Infrared Pedestrian Target Tracking Method With Siamese Network

Infrared pedestrian tracking plays an important role in night driving,military reconnaissance and other fields.Compared with the common color video,infrared video is easily affected by the surrounding environment.In the infrared pedestrian tracking,the appearance contour and gray distribution of pedestrian target often change greatly,which leads to the difficulty of tracking.To this end,the paper deeply researches convolutional neural networks and uses them for infrared pedestrian target tracking.This paper deeply analyzes the existing target tracking algorithms.In the target tracking system using traditional algorithms,the tracking effect is not comprehensive enough to only have strong effects on several attributes.The target tracking algorithm using deep learning shows outstanding advantages in success rate and precision.This paper uses the infrared pedestrian dataset to test the most mature tracking algorithm Siam RPN(Siamese Region Proposal Network)that using siamese networks.When factors such as waves affect the infrared tracking effect,the algorithm shows poor tracking performance.The se factors are common attributes of infrared pedestrian tracking video.This paper uses the infrared pedestrian dataset to test the most mature tracking algorithm using deep learning.It is found that the algorithm shows poor tracking performance when the similar heat source targets cross,the target size changes greatly and the video contains more background clutter,which are common attributes of infrared pedestrian tracking video.Aiming at the problems found,this paper proposes an infrared pedestrian tracking method based on video prediction and Siamese network.The proposed tracking method combines a video prediction network with Siam RPN.The prediction network can predict the target features of the current frame through the target images of the past frames.The output image of the prediction network is used as the target tracking template to improve the similarity between the tracking template and the detected target,and enhance the model matching ability in the target tracking,so as to improve the tracking ability of the infrared pedestrian target.This paper designs nine different network structures by changing the number of network layers,the number of target images and image frames used for prediction,and the network's tracking strategy.And through comparative experiments,the network structure with the best performance was selected.On the PTB-TIR dataset,an objective comparison is made with eight mainstream target tracking algorithms including Siam RPN.These algorithms are evaluated on nine attributes based on the tracking success rate and accuracy.The experimental results show that the tracking success rate and accuracy of the network proposed in this paper on various attributes such as intensity change,occlusion,and background clutter are greatly improved compared to other networks.It shows the good performance of infrared pedestrian tracking,and will have a broad application prospect in this field.