Research On The Key Technology Of Visual Object Tracking In The Challenging Conditions

With the rapid development of computer technology,as the key technology of linking up image processing and scene understanding,visual object tracking is widely used in some fields.In the past decades,due to the introduction of new technologies and the enhancement of computing power,researchers have suggested a variety of excellent tracking algorithms to promote the rapid development and progress of the field of visual tracking.However,due to the complexity of the target environment and the instability of the target itself,the target tracking technology is still an open research topic.Especially,it is still a very challenging problem to achieve accurate,robust and timely tracking in complex scenes.On the basis of detailed analysis of the working mechanism of the existing tracking technologies and the latest academic achievement in this field,this paper conducts in-depth research on the tracking strategy under complex environment.The main contributions and innovations of this paper are as follows:1.This dissertation suggests a Parts and Spatial-Temporal contexts based Kernelized Correlation Filters tracker,abbreviated as PSTKCF tracker.Firstly,a tracking confidence evaluation mechanism is proposed to measure the tracking confidence of the local models and global model.Base on the proposed evaluation mechanism,a local model is designed for the weighted fusion of multiple reliable parts according to the confidence degree,therefore reducing the risk of model contaminated by interference information.In addition,a global model is proposed by combining the global information and global-local interaction information to estimate the final state of the target.Meanwhile,the unreliable parts are reset to ensure the quantity and quality of the reliable parts in the global model.Finally,a simple and effective scale estimation mechanism is proposed to avoid the influence of searching preset scale pool at the cost of tracking speed.Experiments on standard data sets show that the target appearance model constructed by our proposed method can robustly adapt to the significant changes of target and track the target accurately.2.This dissertation proposes a Correlation Particle Filters based tracker via Occlusion and multi-Peak Detection,abbreviated as CPFOPD tracker.Firstly,a part-based strategy combining artificial sampling and random sampling is proposed.According to the different object,this strategy can guide the particles to be dropped within the specific range,which reduces the disadvantages of blind sampling of the existing particle filters and also enhances the robustness of the tracking system in the face of abrupt changes of target appearance.Then,considering the same motion trajectory between the local samples and global target,a target estimation method based on tracking confidence detection is presented.In the estimation model,the tracking system optimizes particle distribution and reduces the number of particles required by guiding particles towards the target potential location.The combination of particle filter and correlation filter makes up for shortcomings each other and plays a significant role in improving the performance of tracking.Experimental results show that our proposed CPFOPD tracker is better than other advanced algorithms when the target is affected by illumination,occlusion,and motion blur.3.This dissertation presents a Double Attention and Multi-Convolutional-layer based on Siamese Network tracker,abbreviated as DAMCSNet tracker.Firstly,in order to eliminate the interference of background information to the target,a scene attention detection network is suggested.The color histogram of the target in current frame is used to calculate and obtain the scene attention map.Next,a feature fusion network is proposed to realize the multi-resolution features integration based on scene attention map,which makes the tracker focuses more on the valuable target area.Then,combining with the attention mechanism,a channel attention enhancing network is proposed by selectively describing the target from the view of channel level to improve the representation ability of target model.When the target is by factors such as illumination,rotation and motion blur,our proposed tracker MC2 SNet can better estimate the state of the target compared with other state-of-the-art algorithms.In summary,this paper carries out in-depth research on visual object tracking in complex environment in many aspects,and realizes tracking tasks of video sequences with the help of the technologies of correlation filter,particle filter and deep learning respectively.In addition,our three tracking methods provide new and different ideas for the tracking field.Moreover,considering the real-time requirements in practical application,many effective and efficient design proposals are adopted to improve the computational efficiency in the period of design.For evaluating the comprehensive performance of the proposed algorithms,the dissertation also organizes extensive comparative experiments on several public visual object tracking datasets.The experiments results show that our proposed three tracking methods achieve excellent tracking performance.