Research On Multi-Template Matching Algorithm For Target Tracking In Complex Scenes

As the critical technique in the target tracking systems, the main task of image matching is automatically recognizing and accurately locating the target in the image sequence. With great significance and practical value, image matching technique is a core subject in the domain of computer vision, and it has been applied to many fields including military, medicine, business, traffic, scientific analysis and so on. Based on the application of target tracking in the complex scene, this paper makes an in-depth theory research and analysis of image matching algorithms. By comparing and analysing the principles and performance of different algorithms, an intensity correlation-based multi-template matching method is proposed. The experiments prove that this method has considerable advantage in restraining the matching error and improving accuracy of the system. Additionally, this method has a good adaptability of target appearance variation.Through the in-depth study on image matching theory and by comparing and analysing the related research, this paper adopts the intensity correlation-based template matching algorithm as the chief method to track the target in the complex scene. This kind of image matching algorithm has good adaptability to various scenes and low complexity, and it's easy to be realized in parallel processing. This paper makes plenty of theoretical researches and experimental analysis aiming at the intrinsic defect of the algorithm in the process of image sequence matching.In terms of algorithm optimization, this paper researches the influence of matching error sources, correlation surface shape, and template strategy on algorithm performance. Additionally, it proposes a method to estimate matching results through analysing the transformed correlation surface, and introduces the peak area (PA) description to forecast the decrease of result accuracy. On the basis of wide comparison and analysis of related foreign and domestic study, chapter three emphasizes a geometric transform-template group matching algorithm, and validates its performance in experiments. At last, it puts forward an improvement solution of this method through the correlation surface analysis.Comparing with the traditional template matching method which uses single template (including the target subimage) to match the real-time image sequence, this paper proposes a multi-point template matching algorithm. This algorithm uses multi-point locations (including the target location) to extract several templates from the same frame of the image sequence, and then finds the matching results through intensity correlation-based template matching algorithm respectively. After a correctness selection processing of these results, it makes a geometric correction of those wrong matching points. Here we adopt timely update template strategy in this algorithm. The templates of each location are updated when the correlation surface shape is not satisfactory.Through a series of experiments in MATLAB simulation environment, the results have demostrated that the multi-template matching algorithm shows better performance in precision and stability in comparison with others.