Deep Convolutional Features For Correlation Filters Visual Tracking

Computer vision tracking technology has been widely used in the fields of intelligent transportation,monitoring,human-computer interaction,etc.With the continuous improvement of computer performance,artificial intelligence and other technologies,visual tracking has developed rapidly.However,when the existing tracking algorithms encounter complex environment and changes of the object's own apparent model,there is a problem that the high-precision tracking algorithms are not easy to meet the real-time requirement,and the accuracy of the fast tracking algorithm is not high.Therefore,this thesis combines deep learning and correlation filters visual tracking framework to ensure real-time tracking as the premise,and the tracking accuracy is improved as much as possible.The tracking algorithms are improved from deep convolutional features selection,correlation filters training,object position prediction and model update.And four visual tracking algorithms are proposed.The main contributions are as follows:(1)An adaptive convolution feature selection tracking algorithm is designed to reduce the deep convolution feature redundancy and improve the tracking speed and accuracy.The convolutional channels are selected by using the feature mean ratio of the target area and the search area.And then the tracking is performed by using the effective convolution feature.The designed algorithm was tested on the OTB-100 dataset.The experimental results show that the average distance accuracy is 86.4%and average tracking speed is 29.9 frames per second,meeting the requirement of-real-time tracking.(2)In order to take full advantage of the different convolutional layer features,an adaptive convolutional features switching algorithm is designed.This method uses the adaptive convolutional feature selection algorithm to select the convolutional channels from two mid and high convolutional layers.Then,the suitable single-layer effective convolutional feature is selected by the peak-to-sidelobe ratio of each frame to track adaptively,which improves the accuracy of tracking under the premise of real-time tracking.The designed algorithm was tested on the OTB-100 dataset.The experimental results show that the average distance accuracy is 89.3%,and the average speed is 25.8 frames per second.This method meets the requirements of-real-time tracking,and the tracking accuracy is improved.(3)According to the characteristics of each channel convolutional feature,a tracking method with channel reliability and fine location is proposed.This algorithm selects the single-layer convolution feature suitable for object tracking,and then uses the adaptive convolution feature selection algorithm to select the effective convolution feature,which reduces the feature dimension and improves the tracking speed.In order to avoid the influence of convolution feature layer and channel number reduction on tracking accuracy,a channel weighted correlation filters algorithm is constructed to improve object tracking accuracy.Then the peak-to-sidelobe ratio is used to estimate the accuracy of object location.And a fine location method via minimizing mean frame difference is designed to reduce the prediction position error.The algorithm was tested on the OTB-100 dataset.The experimental results show that the average distance accuracy is 91.3%and tracking speed is 31.9 frames per second.The accuracy and speed of the tracking were further improved.(4)In order to improve the overall robustness of the tracking algorithm,a continuous convolution tracking algorithm based on channel pruning is proposed.The method uses the channel pruned convolutional model to improve the calculation speed of the convolutional features.Then,the channels are pruned further to reduce the dimensions of the selected features.The effective convolution features,histogram of oriented gradient and color name features are fused to track on the continuous Cconvolution operators framework.And an adaptive iterative strategy is designed to reduce the computational complexity and improve the real-time performance.This algorithm is tested on the OTB-2013,OTB-2015,VOT-2016 and VOT-2017.The average distance accuracy is 91.3%on OTB-100.The accuracy and robustness of the tracking are further improved.