Research On Object Tracking Algorithms Based On Deep Learning

Object tracking is an important task of computer vision.It is a continuous inferential process of target state.Many people pay attention to tracking field because of its wide application.Nowadays,it is widely used in road monitoring,target recognition and classification,military detection and other fields.In recent years,people have put forward lots of new tracking algorithms which is expected to reach a balance between real-time and accuracy.However,these new tracking algorithms still confront many challenges,such as the sudden movement of the object,illumination change,occlusion,deformation,pose changes and so on.These factors may lead to target drift or even the failure of tracking.Therefore,how to make tracking more robust is still a challenging topic.Deep learning is a new technology in machine learning algorithm.Its motivation is to create a neural network to simulate human brain for analysis and learning.In recent years,it has developed rapidly and is widely used in the fields of detection,identification,tracking,etc.Based on the existing tracking theory and results,we have conducted in-depth research on traditional tracking algorithms based on deep learning and proposed some improved algorithms for their existing problems.The main research work and innovations of this paper are as follows:(1)This paper first gives the background of object tracking and deep learning.It focuses on the basic theory of object tracking.Then we improved some of the problems existing in the tracking algorithm in recent years.In the experimental part,the video sequence on the data set and the self-acquired video sequence are selected as experimental objects,and the algorithm is compared with the mainstream algorithm.(2)A sparse object tracking based on deep learning and support vector machine algorithm is proposed.Firstly,the algorithm constructs a deep network based on the auto-encoder and then adding the sparsity constraint in the deep network to sparse connection matrix between the hidden layer and the output layer.As a result,the algorithm optimizes the parameters of the deep network and improve its efficiency.It means that more essential features of the target will be extracted with this network.In the prediction of the target,the algorithm introduces the difference between target and background into the particle filter and design a scoring device based on support vector machine,so that particle performance is enhanced and the risk of drift in the process of tracking target is reduced.Experiments on different video sequences have been earied out for many times.According to the results of experiments,we can come to a conclusion that our algorithm has higher accuracy and better robustness,especially under the circumstance of occlusion,illumination change and similar background.(3)An object tracking algorithm based on K-sparse depth network and improved particle filter is proposed.By adding K-sparse constraint to the stacked denoising auto-encoder,the K neurons with the highest activity in the hidden layer are activated,and a K-sparse depth network is constructed.The features extracted by this network are easier to classify.At the same time,the high-weight particles and the low-weight particles are linearly combined in the resampling process,and the resulting particles are completely new,and the particle diversity is improved.The test results show that the algorithm can accurately track the target under the circumstance of illumination change,occlusion and deformation.(4)An object tracking algorithm based on multi-strategy search is proposed.The algorithm firstly generates target candidate regions by combining selective search and particle filter.In the feature extraction stage,the feature is collaboratively extracted by the convolutional neural network and the stacked denoising auto-encoder.Finally,the SVM is combined to generate the particle confidence and the particle with the highest confidence is used as the tracking result.Experiments show that the proposed algorithm has better robustness and accuracy than other algorithms under the circumstance of scale variation,partial occlusion,out of view and rotation.