Research On Pedestrian Visual Tracking Algorithms Based On Siamese Network

Object tracking algorithm,based on fully convolutional Siamese network,has been rapidly developed due to its good performance balancing capacity.However,in the actual tracking processing,it still faces many challenges,such as objects’ fast motion,deformation and occlusion.Wherein,pedestrian object tracking is an important branch of visual object tracking,which also has these difficulties in the process of tracking.Therefore,to solve the occlusion problem in pedestrian object tracking,This paper focuses on the Update Net network algorithm,proposes two improved schemes based on the existing object tracker used by fully convolutional Siamese network,and verifies them by common benchmark datasets.The specific research contents are as follows:(1)For pedestrian tracking task,in the first improved scheme,we use the Siam RPN algorithm as the benchmark tracker and improve it.On this basis,in order to solve the occlusion problem in the process of pedestrian tracking,this scheme integrates the Update Net network module and proposes a pedestrian visual tracking algorithm based on improved Siam RPN.Firstly,the uneven parameter distribution of Siam RPN algorithm is solved by applying the depth-wise cross-correlation module,while the hyperparameter values of the algorithm’s candidate proposal strategy are adjusted to optimize its performance;Secondly,in response to the drawback that fixed object features in the template frame during the Siam RPN algorithm tracking process,this scheme integrates the Update Net network module to update it online;Finally,the algorithm is verified on the common benchmarks,and a visual comparison is made using the pedestrian video sequences in the benchmark.Compared to the improved Siam RPN tracker,the results of experiment on the OTB2015 benchmark show that the proposed algorithm improves the Area Under Curve(AUC)score and distance precision relatively by 1.4% and 2.3%,and the two corresponding indicators on occlusion attribute are improved by 2.2% and 3.0%,respectively.At the same time,the tracking speed reaches 140 frames per second,which achieves the real-time requirements.(2)The second improved scheme uses the Siam Mask algorithm as the benchmark tracker for pedestrian tracking tasks.To address the issue of pedestrian occlusion,this scheme proposes a pedestrian visual tracking algorithm based on Siam Mask by integrating the Update Net network module.The Siam Mask algorithm performs object tracking and object segmentation tasks simultaneously through the Siamese network framework,and then relies on the predicted object segmentation mask to achieve more accurate object tracking and locating.This scheme integrates the Update Net network module into the Siamese network architecture,and improves the overall network’s anti-occlusion ability by updating the target features in the Siam Mask template frame online,thereby predicting a more accurate object segmentation mask.By calculating the minimum bounding rectangle of the prediction mask,the object can be tracked more precisely.The proposed algorithm was verified by the public benchmark dataset.The experimental results on the VOT2018 benchmark show that the Expected Average Overlap(EAO)of the proposed algorithm is 0.364,which is 2.3% higher than the benchmark algorithm Siam Mask.The robustness is improved by 5.6%,and it also fulfills the real-time requirements.In summary,the two improved algorithms proposed in this paper can effectively cope with the occlusion problem in the process of pedestrian tracking,better balance the tracking speed and tracking performance of the original benchmark algorithm,and provide two algorithm selection schemes for the development of pedestrian tracking,which can be selected according to the real-time requirements in the actual tracking task.