Research On Long-term Tracking Algorithm Of Human Target With Significant Posture Changes

As one of the most challenging key technologies in the field of computer vision,the object tracking technology has been widely used in many fields such as intelligent monitoring,human-computer interaction,driverless,virtual reality and even military.Due to the complexity of tracking scenes and changes of human targets,long-term human target tracking with significant posture changes has been a difficult problem to be solved by existing tracking algorithms.Most of the current tracking algorithms include model updating links,and carry out online update training on the observation models using samples which are obtained through simple sampling.However,the samples collected by simple sampling contain less information and have higher redundancy.Therefore,the trained observation models have insufficient classification ability.This article is based on the current research of generative adversarial nets,and improves the model updating links in the existing object tracking algorithms.A negative sample enhancement method and a multi-pose positive sample generation method based on the generation of confrontation network are proposed,and respectively integrated into the object tracking algorithm to improve the tracking effect of the tracking algorithm in long-term human target tracking tasks with significant posture changes.The main research work of this paper is as follows:(1)This paper analyzes the basic ideas of the existing object tracking algorithms based on deep learning method,then studies the frameworks and overall solutions of the algorithms.We analyze the advantages and disadvantages of the algorithms,and point out the importance of model updating in the object algorithm.(2)We delve into the model updating sections of the existing object tracking algorithms and analyze the source of the samples used in the observation model training.A negative sample generation method based on DRAGAN(Deep Regret Analytic Generative Adversarial Networks)is proposed to enhance the samples used in the model updating section.The negative samples generated using DRAGAN is more intrusive,and reduce redundancy,leading to the improvement of the classification ability of the tracking algorithm observation model.(3)We analyze the tracking drift caused by posture changes in human target tracking.From the perspective of image conversion,we study the changes between human target postures and transform them into cross-conversions between different domain images.A multi-pose sample generation method for human target based on StarGAN is proposed.StarGAN is used to generate multi-pose samples corresponding to human targets,and these samples are used as positive samples in the model updating link of the tracking algorithm.In this paper,the negative sample enhancement method and the multi-pose positive sample generation method based on the generative adversarial nets are respectively integrated into the object tracking algorithm based on deep learning.We conduct comparative experiments on these two improved algorithms with the OTB human target tracking data set to verify the effectiveness of both methods.The improved object tracking algorithms,which integrate sample generation methods based on generative adversarial nets,can suppress tracking drift in long-term human target tracking tasks with significant pose changes,improve tracking performance,and enhance robustness.