Dynamic Object Detection And Tracking Based Energy Minimization

Computer vision based video monitoring technology is usually used to analyze image sequence and their relations of certain objects by using image processing methods so as to detect and track the dynamic objects in a specific scene.However,the existing methods still have some problems such as false detection positives,false negatives and drift existing for pedestrian detection and tracking due to the complexity of the scene and the moving characteristics of the pedestrian objects.This thesis mainly aims to research the detection algorithms and the tracking algorithms of dynamic target,and proposes a continuous energy minimization framework for dynamic targets.The framework considered many related factors of tracking multiple targets to model the system comprehensively and improve the objects tracking accuracy.The main contents are as follows:Firstly,the tracking-by-detection approach is analyzed and used in this research by analyzing the commonly used tracking methods and comparing their characteristics and application environments.Secondly,to detect the objects,this thesis proposes an improved detection algorithm based on orientation gradient histogram.The features of the gradient histogram of the target orientation for the input image are extracted,and classified by support vector machine.In feature extraction,motion segmentation is carried out firstly,then the segmented regions are input to the HOG classifier to reduce the amount of computation.As for target tracking,the extended Kalman filters were used to get the original tracking solutions.Thirdly,an optimal model of continuous energy minimization is proposed to improve the performance of the solutions.The model fully considered some critical aspects of target tracking such as occlusion,appearance and dynamics of the multiple objects.In addition,the tracking performance is improved by using standard conjugate gradient and six jumping movement type optimization model to get the lowest energy corresponding to the tracking scheme.Finally,the accuracy and precision of the multi-target tracking are calculated,and the tracking effect of the improved method is evaluated quantitatively by using the multi-object tracking evaluation index CLEAR MOT.