Object Tracking Algorithm Based On Feature Level Fusion

Recently,the technology of vehicle to everything(V2X)has been greatly developed,and the object tracking algorithm is one of the significant and basic components of the V2X.Real-time,precise,and adaptive object tracking al-gorithm is of great significance to the V2X system.However,the widely-used object tracking methods have the following shortcomings:poor real-time per-formance,low precision,and low adaptability,which means that the require-ments cannot be satisfied in both theory and practice.Therefore,the research of object tracking algorithm based on feature level fusion has theoretical and engineering significance.To solve the shortcomings,this thesis proposes var-ious solutions and strategies for object tracking algorithm.The main contents of this thesis are as follows:1.Convolutional feature selection algorithm for real-time object tracking.First of all,object tracking based on feature level fusion needs to extract object feature.So,how to select few and effective feature in order to meet the requirements of precision and real-time performance is a key basis for the fea-ture level fusion.Therefore,a convolution feature selection algorithm for real-time object tracking is proposed for solving the poor real-time performance.Specifically,a new objective function based on adaptive weight is formulat-ed to evaluate and select the convolution features.In addition,quadratic pro-gramming is introduced to solve the proposed objective function,which im-proves the optimization efficiency.Through the proposed feature selection,the relevant convolutional features are selected,and the remaining convolutional features are removed.Compared with existing algorithms,our proposed con-volution feature selection algorithm improves the tracking precision,reduces the computational complexity and guarantees the real-time performance of the tracking algorithm.2.Hierarchical convolutional feature fusion algorithm for precise object tracking.The above scheme deals with the single convolution feature.However,how to fuse the hierarchical convolutional features at feature level to obtain the fusion feature is a key issue to improve the tracking precision.Therefore,a object tracking algorithm based on hierarchical convolutional feature fusion is proposed for precise object tracking.Specifically,a new hierarchical con-volution feature fusion framework is proposed.First,the hierarchical convo-lution features are cascaded at the feature level to obtain the high-dimensional cascading features,and then the feature dimension is reduced by using the convolution layer.Next,the channel attention network is added to weight the cascading features after dimensionality reduction.Finally,the discriminative correlation filter is used to locate the target.The experimental results show that our algorithm performs better with low-dimensional features in visual tracking database and V2X scenarios.The tracking system needs to be adaptively adjusted when the appearance of object changes,and the appearance changes of object also need a stable and adaptive mechanism in the tracking system.For the adaptability,this thesis studies the following two aspects:3.Training sample set management algorithm with non-ideal samples.In above-mentioned two scenarios,it is assumed that the ideal case is that the training samples are correct.However,in real V2X scenarios,training samples are not all accurate,and incorrect training samples(non-ideal sam-ples)degrade the tracking performance.To solve the problems,we propose a new training sample set management method based on adaptive sample weight,which can evaluate and remove the incorrect training samples.Specifically,similarity learning is used to evaluate the quality of training samples online,and the obtained similarity score is used to judge whether the training sample is correct or not.If the similarity score is below a certain threshold,the current training sample is considered incorrect.Namely,the current sample weight is set to zero,which means that the training sample is removed from the training sample set.The experimental results show that the proposed scheme achieves superior performance in visual tracking database and V2X scenarios.4.Scale-adaptive object tracking algorithm.In addition to the need of locating the object precisely,the scale estimation is an important component.Either too large or too small the estimated object scale will affect the locating precision.Conversely,the estimated deviation of the object location will affect the performance of the scale estimation.The s-cale estimation methods in the above algorithms use only single strong feature to estimate the object scale,which leads to the low precision and discriminabil-ity.To solve the problems,we propose a novel scale space multi-kernel tracker,which can achieve more accurate scale estimation.Specifically,the scale esti-mation is performed based on the advantages of multiple kernels and multiple complementary features,and we employ a kernel for each feature.The experi-mental results show that the proposed scheme has significant improvement on image sequences annotated with scale variation.Finally,we summarize this thesis and present the future work.