Research On Visual Object Tracking Algorithm Based On Machine Learning

Visual object tracking is an indispensable part of computer vision system,it has a strong practical value and broad prospect in intelligent video surveillance,medical treatment,human-computer interaction and traffic management.Although many new algorithms are emerging every year,while the reality is complicated.The object is often interfered by factors such as occlusion,illumination change,deformation and fast motion.Therefore,solving these problems has become the main task of visual object tracking to researchers.Based on the analysis of the current research status of visual object tracking,two efficient visual tracking algorithms are proposed,namely compressed sensing object tracking algorithm based on weighted multi-instance learning and visual object tracking algorithm based on deep reinforcement learning.Compressed sensing technology has been paid many attention because of its low cost storage space and high efficiency of data transmission.The compressed sensing object tracking based on the weighted multi-instance learning is proposed in this paper.Firstly,Haar-like features of the input image are extracted.Then design a sparse matrix which randomly project the input image features.The features after the projection can not only fully characterize the image information,but also speed up the tracking speed.Then,sampling multiple positive and negative samples via weighted multi-instance learning method,and their Haar-like features is extracted to train the naive Bayes classifier to distinguish foreground and background,finally realize tracking.A vision tracking algorithm based on deep reinforcement learning is proposed in this paper,image features are extracted by VGG-M,at the last layer of the network,a policy function is learned according to the information that the agent and environment interacts.The agent performs any of the seven actions according to this policy function until the future total discount reward function value no longer increases,then stops the execution of the action and obtains the tracking result of the current image,repeating this until the last frame of the image sequence.Finally,the effectiveness of the algorithm was verified on the OTB-50 database,this database contains 50 image sequences,which can well verify the performance of an algorithm,then compare it with the current popular algorithms on this database.The quantitative and qualitative experimental results show that both algorithms can effectively deal with issues such as occlusion,illumination changes,and rotational motion,it also has strong anti-interference and has a good robustness.