| Target tracking refers to using a specific method to mark the spatial position of the target object in the video image sequence to obtain the motion trajectory of the target object.As a research hotspot in the field of computer vision,target tracking technology can be applied to the fields of intelligent robots,unmanned driving and so on.This thesis focuses on the single target tracking technology for the purpose of taking into account the tracking real-time performance of target tracking system and improving the robustness of the system in the face of various environmental factors,and completes the integration implementation of the target tracking system in the intelligent robot platform.Therefore,the research content of this thesis has important practical significance and application value.Firstly,this thesis introduces the research status of target tracking at home and abroad,summarizes the research points of related work in this field and the shortcomings in the current development stage,and clarifies that the research focus of this thesis lies in the design of network models and optimization strategies.Based on the above contents,this thesis completes the overall design scheme of the target tracking system,and introduces the relevant datasets,performance evaluation indexes and target initialization methods involved in the research process.Aiming at the problem that the traditional Siamese network tracking methods usually only use the deep features containing semantic information to complete similarity matching,and do not make full use of the shallow features containing detailed apparent information and the middle features that play a transitional role between the shallow apparent features and the deep semantic features,so the ability of the tracking system to distinguish between targets and backgrounds needs further improvement in the face of complex situations such as target occlusion,motion blur and low resolution,a Siamese network tracking method based on shallow-middle-deep three-level feature fusion is proposed in this thesis.Firstly,the corresponding feature maps are extracted from the shallow,middle and deep network layers and fused at the feature level to obtain the complementary feature map,so as to achieve a more discriminative feature expression for the tracking target.Then,a pair of similarity score maps are obtained from the complementary feature maps of the two branches of the Siamese network through correlation calculation.Finally,the similarity score maps are fused to get the final score map,so as to achieve accurate positioning and tracking of the target.The experimental results show that the Siamese network tracking method based on shallow-middle-deep three-level feature fusion fully utilizes the advantages of image features at different network levels,and has good tracking performance on the basis of ensuring real-time tracking.Aiming at the defects of the similarity matching mechanism of the Siamese network and the problem that the traditional Siamese network tracking methods usually use cosine window filtering leading to its limited suppression effect on the interference of the target similarities,an adaptive rectangular window filtering optimization strategy based on clustering is proposed in this thesis.Firstly,construct a plane rectangular coordinate system in the score map to quantify the score map,find the peak point in the score map,and use the peak point as the initial cluster center to complete the clustering.Then,by calculating the distance between the cluster center and the farthest cluster point,the size of the rectangular window is controlled by the distance to realize the adaptive change of the size of the rectangular window.Finally,due to the strict time series relationship between adjacent video frames and the continuity of the target motion state,the motion direction of the target can be predicted.Because the actual motion of the target has a mapping relationship with the change of the position of peak point in the score map,the score map can be filtered along the target motion direction to realize the adaptive change of the filtering direction of the rectangular window.The experimental results show that the adaptive rectangular window filtering optimization strategy based on clustering can effectively overcome the interference of target similarity and further improve the robustness of the tracking system.Finally,the control system of intelligent robot based on target tracking is constructed,and the integration of related functions to control the motion state of the robot by using the target tracking results is completed.The relevant experimental results are recorded and analyzed,which proves the effectiveness and practical value of the target tracking method proposed in this thesis. |
