Reserch On Tracking Object Location In Siamese Network Based Tracking Algorithm

Object tracking algorithm has been widely used in security monitoring,intelligent robot,intelligent driving,and other fields.This demand also makes object tracking algorithm become one of the research hotspots in the field of computer vision.Because of the excellent performance of deep learning network of the field of image algorithm and the continuous improvement on the computing power of hardware,the object tracking algorithm based on deep learning has also become a research hotspot in the field of object tracking.Most of the current strongest object tracking algorithms based on deep networks belonged to the siamese network series of algorithms.The research content of this paper is the improvement on the object location of the siamese network based tracking algorithm.Aiming at the imbalance of foreground and background samples and the poor quality and utilization of anchor in the siamese tracking algorithm that uses the regional proposal network(RPN)to locate object,this paper introduces the feature alignment network of anchor free object detection into RPN and proposes a tracking algorithm based on siamese network which can use the prior knowledge of object size of the previous frame.The experimental data show that the foreground and background samples of RPN is more balanced,and the quality and utilization of anchor boxes is also improved.Aiming at the sampling bias problem after introducing the feature alignment network into RPN,the exact sampling convolution is used to eliminate the sampling error of the feature alignment,this paper uses accurate sampling convolution to eliminate sampling error.Based on the idea of dividing and conquering,this paper divides the regression task of the bounding box into two subtaskses: locating the center of the target before determining the target bounding box,and guides the generation of the anchor by using the distribution of the object size change rate between the continuous frames,which improves the quality and utilization rate of the anchor again.The experimental data show that the improved RPN can improve the object location performance of the siamese network tracking algorithm.Aiming at the problem that the number of positive samples decreases after increasing the intersection over union(Io U)threshold of the front/background classification for training RPN,the paper uses the idea of cascaded localization to make the number of positive and negative samples remain relatively balanced with higher Io U threshold.At the same time,because the paper only uses the convolution layer to complete the regression of the object bounding box,the timeliness is higher than the traditional cascade object location strategy.The experimental data also show that the object location accuracy of the further improved tracking algorithm is improved again.In this paper,the object location method of siamese network tracking algorithm is improved.Firstly,the feature alignment network is introduced to utilize the prior of object size.Secondly,the sampling error of the feature alignment network is eliminated by precise sampling convolution and the network is improved based on the divide and conquer idea.Finally,the location performance of the tracking algorithm is improved again by the cascade location idea.