Real-Time Movement Detection And Object Tracking In Video Sequences

Research of digital image processing and computer vision is developing very rapidly in recent years. It is now widely used in military and civil applications, and is becoming the most important way for intelligent machines to access external information and understand the world. Movement detection and object tracking, topics of this thesis, are the two most important applications of computer vision.The research background is mainly on providing key information for world modeling, path planning, navigation, and other senior decision-making of an Autonomous robot in the indoor environment, especially for tasks such as environment monitoring, target following, obstacle avoidance. As two relatively independent applications, movement detection and object tracking are presented respectively in this thesis, both include algorithm theoretical work and experimental verification.For movement detection, background subtraction algorithm based on multi-Gaussian background model is used as the core algorithm, and Gaussian filter and morphology method are used as the supplement methods. In this thesis a complete movement detection solution is provided and its feasibility and efficiency is verified by experiments. Due to the shortcomings of the algorithm, large background area is miss-detected periodically, so an improved model update algorithm is proposed, which can entirely solve this problem. For scenes of camera moves and shadow interferes which are difficult for movement detection, a new model reconstruction algorithm and HSV space-based shadow filter algorithm are proposed. Their efficiency and effectiveness are verified by experiments.For object tracking, this thesis uses the Mean Shift algorithm as the core algorithm. Based on the proposed 2-dimensional histogram, the target template is established according to its color and gray-scale distribution, and Bhattacharyya coefficient is used to measure the matching similarity. Both the theoretical analysis and experimental results are provided to verify its feasibility. In scenes of the target's features is not distinguished from background features obviously and scenes that contains intense movement, the shortcomings of the tracking algorithm could cause failure results. To solve these problems, an improved similarity evaluation method and a Kalman predictor assistance method are proposed. Their advantages are verified by experiments. The relationships between movement detection and object tracking algorithm are also discussed. Some preliminary exploration for the multi-task and cross-integration work in computer vision field is presented. We also propose a new movement analysis method based on the historical movement detection results, which can estimates the velocity and direction of the movement. The detecting result can be used in tracking problem as the target area. In this way, the semiautomatic tracking algorithm can be improved to a complete automatic solution.