Research On Object Tracking Algorithm Based On Deep Learning

Object tracking is a research hotspot in the field of computer vision,and it is also an important part of visual applications such as automatic driving,video surveillance,and human-computer interaction.Due to the complexity of the object tracking task itself,and during the tracking process,faced with changes in the object itself and environmental changes,object tracking presents a huge challenge.This thesis conducts in-depth study on the mainstream deep learning object tracking algorithms in recent years,including deep feature object tracking and deep network object tracking algorithms.In order to improve the accuracy and speed of object tracking,the object tracking algorithm is optimized based on SiamFC object tracking algorithm(Fully-convolutional siamese networks for object tracking).1.Aiming at the problem of insufficient ability of AlexNet feature extraction in the basic network of SiamFC algorithm,an improved algorithm combining dual input shallow feature fusion and attention mechanism is proposed.First,this paper proposes a dual-input shallow feature fusion structure,which serves as the head end of the network.Then,the spatial attention mechanism module and the channel attention mechanism module are introduced.Finally,the network is reconstructed according to the feature fusion structure and attention mechanism.The SiamFC-ours improved algorithm uses the ILSVRC2015-VID data set to train the model and test it on the OTB2015 dataset.2.Aiming at the problem that tracking objects are easily lost when they are occluded and subjected to similar interference,an improved algorithm that optimizes training data and introduces a re-detection mechanism is proposed based on the SiamFC-ours algorithm.The algorithm is improved from two aspects of training data and tracking strategy:first,two high-quality object tracking datasets of GOT-10k and LaSOT are combined to increase the object category of training data,adding difficult negative sample pairs from the same category and different categories.Perform random data erasure,color dithering,blurring and other online data argument to improve the training data;then,give the confidence index of the response graph,enable the re-detection mechanism at low confidence,and update the template at high confidence to achieve tracking Strategy improvement.Ours improved algorithm uses optimized data to train the model,and uses an improved tracking strategy when testing on the OTB2015 dataset.