Object Tracking Algorithm Based On Siamese Convolutional Networks

As a basic task of computer vision,object tracking is one of the hot research fields.Visual object tracking has strong application value in automatic driving,augmented reality,and autonomous robots,and thermal infrared object tracking can be used to achieve night surveillance and night assisted driving because it is not affected by light.In the current object tracking field,the siamese trackers show great potential in terms of balancing accuracy and speed,but there is still much room for improvement in the face of complex and changing tracking scenarios.Firstly,most of the existing siamese trackers only consider the features of the first frame and rarely benefit from the inter frame information.When facing challenges such as occlusion and fast motion,the lack of latest motion information can degrade tracker performance.Secondly,there is only brightness component information in the infrared image,lacking color information,which is easily interfered by similar targets and clutter background.For the above problems,this thesis focuses on the application of siamese networks in object tracking and performs algorithm optimization research on visible object tracking and thermal infrared object tracking.The main contributions are as follows:(1)A real time object tracking algorithm(Siam3D)based on spatio temporal convolution and siamese networks is proposed.This thesis explores the improvement of the feature representation of the initial template frame by the rich information in the latest continuous frames.Specifically,the latest frames after spatio temporal convolution is used to generate an attention map,and then the attention map and the features of the first frame are multiplied to obtain an updated template.Using attention maps,templates can adaptively handle occlusions and deformations of objects.The spatial convolutions of all frames in the model are shared,so the feature results can be reused and the added update module hardly adds tracking time.This module can be embedded in different siamese trackers.(2)An infrared object tracking algorithm(Siam DCT TIR)based on frequency domain transformation and feature selection is proposed.In order to speed up the infrared tracking and effectively use the brightness component information,this thesis carries out the discrete cosine transform on the infrared image and then further extracts the depth feature.By reducing the convolutional layer and pooling layer,the tracking speed is improved and the practicability of the tracking model is increased.At the same time,this thesis introduces the feature weighting mechanism of template frame to search frame to highlight the target and suppress the background.In addition,due to the small size of the infrared dataset,a convenient and easy to use data augmentation and pre training scheme based on grayscale is proposed,which improves the training effect of the model.In this thesis,a large number of experiments such as module effectiveness analysis,attribute analysis,and qualitative analysis are performed on authoritative data sets including visual object tracking and thermal infrared tracking,which verifies the effectiveness of the proposed object tracking algorithms.While ensuring real time tracking speed,the improved algorithms proposed in this thesis have improved accuracy compared to the baseline algorithms.