Research On Deep Learning And Its Applications In Video Object Tracking

Recently,deep learning has made a great breakthrough in the field of computer vision and there are growing interests on the use of deep learning for computer vision tasks.At present,convolutional neural network(CNN),which is a key neural network structure on deep learning,still faces some problems such as poor generalization ability and high computational complexity.In this paper,we focus on deep learning and its applications in object tracking.The main contributions of this paper can be summarized as follows.(1)Based on the state-of-art development of deep learning and its applications in object tracking,various problems on CNNs and object tracking are analyzed.(2)The design of the pooling layer in deep neural networks is studied.The pooling layer is a standard component for constructing deep CNNs,where stochastic pooling has shown its superiority in preventing overfitting for training CNNs.With the use of the negative-response activation functions,the original stochastic pooling cannot work.It is necessary to generalize stochastic pooling for negative-response activations to improve the generalization ability of CNNs.In order to solve this problem,the generalized stochastic pooling is proposed,where the shifting approach works best.The proposed generalized stochastic pooling is not strictly equivalent to the original form for non-negative-response activations,such as ReLU,but it does not result into evidently performance degradation.Experimental results show that the proposed generalized stochastic pooling can be well employed for the design of CNNs with negative-response activations.(3)The design of the stochastic pooling in CNNs for the purpose of object tracking is studied.CNN is an important feature extractor for the task of object tracking.Due to the requirement on the performance and computational complexity in practical tasks,it is necessary to maintain the high accuracy with lower computational complexity for CNNs.In order to improve the generalization ability and decrease the computational complexity for the siamese networks,a novel object tracking network is designed,which is based on the use of generalized stochastic pooling in Tiny Darknet.(4)In order to test fully-convolutional siamese networks model based on Tiny Darknet,we present a detailed testing procedure.Experimental results on the OTB-13 benchmark show that our object tracker can obtain slightly lower tracking accuracy with significantly-reduced complexity.Actually,the proposed method can achieve 30% increases in frame-rate compared with the original AlexNet fully-convolutional siamese networks.