基于改进卷积神经网络的在线视觉目标跟踪方法

Online Visual Target Tracking Method Based on Improved Convolution Neural Network

针对基于神经网络的在线目标跟踪方法容易丢失历史信息的问题,提出了一种基于改进卷积神经网络的在线视觉目标跟踪方法.首先利用卷积神经网络来表示物体的外观,基于树结构用多个卷积神经网络协作来估计目标物体的状态,并且为在线模型的更新设定理想路径.其次,通过维护树结构中不同分支的多个卷积神经网络来有效处理物体外观的多模态,并通过树路径的平滑更新来维持模型的可靠性,同时在卷积层共享所有的参数,从而以最小的额外代价充分发挥多模型的优势.最后在多模型状态附近进行抽样,依据加权平均确定最佳样本估计值.仿真实验表明,与现有算法相比,该算法具有更好的在线跟踪性能.

Aiming at the problem that the online target tracking method based on neural network is easy to lose the historical information,an improved online visual target tracking method for convolution neural network is proposed.Firstly,the convolution neural network is used to represent the appearance of the object.In the tree structure,the convolution neural network is used to estimate the state of the target object,and the ideal path is set for the updating of the online model.Secondly,the multimodal of the appearance of the object is effectively processed by maintaining the reliability of the model by smooth updating of the tree path,and all the parameters are shared in the convolution layer by maintaining aplurality of convoluted neural networks of different branches in the tree structure.With the smallest additional cost to give full play to the advantages of multi-model.Finally,a sample is sampled near the multi-model state to determine the best sample estimate based on the weighted average.Simulation results show that compared with the existing algorithms,the proposed algorithm has better on-line tracking performance.