Research On Visual Object Tracking Algorithm Based On Deep Learning

Visual tracking is an important research topic in computer vision,and has a wide range of applications.In recent years,deep learning has achieved remarkable results in the field of computer vision.The development of deep learning not only breaks through many problems that are difficult to solve with traditional algorithms,improves the cognitive level of computer for images and videos,but also promotes the progress of related technologies in the field of computer vision.The tracking algorithm based on deep learning has also made great progress and achieved good tracking performance.However,many problems still restrict the further improvement of tracking performance,such as inaccuracy target location determined by the response map,poor discriminating ability,inaccurate initialization information,etc.This thesis makes in-depth research on these problems in single object tracking,and proposes three visual tracking algorithms based on deep learning.The main work and innovation of this paper are as follows:(1)Most of the early correlation filter-based object tracking algorithms only use hand-crafted descriptors to extract features,which limites the feature representation of target,so the location determined by the response map is not accurate,especially under the interference of occlusion and background clutters,and the tracking performance is often unsatisfactory.In this paper,a correlation filtering tracking algorithm based on salient region weighting is proposed.The improved residual network SE-Res Net is used to extract the multi-resolution features,and the visual saliency is introduced into the correlation filter,used to weight the response map to improve the accuracy of target location.This algorithm is evaluated on visual object tracking(VOT)challenge,the expected average overlap(EAO)scores of this algorithm on VOT2016 and VOT2017 reach 0.4157 and 0.3412.Experimental results show that the proposed tracking algorithm has good tracking performance.(2)Most of the object tracking algorithms based on Siamese network only learn the similarity measurement model via off-line training,the exemplar branch has insufficient discriminant ability to adapt to the constantly changing appearance of the target in subsequent frames.To solve this problem,a Siamese network object tracking algorithm based on adaptive background superposition initialization is proposed.Firstly,an adaptive background superimposed initialization strategy is proposed and used in exemplar branch of Siamese networks to enhance the discriminant ability of exemplar branch.Secondly,a 13-layer convolutional neural network is proposed and applied as the backbone of our Siamese network.Thirdly,a channel attention mechanism is introduced to automatically re-weight feature maps in order to obtain more effective features for tracking tasks.This algorithm uses the GOT-10 k database for training.This algorithm is evaluated on object tracking benchmark(OTB)and VOT challenge,the area under the curve(AUC)of success plot of this algorithm reaches 64.5% on OTB-100,and the EAO scores on VOT2016 and VOT2017 reach 0.3011 and 0.2397.Meanwhile,the proposed method has a fast running speed,which shows the excellent tracking performance of the algorithm.(3)To solve the problem of inaccurate initial boundary frame provided by detection algorithm in pedestrian tracking system,this paper proposes a pedestrian tracking algorithm based on mask superposition initialization.Firstly,Mask R-CNN is used to acquire the exemplar image and the mask image of the target,and the lightweight convolution neural network is used to extract features of them.Then,channel attention is used to adjust the extracted features,and the adjusted features are fused proportionally to enhance the discriminant ability of exemplar branch.Finally,the algorithm is evaluated on the pedestrian tracking sequences in OTB dataset,the area AUC of success plot of this algorithm reaches 62.2%.The experimental results show that the algorithm can alleviate the problem of tracking performance degradation caused by inaccurate initial bounding box and has excellent tracking performance.