Research On Video Object Tracking Algorithm Based On Self-Supervised Learning

Video object tracking is one of the basic tasks in computer vision and is widely used in security,human-computer interaction,virtual reality,mobile robots,autonomous driving and other fields.According to the different feature extraction methods,video target tracking algorithms are divided into two-stage and one-stage.The two-stage video object tracking algorithm contains two computationally intensive modules,the detection network and the feature extraction network,so it is difficult to achieve real-time performance.The one-stage method that appeared later,put target detection and feature extraction on a shared weight convolutional neural network for multi-task learning,but it lost tracking accuracy while reducing time consumption;especially in scenes with dense targets,the mutual occlusion will cause the identity tags(Identity,ID)assigned to different targets to switch,resulting in a decrease in the ID-Switch indicator.In this regard,this paper improves the one-stage method based on the feature extraction of self-supervised learning,and mainly carries out the following work:(1)The central differential convolution is introduced into self-supervised learning,and the influence of different convolution weights on downstream tasks is explored.Without increasing inference time,it improves the performance of self-supervised learning on downstream classification tasks by 4.14%.Experiments show that the features extracted by self-supervised learning based on central difference convolution are more robust than those extracted by self-supervised learning based on ordinary convolution.(2)A feature extraction model that can be used for video target tracking is trained using self-supervised learning based on central difference convolution;due to the characteristics of self-supervised learning,the range of data sets that can be used by the paper method is greatly increased,and some targets can be detected.datasets are included in the scope of use.The final experiment proves that the feature extraction network based on self-supervised learning training is 3.3% higher than the ordinary method when using the same labeled data.(3)Aiming at the problem that the one-stage video target tracking algorithm causes the tracking accuracy to decline,this paper improves the feature learning method on the basis of the one-stage method;the feature extraction network trained by self-supervised learning is used for each target in the dataset.Generate feature labels to guide the detection network to learn features.The improved method reduces the frequency of ID-Switch by 9.02%.