The Research On Meta Learning Based Multi-domain Network Target Tracking Method

Target tracking is an important research issue in the field of computer vision.It has broad prospects and needs in the fields of intelligent monitoring,automatic driving,humancomputer interaction and national defense and military.Due to the deformation,viewpoint change,occlusion and motion blur of the moving object,the target tracking is still facing great challenges.Target tracking method based on multi domain convolutional neural network(MDNet)has been widely concerned because of its multi domain network structure and good performance,but it still faces the shortcomings of slow training speed and weak generalization performance.In this paper,MDNet and its improved tracking model are taken as the research object,combined with the meta learning method based on deep learning model,the advantages of its fast adaptive ability are used to learn the depth features with more discriminative ability and generalization ability,so as to improve the tracking performance.The main contents of this paper are as follows:(1)Aiming at the problem that the performance of tracking is decreased due to the inconsistency of the pre training set and the tracking video,a multi domain target tracking method based on prototype attention is proposed.The real-time multi domain network target tracking method(RT-MDNet)is taken as the research object.In the training process,the prototype network method is introduced to make the model domain adaptive.The positive and negative samples based on the tracking results get the attention of the target prototype,and then integrate them with the feature map of the video to be tracked by channel adaptively,which makes the model get the target representation with stronger discrimination power on the large data set,thus enhancing the performance of the tracking algorithm.The experimental results on OTB100 and Tracking Net data sets show that the accuracy and success rate of the proposed method are improved by 2.2% and 2.7% respectively,and have competitive ability compared with other representative methods.(2)Aiming at the problem that the performance of model agnostic meta learning(MAML)method is limited by randomly initializing the network head,an improved method HILDEMAML based on linear discriminant is proposed to initialize the classifier head.The closed form solution of the linear ridge regression model is used to initialize the inner loop classifier head.And the closed form solution adopts Woodbury theory,which significantly reduces the computational complexity of matrix inversion in the solution model,and greatly improves the solution speed.The initialization classifier head with closed form solution enables MAML to learn quickly to adapt to new tasks,and prevents the optimizer from paying too much attention to the updating of classifiers in the initial iteration,thus speeding up the updating of network body parameters.Experimental results on four benchmark datasets of small sample learning show that the proposed method significantly improves the performance of MAML and its variant algorithm MAML++,and is superior to other representative methods in the field of meta learning.(3)Aiming at the problem that meta-tracker based on MDNet can not adapt to new tasks quickly because of random initialization,a multi domain network tracking method based on linear discriminant initialization meta-learning(HILDE-SDnet)is proposed.Using the linear discriminant initialization head instead of the original random initialization head can quickly adapt to the new task learning,and enhance the robustness of the tracking model to the target deformation.The experimental results show that this method improves the initialization speed of the tracker and has better tracking performance.