Research On Object Tracking Algorithm Based On Deep Learning And Particle Filter

Object tracking is a hot topic in the field of computer vision,which has a very important research value in intelligent monitoring,unmanned driving,man-machine interaction,national defense and military,etc.Its main task is to obtain the position and motion trajectory of the interested object in video or image sequences,and to provide basic information for further semantic layer analysis.According to the number of moving objects in the scene,object tracking can be divided into single object tracking and multi-object tracking.Compared with single object tracking,multi-object tracking is more complex and involves data association.Single object tracking can be simply understood as only filtering the continuous data of a single object in the video.It focuses on the design of complex appearance model or motion model to distinguish the object and background.With the rapid development of deep learning,object tracking based on deep learning has gradually become the research mainstream of scholars at home and abroad,and has achieved some good research results.However,due to the uncertainty of object motion and measurement value,object tracking is a more difficult state estimation problem.Under the framework of deep learning,the paper uses the idea of particle filter to investigate tracking-by-detection tracking algorithms,which are applied to single object and multi-object fields respectively.The main research contents and innovations of this paper are as follows:(1)The long short-term memory network(LSTM)model based on particle filter is constructed.Aiming at the problem that traditional LSTM is difficult to deal with high-dimensional random sequential data,a new long short-term memory network(PF-LSTM)is proposed in this paper.The network draws on the idea of particle filter,which uses a set of weighted particles to approximate the latent variables,and updates the latent state distribution through the gating mechanism of LSTM network according to the Bayesian rules.(2)The application of PF-LSTM in single object tracking is studied.Aiming at the problem that the existing object tracking algorithms cannot deal with the uncertainty of object motion well,based on PF-LSTM model,this paper proposes a depth particle filter tracker(DPFT),which can effectively model the uncertainty of video sequence.(3)The application of PF-LSTM in multi-object tracking is studied.Aiming at the problem of data association in multi-object tracking,based on DPFT model,this paper designs a model of data association combination problem.The model includes the object existence probability estimation model and the data association algorithm based on LSTM,the birth and death of object can be completely learned from the data,which realizes the end-to-end learning of online multi-object.(4)The effectiveness of the proposed algorithm is evaluated.The experimental results of DPFT algorithm on two benchmark datasets OTB100 and VOT2016 show that the accuracy rate of object tracking is 82.1%,and the success rate is 62.3%.The experimental results of multi-object tracking algorithm on MOT16 dataset show that the MOTA and MOTP of the algorithm reach 48.2% and 75.2%,respectively,and the performance is better than other state-of-the-art methods.