Research On Object Tracking Based On Multi-scale Neural Network In Complex Background

Object tracking is a fundamental and challenging problem in computer vision as it impacts by many factors,such as target scale changes,fast motion,background clutters and etc.With the continuous development of deep learning and the unique advantages of deep neural networks,how to design a more accurate tracking model through deep neural networks has become a hot research topic in the current object tracking area.In this paper,we propose a multi-scale neural network,and the contributions are as follows.It is difficult to predict the target scale in the actual object tracking scene.Many existing object tracking algorithms conduct a large calculation on scale prediction in a small scale range.In this paper,a scale prediction model based on multi-scale features is proposed.The model builds a multi-scale feature matrix to model different scales of the target,and maps the targets of different scales to network layers of different scales to encode multi-scale target.Based on this model,more accurate target scales can be regressed.In order to enhance the representation ability of the target feature,we propose a feature representation model based on the local feature fusion.The important information in the image is highlighted by adding channel attention and spatial attention.This module uses residual structure to achieve local feature fusion,it can get more abundant target features.Furthermore,the maximum loss function of IOU(Intersection Over Union)is combined to guide the network in detail learning,which generates expressive target features and achieves a more accurate tracking prediction.Due to the less discriminative ability of existing object tracking networks,target drifts often occur in object tracking.In this paper,a feature metric model based on Interframe association is proposed.This model reuses target features in historical frames,and calculates the similarity between the current frame and historical frames by using the cosine distance to select the best proposal,which greatly improves the discriminative ability of the network and yields a more accurate prediction.Finally,the individual modules of the proposed method,including the scale prediction module of multi-scale features,the feature representation module based on the attention mechanism and the feature metric module based on temporal information,are verified through the ablation study.In addition,comparative experiments are conducted against numeorus state-of-the-arts on three public datasets,including VOT2019,VOT2018,and OTB.The results demonstrate that the proposed method is superior to most current object tracking algorithms.