Research On Object Motion Tracking Based On Monocular Camera

In the field of computer vision,deep discussion and exploration about object motion tracking algorithms based on monocular video have always been one of important research topics and hot spots.It can also be called visual tracking.As the current domestic and foreign research status shown,the tracking field has made some new advances with the introduction of machine learning.There is a large number of sampled data that make deep learning occupy a dominant position,but the accurate sample that the tracking algorithm can obtain in the video is only the labeled data of the first frame,so the results obtained by offline training alone will bring great errors to the tracking;the network structure in the tracking algorithm based on deep learning is more complicated than usual,and the model parameters are various.The computational complexity of the online training network increases accordingly,which largely forces the efficiency of tracking to be slow.In summary,the difficulty and emphasis of designing visual tracking algorithm is the balance between accuracy and efficiency.Aiming at the above problems,this paper studies and implements the object motion tracking method based on deep learning under the condition of monocular camera and the main contributions are as follows:1)Improved algorithm based on fully convolutional siamese network.An online update mechanism is proposed to integrate the consistency constraint of the target motion with the fully convolutional siamese network for visual tracking,which makes the target motion tracking more robust and real-time.The model is added loop constraints between consecutive multiple frames,on-line learning the differences in characteristics of tracked object during movement.Considering the cumulative error of model updating in tracking,a detector for estimating the best template of the next frame is set to verify whether the model is consistent with the features of the target continuously,thereby reducing the cumulative error of online learning.2)Tracking algorithm based on improved discriminant model.In order to enhance the discriminative function of the network,a tracking algorithm is proposed which based on discriminant models.The algorithm uses target-specific features for discriminative learning through score fusion to help the classification network deal with interference factors and background noise.A feature pyramid model is added to the algorithm framework,and the corresponding different feature spaces are selected according to the scale of the candidate region.Because the parameters of different layers are flexible and supervised to achieve end-to-end training at the same time,the region of interest derived from the additive training can effectively combine shallow and deep object features.The experiments in this paper are based on the OTB-2013 and OTB-2015 evaluation data sets.The algorithm test results are visually expressed by various indicators on different attribute video sequence sets.Their accuracy and success rate indicators are similar to those based on deep learning or correlation filtering.The algorithm results perform well in comparison,and verify the effectiveness and robustness of the algorithm in this paper from qualitative and quantitative analysis.