Research On Online Visual Object Tracking Algorithm Under Complex Dynamic Scenes

Object tracking is one of the hot topics and important research directions in the field of computer vision. It has been widely applied in the fields of video surveillance, intelligent transportation, robot navigation, intelligent vehicle drive assistance and human-computer interaction, etc. Over the past few decades, online object tracking technique has achieved much progress. However, the objects under complex dynamic scenes usually undergo kinds of uncontrollable apparent and motion changes, which make it still a very challenging task to design a robust, stable and effective online object tracking algorithm. Our research work is carried out about the video object tracking in complex dynamic scenes, and the main contents and research results include the following aspects:Firstly, in complex dynamic scenes, it's often unable to accurately describe the changes of objects with fixed features or feature set. In order to solve this problem, a robust tracking algorithm based on adaptive feature selection is proposed, which can adaptively select the best distinguishing features to describe the object according to the changes of object and background. For the problems that the candidate feature pool of the online AdaBoost algorithm are not robust and classifiers are easily effected by the influence of improper samples during the update process, a construction mode of the candidate feature pool which combines color and pyramid gradient orientation histogram features is proposed. Occlusion detection of the tracking result of each frame is executed in order to avoid the phenomena of drift. Lots of comparison experiments show that the proposed algorithm can achieve more robust and accurate results than other object tracking algorithms when the objects confront the situations like illumination variation, scale change, part occlusion, background disturbance and pose variation, etc.Secondly, a novel multi-scale Mean Shift tracking algorithm based on multiple appearance models is proposed to deal with the problems caused from the relative simplicity of the single appearance model and the absence of the update strategy under the original framework of Mean Shift tracking. Multiple appearance models are obtained by using Sparse Principal Component Analysis with feature template sets. Then the Mean Shift trackers are running in parallel under multiple scales through taking each appearance model as the reference model. In addition, the problem that the gradient histograms feature is easy to be trapped into local extreme points under Mean Shift tracking framework has been studied, and generalized gradient vector field is adopted to describe the shape information of object. Experimental results show that our algorithm is more robust and stable against kinds of challenging situations like pose variation, background disturbance and abrupt motion in comparing with other competing tracking algorithms.Thirdly, in order to overcome the problems associated with severe changes in pose, motion and occlusion under complex dynamic scene, a novel tracking algorithm including a discriminative model based on superpixels and a generative model based on global color and gradient features is proposed. Through combining these two models, the distinguishing and invariance of target appearance features description are increased. Furthermore, an update strategy of codebook based on second time discrimination of the tracking results is proposed to avoid the codebook is updated by the wrong results. Experimental results demonstrate that the proposed algorithm performs more stably and robustly against several state-of-the-art algorithms when dealing with complex situations such as pose variation, part occlusion and background disturbance.Finally, the object tracking algorithms based on Sparse Representation only consider the objects apparent features and ignore the object internal spatial relationship during tracking process. To solve this problem, a robust tracking algorithm based on Hough voting with local features is proposed. In our algorithm, the problems that how to establish the matching relationships between local patches for voting and the vectors in codebook, and how to assign a weight to each patch for voting under generalized Hough Transformation framework have been discussed. The former is solved through Sparse Representation theory, and the latter calculates the voting weight according to the category distribution significance of these local patches and the voting significance of the relevant vectors in codebook. Comparison experiments results demonstrate that our proposed algorithm performs more stably and robustly to cope with the situations like illumination variation, motion blur, abrupt motion and background disturbance against other tracking algorithms.Above all, four kinks of object tracking algorithms have been proposed to solve the situations of different variance of appearance and motion confronted in the process of online visual object tracking under complex dynamic scenes.