Research On Object Tracking Algorithm Based On Deep Learning

In this thesis,in order to solve problem that the context-aware correlation filtering algorithm(CACF)based on artificial features has poor tracking performance under complex environments such as out of view,motion blur,and illumination variation,a convolution regression network visual target tracking algorithm was proposed.And a re-detection mechanism is introduced on this basis to solve the problem of tracking failure caused by accumulation of inaccurate model predictions.Finally,we use multiple convolutional layer features for tracking to improve the accuracy of the algorithm.The main work of this paper is as follows:1)A convolution regression network visual target tracking algorithm(CRN)was proposed to effectively deal with the problem that poor tracking performance of traditional artificial feature algorithms under complex environments such as illumination changes and low resolution.In the traditional correlation filter algorithm,the parameters of the filter were set manually.In this thesis,the ridge regression as a layer in the convolutional neural network.By designing its back propagation parameters,a convolutional regression network is finally formed.The whole network was trained end to end on the data set.The convolutional neural network was designed for the object tracking task,and the feature expression ability is stronger.2)Based on the Convolution Regression Network Algorithm(CRN),a re-detection mechanism was introduced.In order to solve the problem that model drift in the case of severe occlusion,a re-detection mechanism is established.A threshold is set by the response value during the tracking process to supervise whether to perform re-detection.When the response value is less than a given threshold,re-detection is activated.In the re-detection stage,a support vector machine classifier is trained to distinguish the target from the background.To avoid falling into local optimum,the re-detection model will expand the search range when detecting.3)We propose an algorithm which use multi-convolution layer feature to track(MCT).The deep layer of the convolutional neural network contains a lot a semantic information,while the shallow layer retains the finer-grained spatial details which are very effective for accurate positioning.Therefore,different network layer features are extracted and merged for tracking.At the same time,a scale adaptive model is introduced for the target scale change,and a one-dimensional scale filter to select appropriate target scale,finally we propose a response value peak distribution function and establish a noise threshold to control whether the model is updated,and ensure that no meaningless update is performed.