Object Tracking Based On Correlation Filter And Deep Model Compression

Object tracking is an important research topic in the field of computer vision.Recently,the correlation filter based object tracking algorithm has been paid much attention in the field of object tracking because of its high accuracy and high speed.In particular,the combination of deep learning technology and correlation filter makes great progress in the performance of correlation filter based object tracking algorithm.However,with the improvement of algorithm precision,its robustness and real-time performance are still difficult to meet the requirements of real scenes.On the basis of correlation filter,deep learning and deep model compression,this paper focuses on the tracking algorithm's robustness and real-time requirements,as well as the efficiency of target feature expression.The specific research contents are summarized as follows:1)An adaptive object tracking algorithm based on context-aware correlation filter and Siamese network is proposed.This algorithm makes use of the complementary advantages of the object tracking algorithm with different tracking mechanisms,namely the context-aware correlation filter and the fully convolutional Siamese network,to alleviate the shortcomings in the tracking process and construct a robust tracking algorithm.Specifically,our tracking algorithm consists of three components: the context-aware correlation filter network,the fully convoluted Siamese network and the tracking result reliability detector.The off-line training of the context-aware correlation filter network and the full convolution Siamese network is used to obtain a more suitable feature representation of the object tracking task.The existence of the tracking result detector prevents the model from introducing the wrong target information and avoids the interference between the two different mechanisms.In the tracking process,the tracking result detector determines the reliability of the tracking result of the correlation filter,updates the model adaptively,and starts the Siamese network to cooperate with the context-aware correlation filter network for robust tracking.2)A real-time correlation filter based tracking algorithm for joint model compression and transfer is proposed.Although the use of deep convolution featureimproves the performance of correlation filtering algorithm,it also brings problems:high memory storage,high feature extraction time,and high dimensional convolution feature increases the learning time of correlation filter.These problems result in high-precision tracking algorithms that cannot be deployed on platforms such as those with single-core CPU and low memory storage.In this work,under the framework of knowledge distillation,we take the existing high-dimensional deep convolution model from the image classification task as the teacher network,and compress and migrate the teacher network joint model into a lightweight student network suitable for the relevant tracking task.The network has a small capacity,which is combined with the correlation filter to track the target,which speeds up the time consumption of feature extraction and the learning time of correlation filter.Experiments show that the lightweight student network can maintain almost the same tracking accuracy while enabling the algorithm to achieve real-time speed on a single-core CPU.