Research On Target Tracking Algorithm Based On Siamese Network

In real life,vision is the most important way for human to obtain information,far more than sound and smell.Computer vision technology has been successfully applied in unmanned driving,auxiliary medical,intelligent security and other fields.Dynamic video target detection and recognition has a wider range of applications.The goal of visual tracking is to detect continuously moving objects in a series of sequential images,obtain the motion data of the objects,further extract the motion trajectory of the objects,and analyze the motion of the objects.In recent years,the success of convolution neural network has provided the foundation for many deep learning-based trackers,and the deep features extracted from the pre-training model on the large-scale image network dataset are widely used in the target representation,At present,there are two kinds of target tracking algorithms based on deep learning,one is tracking algorithm based on correlation filter,the other is end-to-end neural network tracking algorithm.The target tracking algorithm based on Siamese neural network is a typical end-to-end deep tracking algorithm.This paper mainly studies the target tracking algorithm based on Siamese neural network and its improved algorithm.(1)The target tracking algorithm based on Siamese neural network is widely recognized in the field.In this paper,we propose a Siamese network tracking algorithm based on multi-layer fusion,which combines shallow texture features and deep semantic features to optimize the tracking performance.The main structure of this method is two parts,one is that the top-down module transmits the high-level semantic information to guide the learning of the low-level features.The other is that the lateral module which transform the bottom-up features and integrate the horizontal features.The experimental results show that the accuracy of tracking is better.(2)ln addition,a target tracking algorithm based on corner detection and kalman filter is proposed and extended to multi-target tracking.Once the tracking fails,the siamese neural algorithm can no longer continue to track,and the corner detection algorithm can be used to complete the global search of the image.Even if the tracking fails,the tracking can be recovered in the subsequent detection.The tracking algorithm based on the target detection is greatly affected by the video noise.The kalman filter solves the motion estimation problem of the tracking and makes the tracking more smooth and stable.(3)Multi-target tracking is larger and more complex than single-target tracking.On the third chapter,we optimize the speed of the algorithm,and combine the advantages of siamese network and corner detection,we propose a multi-target tracking algorithm based on siamese network and corner detection.Firstly,the algorithm uses corner detection to detect the global target,cuts the target to be tracked after the detection result is obtained,and then uses the siamese network to continue tracking.Once the tracking fails,the target information is retained,which is matched in the subsequent target detection,and the tracking target is continued when it appears again.The fire module is used to greatly reduce the number of parameters in the network,the network is used to detect multiple targets at fixed intervals,and the siamese network is used to track multiple targets in parallel,and the tracking process uses kalman filter to remove the tracking noise.The experiment verifies that the multi-target tracking algorithm can accomplish the multi-target tracking task well.