Research On Object Tracking Algorithm Based On Correlation Filter

In recent years,object tracking has become one of the research hotspots in the field of computer vision.This paper summarizes the main problems still existing in tracking algorithms.One is that there is no one can applicable to all scenarios and tracking drift or failure often occurs when faced with complex environments.Another is that the increasingly complex tracking model causes problems such as too many training parameters,over-fitting or greatly reducing the tracking speed,although bringing about the improvement of precision.Therefore,for further improving the robustness of the algorithm under different challenge factors,the spatial regularization term in the original algorithm is adjusted adaptively to make the algorithm stable under background clutter and deformation.However,the use of convolutional networks makes the speed of the algorithm still too low.Thus,an acceleration model is proposed in the following chapters,which effectively improves the processing efficiency of the algorithm.At the same time,the dual tracker is used to ensure the accuracy of the tracking algorithm.The research results and main contributions of this paper are as follows: Firstly,an object algorithm based on correlation filter and adaptive spatial regularization term is proposed.The spatial regularization weight function used by the mainstream DCF tracking algorithm cannot make adaptive adjustment according to the video characteristics.The weight value depends only on the Euclidean distance between the pixel and the target position,and thus cannot effectively distinguish which pixel is more likely to be a target.Thus,the confidence map is introduced into the weight function.The improved weight is used to update the weight value after the end of each frame prediction,so that the penalty function can be adaptively updated according to the current target feature,and the effect is that the more similar to the target in the non-target area,the smaller the penalty weight value,and vice versa.In addition,the size of the non-target area that is desired to be suppressed can be controlled by the set weight threshold,and an appropriate threshold can make the algorithm effectively cope with the background interference.By verifying the algorithm on the object tracking standard dataset,the results show that the improved algorithm has a 4.2% improvement in target deformation,which can effectively deal with background interference and non-rigid target deformation scenarios.Secondly,a robust object tracking algorithm based on Extreme Learning Machine and correlation filter is proposed.Based on the C-COT algorithm,the algorithm improves the feature extraction method and the method of optimizing the confidence map.It mainly solves the problems brought by the multi-channel convolution features,including too many filter parameters,slow training process and other issues.Firstly,a new feature extraction model is constructed by using the double-layer ELM-based autoencoder to replace the original convolutional neural network.Since the ELM-based autoencoder itself has some characteristics,so the model can not only quickly extract image features,but also ensure the representation of features.Secondly,after the feature extraction model,an online sequence extreme learning machine is added to construct a target rough position estimation model.The model is mainly used as a tracker to initially predict the target position,and outputs a Gaussian Map representing the target probability.The possible locations of the target are determined by finding the extreme points of Gaussian Map in continuous domain.Thirdly,the algorithm uses the framework of the C-COT algorithm as another tracker,except that the multi-channel features used here are composed of features extracted from the new feature extraction model,HOG features and raw image data.And the algorithm predicts the final target position by a proposed multi-tracker fusion mechanism instead of only by the confidence map in order to avoid the tracking result from falling into the local optimum value.Finally,the effectiveness of the new algorithm is verified on the target tracking standard datasets.The experimental results show that the tracking speed of the new algorithm is about 12 times that of the C-COT algorithm,and it is robust to occlusion,motion blur and similar targets,which can effectively improve tracking accuracy and speed.