Visual Tracking Algorithm Based On Deep Learning And Semi-supervised Learning

Aiming to locate an object in a video,target tracking has gained widely attention of many researchers in recent years.In this thesis,the difference between the classical feature description algorithm and the deep learning for feature extraction are compared.The convolutional neural network is introduced into the classical particle filter algorithm,and the process of creating and updating the target model is proposed.In addition,an improved Particle filter based on a partial resampling algorithm with adaptive threshold is proposed.In a set of experiments the new algorithm is compared with other classical tracking algorithms to validate the efficiency of using deep learning to model the target.Furthermore,an improved tracking algorithm based on the semi-supervised learning and deep learning is proposed,in which the structure and training process of the convolution neural network are improved.For the long time of CNN's off-line training,ReLU is used as the activation function of the convolution layers,which greatly reduces the training time.At the same time,the semi-supervised learning is introduced into the online training process of CNN.The influences caused by the wrong samples on the model parameters are reduced by using the semi-supervised theory to update the training set.In offline training of CNN,we expands the training set by using the data augmentation,and the over-fitting problem is avoided effectively.