Research On Pedestrian Object Tracking Algorithm Based On Fusion Of Flow Net And Siamese Network

The intelligent development of modern society is rapid,and surveillance cameras in public places are becoming more and more popular,and the corresponding pedestrian object tracking technology based on surveillance video also has broad application prospects.Recently,deep learning has developed rapidly and the performance of deep learning-based object tracking algorithms has also improved a lot.However,there are still many challenges in actual application scenarios.This paper analyzes the similarity interference and the occlusion challenge in the object tracking challenge in depth,and proposes a siamese network object tracking algorithm based on optical flow prediction to solve the problem that the object is easily changed when the object is interfered or occluded by similar objects.This problem in turn improves the robustness of the object tracking algorithm.The main contents of this paper are as follows:The algorithm in this paper integrates the Flow Net with the Siamese network to achieve a high-precision object tracking algorithm under the interference and occlusion of similar objects.First,the Flow Net is used to extract the motion information between the searched image frame and the previous frame image to generate an optical flow vector diagram,regression to obtain the object predicted position,crop out the predicted position image block,and enter the search branch in the lightweight siamese network.Make the search area of the tracking algorithm the position where the object may appear,achieve the purpose of preventing similar objects in other positions in the figure from interfering with the tracking algorithm,and solve the problem of the object being interfered by similar objects.Then the optical flow information extracted through the Flow Net is combined with a similarity threshold to determine whether the object is changed.When the similarity value between the predicted object and the template object is lower than the threshold,the optical flow vector diagram of the optical flow prediction module is multiplexed to compare the optical flow vector at the object position between adjacent frames.When the optical flow vector difference is higher than one When the threshold is set,it is considered that the object has been changed incorrectly,and the object is corrected.According to the optical flow vector information of the preceding and following frames,people with similar optical flow vectors and high similarity in the latter frame are reselected as the predicted object,and the tracking object is corrected.It can better deal with the problem that the object is easily changed when the object is mostly occluded and partly occluded by similar objects.In this paper,the backbone network in the siamese network structure is improved to a lightweight convolutional neural network.The 1×1 convolution kernel is used multiple times to compress the number of channels,reduce the number of parameters,improve the calculation speed of the tracking algorithm,and make the siamese network more efficient.Non-linear expression ability and generalization ability have also been improved.The convolutional features extracted by the third convolutional layer and the fifth convolutional layer are merged to enrich the convolutional features,thereby further improving the performance of the tracking algorithm in this paper.The algorithm in this paper is experimentally verified on the OTB(Object Tracking Benchmark)data set proposed by the University of California.Experimental results show that the siamese network object tracking algorithm based on optical flow prediction can effectively solve the problem of object error change caused by the interference of similar objects and the object being occluded.The success rate curve AUC value reaches 0.635 under the test sequence of the OTB occlusion challenge.The accuracy rate is 0.830,which is 18.7% and26.9% higher than the KCF(Kernel Correlation Filter)algorithm success rate and accuracy rate respectively.In this paper,the Flow Net and the siamese network are integrated for object tracking,which greatly improves the performance of the object tracker in the scene of similar interference and occlusion.