Research And System Implementation Of Multi-object Tracking Based On Attention And Cross-frame Correlation

Multi-object tracking technology involves pattern recognition,machine learning,object detection and other research fields,and has a wide range of applications in urban security,automatic driving,intelligent navigation and other fields,having great research value.The main task of multi-object tracking is to connect the targets in a frame and them in the next frames to form continuous tracking trajectories by performing position detection and identification calibration of multiple objects in a continuous video sequence,and keeping the object identities unchanged during the tracking process.Aiming at the problem of feature discrimination decline caused by object interaction occlusion and the identity switching problem caused by the long-term disappearance of objects,this thesis proposes an anti-occlusion multi-object tracking algorithm based on attention from the perspectives of feature extraction and motion prediction.The attention mechanism enhances feature discrimination and motion prediction accuracy,improving the accuracy of multi-object tracking.This thesis also proposes a multi-object tracking algorithm based on cross-frame correlation from the perspectives of similarity calculation and data correlation.It uses a frame as a unit to associate data crossing frames on the basis of the correlation between frames and maintains the identities of the objects that disappear for a long time,improving the stability of multi-object tracking.In this thesis,the research on feature extraction and motion prediction,similarity calculation and data correlation is carried out as follows:1.An anti-occlusion multi-object tracking algorithm based on attentionAiming at the problem of inaccurate motion prediction and decrease of appearance feature discrimination caused by interactive occlusion between tracking objects,this thesis proposes an attention-based multi-object tracking algorithm.A multi-scale dual attention network is used to extract features with focused attention areas,and a tracking attention and occlusion attention mechanism is established to enhance the features of the tracked objects while suppressing the features of the occluding objects.The timing information aggregation network fuses timing information to further enhance the discrimination of features.Through a series of experiments,it is verified that the attention-based multi-object tracking algorithm proposed in this thesis can effectively alleviate the problems of feature discrimination decline and motion prediction inaccuracy in the object interactive occlusion scene,thereby improving the tracking accuracy.2.A multi-object tracking algorithm based on cross-frame correlationAiming at the problem of identity switches caused by the long-term disappearance of the objects in the field of vision,this thesis proposes a multi-object tracking algorithm based on cross-frame correlation.A pyramid network extracts multi-scale information for target detection,a similarity prediction network predicts the similarity between the targets in different frames,and a cross self-attention network strengthens the connection between the trajectories and the objects,and data association is performed in units of frames.When an object disappears,the cross-frame correlation strategy helps to regain the object identity,improving the stability of multi-object tracking.Through a series of experiments,it is verified that the multi-object tracking algorithm based on cross-frame correlation proposed in this thesis can effectively alleviate the problem of identity changes caused by long-term disappearance of objects.3.A multi-object tracking systemOn the basis of the multi-object tracking algorithm based on attention and the multi-object tracking algorithm based on cross-frame correlation,this thesis designs and builds a multi-object tracking system,which realizes the functions of user registration and login,model configuration and deployment,tracking show and historical performance comparison.In the interface of each user,it is possible to choose to call the camera,use local video or use datasets to track multiple objects in complex scenes,and output the tracking information and visualization results of multiple tracked objects;and the user can view and compare his history tracking results.The system interacts with users through a graphical interface,which reflects the application value of multi-object tracking.To sum up,this thesis conducts in-depth research on feature extraction and motion prediction,similarity calculation and data association,and this thesis proposes an anti-occlusion multi-object tracking algorithm based on attention and a multi-object tracking algorithm based on cross-frame correlation,improving the ability to track interactively occluded targets and long-term disappearing targets.Based on the above research,this thesis finally builds a multi-object tracking system.