Object Tracking Algorithm Based On CNN And RNN Structured Processing

Object tracking is one of the important research contents in the field of computer vision.It is widely used in national defense military,security system and daily life.However,complex situations such as motion blur,scale variation,and complex background often appear in actual scenes,making the performance of the tracking algorithm difficult to guarantee.The object tracking algorithm based on deep learning uses convolutional neural network(CNN)to extract features from images.These features have stronger generalization ability,which can effectively improve the accuracy of the algorithm.However,the similar objects cannot be effectively distinguished by CNN-based tracker.To further improve the accuracy and real-time performance of the tracker,the structured processing network(SPNet)combining convolutional neural network with recurrent neural network(RNN)for object tracking is proposed in this paper.The details are as follows:(1)The method of offline pre-training and online fine-tune be used to extract the feature and track the object.Aiming at the problem that the current CNN-based tracker has poor tracking accuracy for similar objects,a method of structural processing using RNN is proposed.CNN is used for inter-class discrimination to provide discriminability between the object and the background.The object be modeled by RNN to distinguish similar objects.Since the traditional chain-structured RNN is not suitable for images,in this paper,the directed acyclic graph RNN are proposed to process images,which enables the network to model long-range semantic dependencies among image units.Thus,the discrimination between the object object and the similar object can be provided,and the accuracy of the algorithm can be improved.(2)RolAlign be introduced by SPNet to accelerate feature extraction while ensuring feature quality,reduce dimension of data.The essence of this layer is a pooling of candidate regions.At the same time,RolAlign layer prevent over-fitting and improve real-time performance of the algorithm.(3)Hard minibatch mining is used in SPNet to improve the accuracy of the algorithm.Similar loss is introduced to measure the difference between similar objects.The tracking results are fine-tuned by boundary box regression,which further improves the accuracy of the algorithm.Finally,the long-term and short-term complementary model updating strategy is used to further improve the performance of the algorithm.Finally,we design experiments to compare the proposed SPNet with current state-of-the-art tracker.The experimental results show that SPNet still has good tracking effect in complex situations such as illumination change,object deformation and occlusion,and has certain advantages in accuracy and real-time.