Object Tracking Based On Sparse Convolution Networks

Object tracking is a key topic of computer vision,which has an extensive application potential,such as video surveillance,automatic drive,human computer interference.Although considerable progress has been made in recent years,it remains some problems include appearance changes and external interference,which cause a challenging task of designing a robust tracker.Recently,convolutional network has been widely used in computer vision task as it achieved significant performance in feature extraction.Deep networks need significant amount of samples and time for offline training,whereas only the state of position in first frame was given.As a result,how to design a robust online tracking algorithm using convolutional networks is a question that is worthy of further investigation.The research of this thesis is also focus on this problem and the main contributions are show as follows,1.The features extracted by traditional convolutional networks are lack of local spatial features and thus sensitive to external interference such as heavy occlusion and scale variation due to pool operation.To address this issue,this paper proposes a sparse convolutional network for online tracking,which aim to extract local spatial features by using sparse representation.Furthermore,all local spatial features can be used to construct the inner geometric layout to handle external interference.2.Based on sparse convolutional network,we proposed a sparse gradient convolutional tracker without pre-training(SGC).Note that the convolution operations of traditional networks are based on raw data,thus these networks are vulnerable to dramatic appearance changes when the structure is not deep enough.In SGC,we first integrate gradient features with a two-layer convolutional network to alleviate this problem and then design a robust online tracking algorithm by using particle filter framework.3.In order to improve the speed of tracker that is based on sparse convolutional networks,we also propose a real-time tracking algorithm by a convolution-based complementary model(RTC).The RTC are comprised of a convolution-based discriminative model(CDM)and a convolution-based generative model(CGM).In CDM,we first integrate the gradient and color features to construct the appearance model and combine the background information to predict the center position.Then,a sparse convolution network is embedded in generative model to extract foreground information and construct spatial structure layout to estimate the scale variation of the target.