Video Object Tracking Based On Deep Siamese Network And Multi-scale Feature Fusion

Object tracking technology is widely used and has a close relationship with the lives of modern people.The visual object tracking is that for any video sequence,the bounding box of the object is only given in the first frame of the sequence,and researchers need to design an algorithm to accurately predict the position of the object in the remaining frames of the sequence.Researchers have been studying object tracking for more than ten years.From the initial generative tracking algorithms,such as traditional object tracking algorithms based on optical flow,particle filtering,Mean Shift,etc.,to discriminative tracking algorithms in recent years,such as algorithms based on machine learning,correlation filtering,and deep learning.Tracking algorithms continue to develop and progress,but occlusion,real-time tracking,and accurate tracking in complex scenes still pose significant challenges to current object tracking tasks.At present,algorithms based on correlation filter and deep learning are hot research topics.In the past few years,the siamese network has become a research hotspot in the field of object tracking due to its simple and effective network structure.Its representative algorithms such as Siam FC,Siam RPN and ATOM are excellent in speed and accuracy.However,the feature extraction networks of these algorithms are relatively shallow networks,such as Alex Net,which do not exert the advantages of deep neural networks.In addition,due to the downsample and other operations of the siamese network during the feature extraction stage,the size of the small object in the feature map is very small,resulting in poor tracking of the small object.In response to the above-mentioned problems of the siamese tracker,this article mainly made the following three improvements.First,a deep feature fusion network Siam DFF based on Res Net is proposed to improve the feature expression ability of the network.Since the padding operation will affect the accuracy of object localization,after each convolutional layer with a padding operation,a crop layer is added to remove the features affected by the padding operation.On this basis,by integrating depth features of different scales,the accuracy and robustness of the tracking are improved.Secondly,an efficient algorithm Siam DCAN based on deep features and channel attention mechanism is proposed.By adopting the channel attention mechanism in the deep neural network,Siam DCAN can capture the interconnection between the channels and selectively emphasize the channel feature maps that are more conducive to distinguishing the object.Finally,for the problem that small objects are difficult to track accurately,we propose an object tracking algorithm Siam FPN,which is based on deep siamese network and feature pyramid network.The feature pyramid network is used to extract and merge multi-scale features,which retains the feature details of small objects to the greatest extent,and effectively improves the tracking accuracy of small objects.The three algorithms were trained on the VID video dataset and tested on the OTB2013 dataset and OTB2015 dataset,and compared with many tracking algorithms based on deep learning and correlation filtering.Experimental results show that our proposed algorithms have greatly improved tracking accuracy,which proves the effectiveness of the above object tracking algorithms based on deep siamese network and multi-scale feature fusion.