Fast Deep Tracking Via Semi-Online Domain Adaptation

Visual tracking is one of the main branches in the field of computer vision,and it has a very wide application prospect in real life.Most of the traditional tracking algorithms use image features of manual design,such shallow learning algorithms are limited to the ability of image feature representation,thus it has some limitations in practical application scenarios.The tracking algorithm based on deep learning can automatically learn feature representation which is more adaptable to the current task from the current data.Therefore,it is more robust for scenarios such as extreme light change.However,the gradient value of each parameter in the network needs to be solved in the process of backward propagation of neural networks,therefore,the tracking algorithm based on deep learning is often difficult to be real-time.To alleviate the problem,a number of real-time deep trackers have been proposed via removing the online updating procedure on the CNN model.However,the absent of the online update leads to a significant drop on tracking accuracy.In order to explore high tracking accuracy and real-time tracker,we propose a semi-online domain adaptive tracking algorithm based on the research of early vision tracking,combining with the deep learning method and correlation filter algorithm.The main research work is as follows:1?This paper proposes a lightweight deep neural network as the basic framework,which has a shorter time to solve the inverse gradient.At the same time,a tracking branch is added to the three convolution layers of the network as the domain adaptive layer,in this way,not only guarantee the multi-scale characteristics of the feature,but also effectively transfer the feature information of the image classification domain to the visual tracking domain,and improve the adaptability of the tracker.2?Combining with deep learning method and correlation filter algorithm,we use deep feature to learn correlation filters,and locate the target quickly by searching the maximum response value of filter response map.3?Different from the previous update strategy,we propose a semi-online update strategy,the first ten frames of video are taken to adjust the domain adaptive layer of the network.In this way,not only can effectively improve the network adaptability of current video sequence to ensure the tracking accuracy,but also does not increase the computational burden of the network to ensure the running speed of the network.In order to verify the effectiveness of this method,the experiment is carried out on the visual tracking public data set(OTB).Through the analysis of the test results in a variety of specific video sequences,it shows that the proposed visual tracker achieves comparable tracking accuracy to the state-of-the-art trackers and runs at real-time speed.