| In recent years,visual object tracking has been widely studied and applied in auto-driving,specific object tracking and positioning,human-computer intelligent interaction,and so on.It has become one of the hot issues in the field of computer vision.The deep correlation filter tracking algorithm has become one of the hot research problems in the field of visual object tracking due to its combination of the deep convolutional neural network with powerful feature representation and the computationally efficient correlation filter tracking algorithm.As a feature extractor of deep correlation filter tracking algorithm,deep convolutional neural network can extract the deep information of object by using strong feature characterization ability.Accurate object features can help the downstream tracker to track and locate target better,and thus improve the tracking accuracy.However,deep convolutional neural network models are often complex in structure and not only have high demands on computational resources,but also have high computational complexity,which inevitably restricts correlation filter from taking full advantage of its speed and making it difficult to meet the requirements of target tracking in real time.In addition,when the actual tracking scenes and targets themselves change a lot compared with the training dataset,the robustness of the pruned feature extraction network trained by off-line pre-training has some limitations.In this paper,based on the above problems,the correlation filter tracking algorithm based on the deep-pruned feature network is proposed.The main contents are as follows:(1)Firstly,a coarse-grained pruning algorithm for deep convolutional neural network based on channel importance,PCIP,is proposed.The algorithm uses the importance of the feature channel as a measure of the importance of the convolution kernel,and uses a continuous and smooth pruning strategy to prune the convolutional neural network from the perspective of global optimization.This can reduce the complexity of the network and computational burden while guaranteeing less loss of network characterization ability and system performance;(2)In order to reduce the size of the feature network,reduce the time required for feature extraction,and improve the tracking speed,the PCIP pruning algorithm is introduced into the feature extraction network of deep correlation filter tracking.The PCIP pruning method of deep correlation filter tracking is given.The initial lightening of the feature network is completed,and the deep correlation filter tracking algorithm with pruned feature network is obtained.(3)In order to solve the pruning rate adjustment problem,an optimization method for the global tracking pruning rate of deep correlation filter tracking is proposed based on the response contribution of the tracking response under the condition that the tracking accuracy is satisfied.(4)To further improve the real-time performance of the tracking algorithm in practical applications,an alternative convolutional kernel is defined to integrate multiple non-important convolutional kernels based on the global tracking pruning rate.Based on the alternative convolutional kernel,a specific method for secondary pruning of the feature network is presented.A lightweight deep correlation filter tracking algorithm based on alternative convolutional kernel is obtained by cascading the secondary pruned feature network and the correlation tracking filter.(5)An online update algorithm based on SSIM for pruned feature network is proposed for the tracking environment change and the target’s own state change.(6)Based on the OTB2013 dataset,the proposed PCIP pruning algorithm,the lightweight deep correlation filter tracking algorithm based on alternative kernel,and the S SIM-based pruned feature network online update algorithm are experimentally validated.The experimental results verify the validity of the proposed algorithm. |
