aiming at the problem of background clutter and occlusion interference,this paper uses feature fusion to enable the tracker to better locate the target under the background clutter interference,and develop a target tracking algorithm based on adaptive feature fusion.1)In the actual tracking process,the environment where the target is located may be more complicated,and there may be areas similar to the target that interfere with the positioning of the target.For example,there are interference items in the background that are similar to the target texture features or similar to the target color features.For such problems,improving the ability to express features is an effective solution.Considering that the deep features extracted by the convolutional neural network have strong representation capabilities,this paper uses the VGG network to extract the target features based on the correlation filtering based on the efficient convolution filter tracking algorithm,and performs weighted feature fusion to solve the background Complexity interferes with tracking accuracy.The algorithm uses the features extracted by the first five convolutional layers of VGG-M for feature fusion,improves the expression ability of the features,and enables the features to still locate the target under the interference of different environmental factors.In addition,considering that different features have different expression capabilities in a changing environment,and the increase in the amount of calculation brought by feature fusion.In order to achieve the real-time tracking effect,two-layer features are selected for weighted fusion to achieve a better tracking effect.Through experiments on the OTB-2015 data set and Temple color 128 data set and comparison with other algorithms,the experimental results prove that the improved adaptive feature adaptive fusion algorithm has a higher success rate than the comparison algorithm and can guarantee a certain tracking speed.2)In the process of object tracking for a long time,it is inevitable that the target will be blocked.Occlusion will cause the loss of target information,and tracking is to locate through sufficient target information.Once the occlusion causes the lack of target information,it will affect the tracking.If the filter is updated during occlusion to introduce error information into the template,the learned error information will cause tracking errors.Continuous updating of the filter may cause more and more error information to be introduced into the model,causing tracking drift.Will cause the loss of the target.Aiming at the problem of short-term occlusion,an occlusion detection mechanism is proposed.When there is occlusion,the confidence response value will change to determine whether there is occlusion.On this basis,combined with the characteristics of LBP features that are not sensitive to deformation,it is further judged whether the change of confidence change rate is caused by occlusion or self-deformation.If occlusion occurs,the filter is not updated.The experiment proved that the increase occlusion detection algorithm for tracking accuracy and success rate has improved over the occlusion problem.It has better tracking effect and more stable.