RGBT Objects Tracking Based On Graph Neural Network

As one of the research focuses in the field of computer vision,object tracking can be widely used in the field of autonomous driving,monitoring and security.Combining visible light images(RGB images)and infrared images(T images)for target tracking can better adapt to various scenes of tracking,and can still achieve good tracking results in complex situations such as insufficient lighting conditions and foggy days.With the development of deep learning technology,the RGBT object tracking method based on deep learning has made great breakthroughs,but how to make more effective use of the useful information in the two modal images is still the key problem.At the same time,due to the absence of relevant datasets,RGBT multi-object tracking has developed slowly.In view of the above problems,this paper combines the graph neural network to study the feature extraction and fusion of RGBT single-object tracking,and designs a feature extraction model based on graph attention method.Aiming at the lack of RGBT multi-object tracking data,the method of infrared data generation is explored to solve the problem of insufficient training data.Considering the problem of RGBT multi-object tracking,the correlation method of RGBT data based on graph neural network is explored to solve the problem of multi-object tracking.The specific research content is as follows:(1)A single-object tracking network based on the Siamese Network is constructed for RGBT video.This paper uses the graph attention method to extract and fuse the information of RGBT video,which can improve the effectiveness of multi-modal data utilization and increase the tracking accuracy and success rate.In order to prove the effectiveness of the proposed algorithm,a horizontal comparison was made on the Las He R and RGBT234 public datasets to prove the effectiveness of the algorithm.(2)Considering the lack of RGBT multi-object data,which is difficult to meet the deep learning training,this paper adopts the method of image generation,and uses the generative adversarial network(GAN)to generate infrared data from the visible multiobject tracking dataset to meet the network training needs.At the same time,in order to measure the generation effect,this paper compares three image generation methods,and compares the Structural Similarity and the Peak Signal-to-Noise Ratio of the generated results.(3)Aiming at the RGBT multi-object tracking,this paper constructs a RGBT multiobject tracking model based on graph neural network.By fusing the visible and infrared image information,the model abstracts the object information in the continuous frame into a graph model,which uses the object spatial features and appearance features as the matching measurement criteria,and realizes the prediction of the associated edge through the information propagation mechanism of the graph neural network.In order to verify the effectiveness of the algorithm,the algorithm in this chapter was tested and evaluated by the CAMEL measured dataset,and a multi-target tracking accuracy(MOTA)of 68.2was obtained.