Study On Correlation Filter Tracking Algorithm Based On Depth Convolution Feature

Object tracking is one of the important research directions in the field of computer vision,which plays an important role in intelligent video surveillance,human-computer interaction,robot visual navigation,virtual reality,medical diagnosis and other fields.The basic problem of object tracking is to determine the object to be tracked in a video or image sequence,locate the accurate position of the object in the subsequent successive frames,and generate the object trajectory.In practical application,the object's apparent model is often changed due to non-rigid changes during the movement of the object,and the object tracking task becomes more difficult due to complex light changes,scale changes,fast movement,motion blur,object being blocked for a long time and interference of similar objects in the background.Although object tracking technology is becoming more and more accurate,object tracking is still challenging in terms of stability and timeliness.In this paper,the correlation filtering tracking algorithm based on depth convolution feature is studied,and the feature-level fusion method is used to improve the object tracking performance.The main work of this article is summarized as follows:1.A correlation filters tracking algorithm based on feature level product fusion of deep convolutional networks is proposed.In order to take advantage of the complementary advantages of convolution features at different levels,feature-level product fusion method is used to accurately locate the position of the object from coarse to fine,thereby improving the robustness and accuracy of the tracking.At the same time,the trajectory fluctuation and peak side lobe ratio are used as tracking confidence indicators to judge whether the object tracking has failed.The principle of learning rate update is introduced to update correlation filters,which effectively solves the problem of instantaneous drift or incorrect prediction of the object position due to insufficient online update.2.A correlation filters tracking algorithm based on depth convolution feature adaptive integration is proposed.By using the idea of integrated learning and the hierarchical convolution characteristic of deep convolutional neural network and the adaptive integration of the convolution characteristic of different levels,the purpose is to combine the different trackers to get a stronger tracker,which can effectively solve the problems such as the performance instability and scale undervariation of the tracker in the long-term tracking process.At the same time,an online learning mechanism is used to continuously update the integration coefficient and related filter parameters of feature maps at different levels in the tracking module,so as to make the tracking effect more stable,robust and reliable.3.An object tracking algorithm based on response graph multi-peak detection is proposed.On the basis of the existing object tracking algorithm,by analyzing the peak distribution of the response graph,the average peak correlation energy and trajectory volatility are used as the tracking confidence index to detect the tracking state of the object,and determine whether it is necessary to track again to correct the tracking position.Compared with the improved algorithm,the proposed object position correction strategy can effectively improve the tracking accuracy.