Research On Application Of Deep Learning And Deep Reinforcement Learning In Video Object Tracking

In video object tracking technology,accuracy and speed are two important indicators to evaluate the performance of tracking algorithms in complex environments.In recent years,compared with the traditional object tracking algorithm,the deep learning model has made a breakthrough contribution to the improvement of tracking accuracy due to its powerful feature extraction ability,but it has slowly speed.Recently,some scholars have applied deep reinforcement learning to video object tracking technology,the speed is improved compared to the deep learning object tracking algorithm,but its accuracy is slightly worse than the deep learning.The problem that tracking speed and tracking accuracy are mutually constrained is the main challenge faced by video object tracking algorithms.Thus,the application of deep learning and deep reinforcement learning in video object tracking are comparative study in this paper.The works of this paper is as follows:(1)The system comprehensively reviews the current research status of deep learning and deep reinforcement learning in video object tracking technology and points out the advantages and disadvantages of various methods.Firstly,aiming at the problem that the accuracy and speed exist condition each other in the online tracking process for deep learning object tracking algorithm relying on offline training,according to the different solutions,it is divided into deep learning tracking with online fine tuning and no online fine tuning,and the advantages and disadvantages of various methods are analyzed.Secondly,aiming at the existing deep reinforcement learning object tracking algorithm,according to the different functions of reinforcement learning,it can be divided into two categories,namely,decision-making object location using reinforcement learning and directly predicting object location using reinforcement learning.At the same time,summarized and analyzed them.(2)Aiming at the problem that the deep learning tracker using non-tracking dataset in offline training is insufficient tracking accuracy in the complex environment,the object tracking that directly using the tracking dataset to offline train the convolutional neural network is studied and implemented.Firstly,a multi-branch convolutional neural network is constructed,which including general feature layers and a specific feature layer.Secondly,it is offline trained using the idea of multi-domain learning,the main purpose is to use general feature layers to extract general feature representation of objects in all training video sequences.Finally,the generic obj ect feature representation extracted from the generic feature layer is used in the online tracking.The experiment that compares the video obj ect tracking algorithm based on convolutional neural network(MBCNN)with four mainstream tracking algorithms(DLT,TLD,Struck,CXT);quantitati've experimental results that based on the overall performance and attribute-based on the OTB50 dataset and the qualitative experimental results on the four video sequences existing in the real scene show that the deep learning video object tracking algorithm using the tracking dataset for offline training has higher tracking accuracy in complex environments.(3)Aiming at the problem that the tracking speed and tracking accuracy are mutually constrained in the object tracking technology,the object tracking based on deep reinforcement learning is studied and implemented.Firstly,the supervised learning and reinforcement learning are used to train convolutional neural networks,it makes the network realize the mapping from a given state to an action.In online tracking,the pre-trained convolutional neural network predicts the appropriate action according to the current object state,thereby moving the position of the rectangular representing the object state until the stop action is predicted,and determining the position of the object in the current frame.The experiment that compares the object tracking algorithm based on deep reinforcement learning(ADCNN)with four mainstream tracking algorithms(SCM,TLD,Struck,CXT),the quantitative experimental results that based on overall performance and based attribute performance on the OTB 50 dataset show that the deep reinforcement learning object tracking algorithm can achieve a good balance between tracking speed and tracking accuracy.At the same time,the qualitative experimental results on five video sequences with multiple challenging attributes show that the deep reinforcement learning object tracking algorithm can achieve robust tracking in most challenging attributes.(4)The applicability analysis of attribute-based tracking algorithm.For the 11 challenging attributes in object tracking,considering that there is no significant difference in the tracking accuracy of different algorithms under certain attributes,but the tracking speed is significantly different.Taking the deep learning tracking algorithm MBCNN and deep reinforcement learning tracking algor:ithm ADCNN as the research objects.Firstly,the experiment and comparative analysis of the applicability of the online fine-tuning and no online fine-tuning based on 11 attributes for MBCNN in the actual situations where the accuracy and speed have different requirements.Secondly,the applicability of MBCNN and ADCNN based on 11 attributes is compared and analyzed by experiments in actual situations with different demands for accuracy and speed.It provides a theoretical basis for users to choose reasonable tracking algorithms on sequences with different attributes according to different demands of accuracy and speed.