Study On Multi-Object Tracking Method Based On Bionic Intelligence

Detecting and tracking of moving objects in image sequence, which is important to the middle-level and high-level processing such as the understanding of target behavior, is one of the most popular research topics in computer vision and digital video area. Detecting and tracking moving objects robustly and fast is the most important issue in the study of the moving objects detection and tracking. To resolve the key issue, this thesis investigates the methods for multi-object detection and tracking based on bionic intelligence, as follows:1. Researches on some popular adaptive modeling methods for dynamic background. Background modeling is the basis of background subtraction, which is one of mostly used methods for moving object detecting. As the real scene is always changing because of some factors such as lighting changes, it is necessary to build a model that could adapt to the changing of background and describe the background accurately. This thesis researches on some popular adaptive dynamic background modeling methods, summarizes the advantages and disadvantages of them.2. A memorizing background modeling algorithm based on the Gaussian mixture background model is proposed. The denoising method, as well as the shadow detection that is based on the HSV color characteristics, is embedded in the algorithm. And a learning rate adaptable updating method is used to segment the objects that move slowly or stop for a short while. In addition, imitating the process that human perceive the environment, a memorizing method is introduced to the algorithm to handle sharp changes of the background.3. A novel state-transition and observation model is proposed, which is used in particle filter to track multiple moving targets robustly. To solve the computation problem of the popular observation model that is used in the particle filter to track moving objects in image sequence, this thesis builds a new state-transition and observation model based on the detection result of the memorizing background model. The experiment results show that using the proposed model, objects could be tracked fast. And even when covered or blocked, each object still could be tracked robustly.