Object Tracking Algorithm Based On Recurrent Neural Network And Bayesian Filter

With the development of artificial intelligence technology,computer vision has been widely applied in the fields of human-computer interaction,monitor security,military deployment and so on.As one of the hot topics in computer vision,the task of object tracking is to locate a target in video sequences across time and identify its identity for building the target trajectory.Object tracking is a very challenging task in real world scenarios due to background disturbance,illumination change,target deformation,occlusion and so on.With the advent of big data era,deep learning method has been greatly developed in the fields of speech processing,image recognition,etc.,and it has also provided a new solution for the problem of target tracking.Under the framework of deep learning,this paper utilizes the idea of Bayesian filtering to investigate two tracking-bydetection tracking algorithms: online object tracking based on recurrent neural networks(RNNs),and online multi-object tracking using RNNs and multiple feature cues.(1)An online object tracking method based on RNN,which is suitable for single object tracking and multi-object tracking,is studied and implemented.Firstly,RNN is constructed to predict the possible target location in the next frame by current position information.The prediction is then updated using the latest observations,and the state of object is determined.For multi-object tracking,it is necessary to determine which object the observation belongs to before updating,and then update the object state with the matched observations.To this end,the Hungarian algorithm is adopted to solve the problem of data association.The cost matrix is constructed by measuring the distance between the object and the latest observation,and the object state is updated by the result of data association.Experiment results show that the proposed algorithm produces better results in both single object tracking and multi-object tracking,and is robustness for target deformation and occlusion.(2)In order to further improve tracking accuracy,we introduce the visual information of object based on the RNN tracking model.An online multi-object tracking algorithm based on RNN is proposed where position information and appearance features are combined.Specifically,we use the Siamese VGG network to extract appearance feature,and calculate the appearance similarity between the target and the detection.By combining this appearance similarity with position distance metric,more reliable matching is obtained to update the target state and improve the algorithm's tracking performance.The experimental results demonstrate that the proposed method can effectively overcome the influence of occlusion,target deformation and scale changes,and shows a better tracking performance in complex scenes.