Infrared Single Object Tracking Algorithm In Complicated Scenarios

Single object tracking is an important research direction in the field of computer vision.The purpose of single object tracking is to model the appearance information of the object and continuously predict the state of the object in successive frames by using the position of the object in the first frame of video.Because infrared thermography technology can capture the information of objects without other light source,infrared object tracking has become one of the hot issues in the field of tracking.However,white noise,low signal to noise ratio and image blur often exist in the infrared image.In practical application.There are complicated tracking scenarios,such as object occlusion,similar object interference and small object,which affect the accuracy of infrared object tracking.With the wide application of deep learning technology in the field of object tracking,there are many object tracking algorithms with good performance.In all these algorithms,the learning discriminative model prediction for tracking algorithm(Di MP)can learn the observation model for a specific object,which has strong ability to distinguish the object from the background.However,because of the fixed size receptive field in the classification branch,the algorithm is not sensitive to the scale change of the object.Because of the inherent image blur,white noise and other problems of infrared image,it can not effectively represent the object when occlusion or similar object interference occurs.Meanwhile,in order to learn the change of the object,the mainstream method updates the tracker online by using the tracking results as training samples.However,when occlusion or similar object interference occurs,it often leads to catastrophic update,causing tracking failure.In order to solve the above problems,the improvement methods are as follows:(1)To solve the problem that DiMP algorithm is not sensitive to the scale change of infrared object,an infrared object tracking algorithm based on receptive field pyramid model is proposed.The receptive field pyramid model contains multiple branches,each branch uses convolution kernel of different sizes to extract features.The network learns features of different scales and aggregates them.Meanwhile,in order to reduce the number of parameters,the convolution kernel of large size is splitted by using the convolution splitting idea,it improves the sensitivity of object tracking to scale changes and alleviates the tracking failure caused by the fixed receptive field.(2)To solve the problem of object occlusion and similar object interference in infrared image tracking,a robust infrared object tracking algorithm based on lightweight channel attention mechanism is proposed.By designing a lightweight channel attention mechanism to avoid the introduction of extra computation,one-dimensional convolution with fixed convolution kernel size is used to calculate the weight of each channel and weight it,so that the object features can be better extracted in the tracking process.Secondly,combined with the average peak correlation energy evaluation index,the oscillation degree of the response graph is used to judge whether the prediction result contains much noise.When there is too much noise,the current frame is discarded to update,so the robustness of the tracker is improved.(3)Based on the single object tracking algorithm proposed in this thesis,the infrared single object tracking software based on pyqt is designed and developed.After deployment in real scenarios,the software is applied to a national key research and development program.The software can save the intermediate results of the tracking process,facilitate users to analyze the tracking process,and effectively realize the application of infrared single object tracking.