Context-aware Base On Correlation Filter For The Research And Design Of Object Tracking

Object tracking is the task of detecting the same target in subsequent frames on the premise that the target is selected in the initial frame of the video.In recent years,many object tracking methods using deep learning have achieved good results.However,the high-dimensional deep features and frequent convolution and pooling operations have caused tracking delays.During the tracking process,the appearance of the target changes and the complete occlusion causes a model drift.In order to solve the above problems,this thesis optimizes the deep features based on the correlation filtering tracking algorithm;and uses the context information and time information to improve the context-aware correlation filter model to improve the accuracy of object tracking.The main research contents of this thesis are as follows:(1)Context information plays an important role in the construction of correlation filters.Many tracking algorithms use cosine windows to mitigate boundary effects,but also weaken the information contained in the context area.In response to this problem,a weighted strategy is used to highlight the contribution of different context areas to object tracking,and a time regularization term is added to ensure the continuity of filter parameter updates in time and prevent suddenly changes in model updates,making the response effect more robust.Compared with the traditional correlation filtering tracking algorithm,the tracking accuracy of the algorithm has been greatly improved,and it can be effectively applied to complex scenes such as target scale variations and rapid rotation.(2)The deep feature happens redundancy in object tracking,causing tracking delay.In response to this problem,a feature compression method using pre-trained auto-encoders is proposed.First the ImageNet dataset be used to train the base auto-encoder.Since the feature spaces of different types of targets are different,multiple feature compressors are then trained on the base auto-encoder.Finally,based on the specific compressor,the network is selected to match the appropriate automatic encoder to compress the features.This algorithm not only maintains the tracking accuracy,but also significantly improves the tracking speed compared to other algorithms that use deep features.Finally,a large number of experiments were conducted on the representative OTB dataset.The experimental results show that the filter model proposed in this thesis can effectively adapt to complex scenes such as scale variations and occlusion,and significantly improve the tracking performance after optimizing the feature space.